diff --git a/.github/workflows/site.yml b/.github/workflows/site.yml new file mode 100644 index 0000000..160cd61 --- /dev/null +++ b/.github/workflows/site.yml @@ -0,0 +1,36 @@ +name: Deploy website + +# Publishes site/ to GitHub Pages. One-time repo setup: +# Settings → Pages → Build and deployment → Source: "GitHub Actions". +# Custom domain later: add site/CNAME with the bare domain, point DNS +# (A/AAAA to GitHub Pages IPs or CNAME to .github.io), done. + +on: + push: + branches: [main] + paths: ['site/**', '.github/workflows/site.yml'] + workflow_dispatch: + +permissions: + contents: read + pages: write + id-token: write + +concurrency: + group: pages + cancel-in-progress: true + +jobs: + deploy: + runs-on: ubuntu-latest + environment: + name: github-pages + url: ${{ steps.deployment.outputs.page_url }} + steps: + - uses: actions/checkout@v4 + - uses: actions/configure-pages@v5 + - uses: actions/upload-pages-artifact@v3 + with: + path: site + - id: deployment + uses: actions/deploy-pages@v4 diff --git a/.gitignore b/.gitignore index b3e1193..61ab52e 100644 --- a/.gitignore +++ b/.gitignore @@ -37,6 +37,7 @@ c/tests/test_schema_gbnf c/tests/test_schema_gbnf.exe c/tests/test_compat_direct c/tests/test_compat_direct.exe +result # oracoli tiny generati (make_glm_oracle.py) e dati benchmark scaricati c/glm_tiny/ diff --git a/README.md b/README.md index 9e33501..aaf7463 100644 --- a/README.md +++ b/README.md @@ -3,6 +3,12 @@

+ Website + Latest release +

+ +

+ Website · English · 简体中文 · 繁體中文 · Italiano

@@ -44,6 +50,16 @@ brightness is routing heat, and every expert routed in a turn flashes white. Hov as a 3-D galaxy — 13,260 characterised experts, 1,041 replicated specialists clustering by topic (poetry, law, Chinese, SQL…). Position is measured routing affinity, not a learned embedding. Drag to spin.

+## The vision + +Frontier models should not be sealed inside datacenters. colibrì exists so that +**anyone curious enough can open one up**: run a 744B-parameter mind on hardware +you already own, watch every expert fire in real time, and change the code that +does it. Not renting intelligence behind an API — *holding* it: probing it, +measuring it, improving it. Every optimisation in this project started with +someone measuring something on their own machine; the engine is deliberately +small enough that the next one can come from you. + ## The idea A 744B Mixture-of-Experts model activates only ~40B parameters per token — and @@ -61,6 +77,18 @@ So the model doesn't need to *fit* in fast memory — it needs to be **placed**: at int4) live **on disk** (~370 GB) and are **streamed on demand**, with a per-layer LRU cache, a learned pinned hot-store, and an optional VRAM tier. +Think of the core algorithm as **a JIT, but for weights**. A compiler JIT never +compiles the whole program — it watches what actually runs and compiles the hot +paths, just in time. colibrì makes the same bet about a 744B parameter space: +parameters are not resident state to be held, they are **data to be staged** +across a heterogeneous storage hierarchy (VRAM / RAM / NVMe), exactly when the +router proves they are needed. Measured routing heat decides which experts earn +which tier, the router runs a layer ahead so prefetch hides the staging latency, +and — like a JIT — the engine learns your workload: the more you run, the hotter +the right experts get. It works because routing has measurable structure (see +the [expert atlas](https://github.com/JustVugg/colibri/issues/175)) — and +structure is cacheable. + The engine is a single C file (`c/glm.c`) plus small headers. No BLAS, no Python at runtime, no GPU required. @@ -203,6 +231,13 @@ COLI_MODEL=/nvme/glm52_i4 ./coli doctor # read-only readiness check The engine at runtime is pure C — python is only used by the one-time converter and the optional API gateway. +**On Windows?** You don't need to build. Download the +`colibri--windows-x86_64.zip` from +[Releases](https://github.com/JustVugg/colibri/releases), unzip it, rename +`colibri-*-windows-x86_64.exe` → `glm.exe` (so the `coli` launcher finds the +engine), install [Python 3](https://www.python.org/downloads/), then run +`coli chat`. Full walkthrough in the [Quick Start guide](docs/quickstart.md#windows). + Prefer a `coli` command on your PATH? From a checkout, `pip install -e .` registers it (the engine itself still lives in `c/` — this is an editable install from the clone, not a standalone wheel). @@ -220,6 +255,18 @@ install from the clone, not a standalone wheel). | Grammar-forced drafts (structured output) | [docs/grammar-draft.md](docs/grammar-draft.md) | | Environment variable inventory | [docs/ENVIRONMENT.md](docs/ENVIRONMENT.md) | +## What's next + +- **Algorithmic research is active.** The current hierarchy is LRU + a learned + pin set; the next step is under way — smarter placement and scheduling, + overlap of CPU and GPU expert execution, and routing-aware speculation. + Everything lands the way this project always works: measured, reviewed, and + merged in the open. +- **More open models.** The tiering algorithm is model-agnostic: any MoE with + routed experts can be staged the same way. GLM-5.2 and OLMoE run today; + support for more open-weight families — **Kimi K2** (Moonshot AI), + **Qwen3 MoE** (Alibaba), **MiniMax** — is on the roadmap. + ## Supporting the project colibrì started as a one-person project on a 12-core laptop with 25 GB of RAM; @@ -260,6 +307,14 @@ The hummingbird weighs a few grams, hovers in place, and visits a thousand flowers a day. This engine keeps a 744-billion-parameter giant alive on hummingbird rations: 25 GB of RAM, twelve CPU cores, and a lot of disk patience. +## Acknowledgements + +colibrì is an engine; the minds it runs are a gift. Thank you to the teams +releasing frontier-class weights in the open — **Z.ai** (GLM), **Moonshot AI** +(Kimi), **Alibaba Qwen**, **MiniMax**, and **Allen AI** (OLMoE) — and to every +contributor who benchmarked, bisected, replicated an atlas run, or sent a patch. +This project is proof of what open weights make possible. + ## License Apache 2.0. GLM-5.2 weights are released by Z.ai under MIT. diff --git a/SUMMARY.md b/SUMMARY.md new file mode 100644 index 0000000..f4de311 --- /dev/null +++ b/SUMMARY.md @@ -0,0 +1,356 @@ +# E5 — MTLResidencySet over the existing malloc'd slabs (experiment branch) + +Branch: `e5/metal-residency-set` (cut from `origin/dev` @ `caa49f7`, per spec — E4 was cut +from `main` @ `72d3d37`; `backend_metal.mm`/`.h` are byte-identical between the two bases, +confirmed via `git diff 72d3d37 origin/dev -- c/backend_metal.mm c/backend_metal.h`). + +## The hypothesis + +E4 (MTLHeap-backed slabs) proved that batching residency declaration kills the GPU stall +(25.9s → 3.9s at cap16, −85%), but changing the *allocation* (heap sub-buffers instead of +malloc'd host memory) brought a +12–13s expert-disk-load tax, suspected first-touch/lock +contention on CPU-writes into GPU-owned heap pages. E5 decouples the two: keep the exact +same malloc'd slabs and per-slab `newBufferWithBytesNoCopy`-wrapped `MTLBuffer`s, and change +**only** residency bookkeeping — declare residency once, ahead of time, on a set attached to +the command queue, instead of once per command buffer via `useResource:`. If the stall +reduction survives without the load-path tax (malloc pages never change ownership), E5 wins. + +## What changed + +All mechanism code is confined to `c/backend_metal.mm` — +`coli_metal_register`/`coli_metal_unregister`'s existing signatures and every call site in +`colibri.c` (expert_load, uring_load_add, qalloc, kv_alloc, map_of_fd) are untouched; the +residency-set bookkeeping lives entirely inside those two functions' existing bodies. The +`colibri.c`/`backend_metal.h` touches are two, both coordinator-sanctioned: the validator +round-1 instrumentation hook (`coli_metal_resset_stats` + the gate-on-only `METAL-RESSET:` +stats line in `profile_print`) and the ported fslab-OOM unwind fix (see "Validator round 1 +fixes" item 4). Still a smaller diff shape than E4, +which needed a new alloc/free API and four new `glm.c` call-site arms because it changed the +allocation function itself. + +Env-gated `COLI_METAL_RESSET=1`, default OFF, runtime `@available(macOS 15.0, *)` guard with +a one-line stderr fallback when requested on an older OS or when residency-set creation +fails. Gate off ⇒ every new branch is skipped and behavior is byte-for-byte the stock path +(verified by inspection: `g_resset_enabled` starts `false` and nothing sets it except inside +the `COLI_METAL_RESSET` `getenv` branch in `coli_metal_init`, so `resset_add`/`resset_remove`/ +`resset_flush` are no-ops and `moe_submit`'s `useResource:` loop runs unconditionally). + +### Lifecycle (`c/backend_metal.mm`) + +- **Init** (`coli_metal_init`, end of the existing pipeline-setup `@autoreleasepool`): if + `COLI_METAL_RESSET=1` and `@available(macOS 15.0, *)`, create one + `MTLResidencySetDescriptor` (`initialCapacity=4096`, a presize hint only), call + `[g_dev newResidencySetWithDescriptor:desc error:&err]`, and `[g_queue addResidencySet:rs]` + — one set, attached once, for the process lifetime. Failure (old OS or creation error) + prints one stderr line and leaves `g_resset_enabled=false` — stock path. +- **`coli_metal_register`**: after wrapping the buffer exactly as today + (`newBufferWithBytesNoCopy`) and pushing the `g_slabs` entry under `g_slab_mtx` exactly as + today, calls `resset_add(b)` **after dropping `g_slab_mtx`** but before returning. + `resset_add` takes a dedicated `g_resset_mtx` (guarding only the set mutations and the + dirty flag), calls `[rs addAllocation:b]` and sets `g_resset_dirty` — **it does not + commit**. No Metal call ever runs under `g_slab_mtx` (validator round-1 fix; E4's audit + round 2 identified mutex-over-live-Metal-call as the leading suspect for its +12s + expert-disk regression). Re-registering a live base (no in-tree caller does today) drops + the replaced wrapper from the set via `resset_remove(old)` before adding the new one + (hazard-audit defensive fix — the set would otherwise retain the old buffer, and its + pages' residency, forever), keeping set membership an exact mirror of `g_slabs`. +- **`coli_metal_unregister`**: erases the `g_slabs` entry under `g_slab_mtx` (stashing the + buffer), then calls `resset_remove(b)` **outside `g_slab_mtx`**, before returning. + `resset_remove` (under `g_resset_mtx`) calls `[rs removeAllocation:b]` **and commits + immediately** — no batching — because the caller frees the host memory right after the + function returns. See UNCERTAINTIES for why this asymmetry is deliberate. +- **`moe_submit`** (the one function whose `use` list — resolved expert weight/scale slabs — + scales with LRU cache size): calls `resset_flush()` at the top (commits any pending adds + from `resset_add`, under `g_resset_mtx` — it never touches the slab lock), then, if + `g_resset_enabled`, **skips** the + `for(auto&b:use) [e useResource:b usage:MTLResourceUsageRead];` loop entirely — residency + is already guaranteed by the queue-attached set. Every other `useResource:` call site in the + file (`bind_gemv`'s weight/scale buffers, `coli_metal_attn_decode`/`coli_metal_layer_decode`'s + `Lb`/`Rb`/`kvbW`/`kvbS`/`inB`/`pnB`/`rwB`/`rbB`, `coli_metal_gemm`'s `wb`/`sb`) is + **left completely unchanged**, regardless of the flag — see "Why only `moe_submit`" below. +- **Shutdown** (`coli_metal_shutdown`): `[g_queue removeResidencySet:rs]` then clears the + globals, ahead of the existing `g_queue=nil; g_dev=nil;`. + +### Why only `moe_submit` skips `useResource:` + +Apple's `MTLResidencySet` class reference (developer.apple.com, fetched during design on +2026-07-18) is explicit: *"Residency sets don't support hazard tracking, so you need to +account for hazards with fences and events."* The SDK header on this box +(`MTLResidencySet.h`, read directly) is **silent** on hazard tracking — the statement comes +from Apple's online documentation and adoption guide ("Simplifying GPU resource management +with residency sets"), not the header (see UNCERTAINTIES for sourcing). Dropping +`useResource:` therefore risks losing whatever hazard-tracking value those calls provided. Rather than apply the residency set uniformly and +argue *in general* that hazard tracking isn't load-bearing, this diff draws the line at the +one call site the mechanism history actually implicates: + +`moe_submit`'s `use` vector holds only **read-only** (`MTLResourceUsageRead`), **indirectly +referenced** slab buffers — the kernel (`moe_gemv`) never touches them via `setBuffer:`; it +dereferences raw GPU addresses (`waddr[e]`/`saddr[e]`) baked into a separately-bound address +array (`bag`/`bau`/`bad`/`bsg`/`bsu`/`bsd`), which is exactly the "indirect reference" case +`useResource:` exists for. No GPU-side write ever touches these buffers, so there is no +write-after-write/read-after-write hazard for Metal's tracking to have been serializing in +the first place; the one real hazard — a slab unregistered+freed+reused by the CPU while an +async in-flight `moe_block_begin` command buffer still references it via a baked-in GPU +address — is a **CPU-write race that Metal's hazard tracking never protected against anyway** +(hazard tracking only covers GPU-side command dependencies visible through the Metal API; a +raw host-memory write via `pread`/`memcpy` is invisible to it regardless of `useResource:`). +That race is, and always was, the engine's own responsibility (slot/generation lifecycle: a +slab isn't freed while an outstanding async handle still owns it) — unrelated to E5. + +Every other call site (`bind_gemv`, attention K/V cache writes) either doesn't scale with +cache size (fixed per-layer dense tensors — no perf benefit to touching) or has real +GPU-side write traffic in the same encoder (`Lb`/`Rb` are written by `a_copy` and read by +`a_score`/`a_clat` within one encoder — currently ordered by explicit +`memoryBarrierWithScope:MTLBarrierScopeBuffers` calls already present in `encode_attention`, +not by `useResource:`'s hazard tracking, but touching them wasn't needed for the hypothesis +and was judged not worth the added surface area). Leaving them untouched keeps the diff's +blast radius matched to the one seam the fix-plan's v5 finding actually names. + +### Deferred-commit design (`resset_add` batches; `resset_remove` doesn't) + +`coli_metal_register` is called from parallel OpenMP loader threads in tight bursts +("warmup fan-out" — same phrase E4's audit used for the same threads). Committing on every +single `addAllocation:` would reintroduce a per-slab cost on the load path, which is exactly +what E4's own +12s regression looked like (mutex held across a live Metal call, serializing +loader threads). So `resset_add` only marks `g_resset_dirty`; the commit is deferred to the +next `moe_submit` call, which flushes once via `resset_flush()` before it relies on the set +for residency. + +This is correct — not just fast — because of an existing invariant the codebase already +depends on for `resolve()` to work at all: a slab's `coli_metal_register` call always +completes — including its trailing `resset_add`, which runs after `g_slab_mtx` is dropped +but **before the function returns** — before any dispatch that references that slab's +pointer can call `resolve()` for it (the caller in `colibri.c` cannot pass a freshly-loaded +expert's pointer to a dispatch before the load — which registers it — returns). After the +validator round-1 mutex split, the flush's synchronization runs through `g_resset_mtx` +alone: `resset_add`'s set mutation + dirty write and `resset_flush`'s dirty read + commit +are serialized by that one mutex, whose release/acquire pairs provide the memory ordering; +`g_slab_mtx` still orders the slab-table bookkeeping (register-before-resolve) exactly as on +stock. So any slab a given `moe_submit` invocation will resolve was `addAllocation:`-ed (and +marked dirty) strictly before that invocation's `resset_flush()` acquired `g_resset_mtx` — +the flush is guaranteed to cover it, regardless of what other threads are concurrently +registering unrelated slabs. The two mutexes are never held simultaneously anywhere, so no +deadlock ordering exists to maintain. + +`resset_remove`, by contrast, commits synchronously and immediately, with no batching, +because the caller (`colibri.c`, in every one of the four slab-realloc call sites, and in +`kv_alloc`) frees the underlying host memory *right after* `coli_metal_unregister` returns. +An uncommitted-but-still-set-member allocation pointing at memory the host has already freed +is a potential use-after-free the GPU could act on — deferring that removal is not a +performance-vs-safety tradeoff, it's just unsafe, so it isn't deferred. (The spec's own +lifecycle wording backs this reading: "`coli_metal_register` → add allocation + commit +**(batch commits where call pattern allows)**" carries a batching allowance that +"`coli_metal_unregister` → remove + commit" does not.) + +## Instrumentation parity + +No existing counter's semantics changed. `coli_metal_moe_times`/`coli_metal_moe_counts` +(`g_t_setup`, `g_t_gpu`, `g_t_kernel`, `g_t_scatter`, `g_moe_ok`/`g_moe_fb`/`g_moe_experts`) +are computed exactly as before — `resset_flush()` runs *before* `ts_start = mnow()` in +`moe_submit`, so its cost is **outside** `g_t_setup`, keeping the orchestrator's A/B harness +reading the same counters with the same meaning across stock/E4/E5. The flush cost is +surfaced separately (validator round-1 fix — the original design left it invisible, a blind +spot for the battery): a dedicated `g_t_resset_flush` accumulator timed around the flush in +`moe_submit`, exported via `coli_metal_resset_stats()` (`backend_metal.h`) and printed by +`profile_print` as its own `METAL-RESSET: flush N.NNs` line — a **separate line following +the `METAL:` line, mirroring E4's `METAL-HEAP:` convention, so the existing `METAL:` line +the harness parses keeps its exact format** — printed **only when the gate is on** (the +function returns 0 when off), so stock output stays byte-identical. The register-side +`resset_add`/`resset_remove` costs have no dedicated counter: they run inside the engine's +existing expert-load wait accounting (the `t_ewait` window in `colibri.c`), noted in a comment +at `resset_add`, so a load-path regression from set bookkeeping would already show in the +existing disk/wait numbers. `[METAL] residency-set: on` / the two fallback stderr lines from +`coli_metal_init` confirm which path a run took. + +## Validator round 1 fixes + +1. **REQUIRED, Metal calls hoisted out of `g_slab_mtx`** (`backend_metal.mm`): the original + design ran `addAllocation:`/`removeAllocation:`/`commit` while holding `g_slab_mtx`, the + lock the parallel OMP loader threads contend on — structurally identical to the + mutex-over-live-Metal-call shape E4's audit round 2 identified as the leading suspect for + its replicated +12s expert-disk regression, and the SDK header notes commit on a resident + set tries to make resources resident "instantly" (real synchronous work; this set is + resident from startup since it is queue-attached for the process lifetime). Fixed by + introducing a dedicated `g_resset_mtx` guarding only the set mutations + dirty flag; + `g_slabs` push/erase stays under `g_slab_mtx` exactly as stock; the two mutexes are never + held together. The register→flush→resolve happens-before argument is preserved — see the + updated "Deferred-commit design" section and the comment at `resset_add`. +2. **REQUIRED, false citations corrected** (this file + the `moe_submit` commit message, + rewritten pre-push): the original text attributed the hazard-tracking and thread-safety + statements to the SDK header (`MTLResidencySet.h`), which is in fact silent on both + topics. The statements come from Apple's **online** `MTLResidencySet` class reference and + the "Simplifying GPU resource management with residency sets" adoption guide (both + fetched 2026-07-18 during design). All attributions now name the actual source; where a + claim rests on design reasoning rather than documentation, it is labeled as such. +3. **REQUIRED, flush cost made harness-visible**: `g_t_resset_flush` + + `coli_metal_resset_stats()` + the gate-on-only `METAL-RESSET:` line in `profile_print` — + see "Instrumentation parity" above. +4. **Pre-existing fslab OOM-unwind bug — now CARRIED ON THIS BRANCH** (follow-up commit, + coordinator-sanctioned second `colibri.c` change): `expert_load`'s fslab OOM path + (`c/colibri.c`, in `expert_load_impl`) freed `s->slab` via `compat_aligned_free` **without** + `coli_metal_unregister` — on stock that leaves a stale `g_slabs` entry whose GPU + exposure ends with the last command buffer that declared it; under E5 the buffer would + additionally be a **permanent residency-set member** referencing freed host memory until + some later realloc of the same slot unregisters by pointer, a strictly longer-lived + exposure than stock's transient per-CB one. Fixed by porting E4's reference + implementation (`6753225`) to dev's non-heap code shape: `coli_metal_unregister(s->slab)` + before the free. The `uring_load_add` analog (E4's audit round-2 "cheap insurance") is + deliberately NOT carried: that arm is `#ifdef __linux__`-gated and `COLI_METAL` is + macOS-only, so it is dead code on every real build target, and unlike E4 this branch has + no allocation-path reason to touch the function at all. + +## Per-seam differences vs E4 + +| Seam | E4 (`e4/metal-heap`) | E5 (this branch) | +|---|---|---| +| Allocation | New: `MTLHeap` sub-buffers via `coli_metal_heap_alloc` | Unchanged: same `posix_memalign` + `newBufferWithBytesNoCopy` | +| Coordinator C source / `backend_metal.h` | `glm.c` touched (new alloc/free API, 4 call sites + `expert_host_release`) | `colibri.c` + header touched only for instrumentation and the OOM-unwind fix | +| Residency scope | Declared once **per command buffer** (`useHeap:`, still inside `moe_submit`) | Declared once **for the process** (queue-attached set), refreshed incrementally at register/unregister | +| Hazard tracking | Heap sub-buffers forced `MTLHazardTrackingModeUntracked` always (allocation-level) | Untouched at the resource level; `moe_submit` alone stops calling `useResource:` (encoder-level), independent of `COLI_METAL_UNTRACKED` | +| Per-buffer vs per-set skip | `[b heap]` (Metal's own `MTLResource.heap` property) checked per buffer — heterogeneous mixes possible if a slab fell back to malloc | Blanket `if (!g_resset_enabled)` — homogeneous by construction, since every registered slab goes through the same `coli_metal_register` path when the gate is on | +| Availability guard | None needed (`MTLHeap` is old API) | `@available(macOS 15.0, *)`, matching this box's macOS 26.5 but required for portability | +| Known regression | +12–13s expert-disk load at cap16 (suspected first-touch/lock contention on heap pages) | None expected — malloc pages never change ownership; **unverified without a run** | + +## What to measure (orchestrator, cap1/cap16, stock vs E4 vs E5) + +1. **GPU stall** (`coli_metal_moe_times` gpu/kernel breakdown) — success: E5 ≈ E4's + −85%-class reduction vs stock at cap16. +2. **Expert-disk load path** (existing load/service-time counters) — success: E5 ≈ stock, + i.e. **no** repeat of E4's +12–13s tax, since allocation is untouched. +3. **tok/s** — should track (1) and (2) together. +4. **md5 within a fixed dispatch composition** — flag on vs off must be byte-identical at a + given cap (the "Output-invariant by construction" hard constraint); flag-on vs flag-on + across cap1/cap16 may legitimately differ (different dispatch composition, per the + fix-plan's "Determinism side-finding"). +5. **`[METAL] residency-set: on` line present in stderr** at flag-on startup, and absent + (or the OS<15/create-failed fallback line) otherwise — cheap sanity check that a run + actually exercised the intended path before trusting its numbers. Also read the + **`METAL-RESSET: flush` line** (gate-on only): if that number is large, the deferred + set-commit cost is eating the stall win from the dispatch side. +6. If the hypothesis holds (E5 stall ≈ E4, E5 load-path ≈ stock, identical output), E5 becomes + the upstream PR candidate and must include the cap-default recalibration flagged in PR + #386's CURRENT-STATE CALIBRATION markers, per the spec's validation plan. + +## Build + +`cd c && make glm METAL=1` and a separate explicit `-Wall -Wextra` compile of +`backend_metal.mm` (the Makefile's `METALXX` line does not itself pass `-Wall -Wextra`, so +the warning surface was checked with those flags added explicitly; current `dev` contributes +one pre-existing `unused variable 'TG'` warning), plus +`cd c && make glm` (plain, non-Metal — the one `colibri.c` instrumentation touch, the `METAL-RESSET` stats line, +is inside the pre-existing `#ifdef COLI_METAL` arm of `profile_print`, so the plain build +compiles none of it), and +`make metal-test` (existing synthetic kernel-correctness unit test — no model, no +`glm52_i4/`, random weights — run once with `COLI_METAL_RESSET` unset and once with +`COLI_METAL_RESSET=1` to numerically exercise `coli_metal_register`/`moe_submit`'s changed +code path, since the task scope excludes running the real model). Exact results in the final +report, not here (build results belong to the report per the task's deliverable split, and +this file is written before the batched build run, per the scheduling constraint). + +## UNCERTAINTIES + +**Everything below is a judgment call, a seam where the residency-set lifecycle interacts +with the existing queue/command-buffer structure, or something unverifiable without a real +model run — flagged per the task's hard requirement.** + +1. **The central design risk: skipping `useResource:` in `moe_submit` gives up Metal's + automatic hazard tracking for that buffer set.** Sourcing (corrected in validator round + 1): the SDK header on this box + (`/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/.../Headers/MTLResidencySet.h`, + read directly) documents the protocol only in terms of residency and says nothing about + hazard tracking either way; the two operative statements are from Apple's **online** + documentation (fetched 2026-07-18): the "Simplifying GPU resource management with + residency sets" adoption guide — *"You don't need to call `useResource`/`useHeap`... for + allocations in a residency set"* — and the `MTLResidencySet` class reference — + *"Residency sets don't support hazard tracking, so you need to account for hazards with + fences and events."* I reasoned through every + code path that touches `moe_submit`'s `use` buffers (read-only, indirectly referenced, + never concurrently written, freed only after the engine's own slot lifecycle guarantees + no outstanding async reference) and concluded removing `useResource:` there specifically + is safe — but this reasoning is **not the same as having run the model**. If any code + path I didn't trace lets a slab get unregistered while an async `moe_block_begin` handle + is still in flight and reading it, this change removes a mitigation (weak as it may have + been) that existed before. **This is the #1 thing to watch for md5 divergence on**, and + the reason the scope was deliberately narrowed to `moe_submit` alone rather than applied + uniformly. +2. **Residency-set mutations are serialized under a dedicated `g_resset_mtx` (validator + round-1 fix — originally they ran under `g_slab_mtx`, the E4-regression shape; no Metal + call runs under the slab lock anymore).** The serialization itself is kept as required + for correctness: Apple's online `MTLResidencySet` class reference states the set's + *"methods aren't thread-safe"* (the SDK header contains no thread-safety statement either + way — citation corrected in round 1; the online doc is the source). What remains + **unverified without profiling a loaded run** is the *cost* of the calls themselves: + `resset_remove`'s synchronous `commit` runs inside `coli_metal_unregister` on the + loader path (its cost lands in the existing `t_ewait` accounting), and the SDK header + says commit on a resident set tries to make added/removed resources resident/non-resident + *"instantly"* — real synchronous work, since this set is resident from startup + (queue-attached for the process lifetime). If `commit()`/`addAllocation:` turn out + expensive on this hardware/OS build, the load path degrades through set bookkeeping + rather than mutex contention — a different, now-decoupled failure mode, but the same + symptom as E4's regression. Orchestrator: check E5's load-path timing against stock, not + just against E4, and read the new `METAL-RESSET: flush` line for the dispatch-side share. +3. **`resset_flush()`'s cost sits outside `g_t_setup`/the `moe_times` breakdown** (it runs + before `ts_start = mnow()`), by design, to keep the harness's existing counters + meaningful — and, since validator round 1, it is **no longer invisible**: the + `g_t_resset_flush` accumulator surfaces it as the gate-on-only `METAL-RESSET: flush` + line (see "Instrumentation parity"). Residual blind spots: (a) the accumulator is a + plain double written from `moe_submit` on the engine thread, matching the existing + `g_t_setup` convention — if `moe_submit` were ever called from multiple threads + concurrently, both counters would be equally wrong; (b) the register-side + `resset_add`/`resset_remove` costs have no dedicated counter and are only visible + blended into the existing `t_ewait`/disk-wait numbers (comment at `resset_add` says so) + — a fine-grained attribution would need a throwaway probe. +4. **`initialCapacity = 4096` on the `MTLResidencySetDescriptor` is an unverified guess.** + It's documented as a presize hint only (no correctness effect either way), chosen to be + "clearly larger than the permanent-weight-tensor + KV-cache + plausible cap16 LRU-slab + count" without actually counting those registrations precisely. Too small just means + internal array growth; not a correctness concern, flagged only because it's a number I + picked without measuring. +5. **Not calling `requestResidency()` proactively.** Apple's guide frames it as an optional + latency-hiding call ("call ahead of time during non-critical moments... to minimize [first + command buffer] latency"), and Blender's Cycles PR (the spec's cited reference + implementation) doesn't appear to use it either per its PR description. Omitted to keep + the lifecycle minimal and match the reference pattern; if profiling shows a + first-command-buffer-after-a-load-burst latency spike, this is the documented lever to try + next, not implemented here. +6. **The deferred-commit correctness argument (item in "Deferred-commit design" above) rests + on a single-writer-before-single-reader program-order guarantee that is true today by + inspection but is not an invariant enforced anywhere in code** (no assertion, no type-level + guarantee) — it's the same kind of implicit ordering `resolve()` itself already depends on + for correctness (a slab must be registered before any dispatch can resolve its pointer), + so this diff doesn't introduce a new category of fragility, but it's worth naming + explicitly rather than leaving implicit. +7. **Async `moe_block_begin`/`moe_block_end` overlap with concurrent `register()` calls** + (background loader threads registering new/different experts while an unrelated MoE block + is still in flight on the GPU) was reasoned through but never exercised in a real + concurrent stress scenario — the synthetic `metal-test` unit test's `run_moe` calls are + single-threaded and synchronous (`coli_metal_moe_block`, not the async `_begin`/`_end` + pair), so it does **not** cover this interleaving. The real engine's `PILOT`/prefetch and + `moe_block_begin`/`_end` overlap path is exactly the concurrency shape most likely to + expose a bug in this design if one exists, and is untested here by construction (out of + scope: no model runs). +8. **`coli_metal_gemm` (prefill path) and `bind_gemv` (attention path) still call + `useResource:` unconditionally, so they get no CPU-overhead benefit from the residency set + even though their buffers are also set members.** This is deliberate (see "Why only + `moe_submit` skips" above) but means E5's win, if any, is scoped to the decode-path MoE + dispatch loop specifically — prefill and attention timing should be unaffected by the flag, + which is itself a testable prediction the orchestrator's harness can check. +9. **API surface verified against this box's actual SDK headers** + (`MTLResidencySet.h`, `MTLDevice.h`, `MTLCommandQueue.h`, `MTLAllocation.h`, + `MTLResource.h` — all read directly, not from memory) and against Apple's own + "Simplifying GPU resource management with residency sets" guide, so the method names/ + signatures (`newResidencySetWithDescriptor:error:`, `addResidencySet:`, + `removeResidencySet:`, `addAllocation:`, `removeAllocation:`, `commit`) are + high-confidence. What is **not** independently verified is runtime behavior beyond what + the docs state and what the synthetic unit test exercises — no substitute for the + orchestrator's real cap-sweep battery. +10. **Pre-existing fslab OOM-unwind bug — carried on this branch** (follow-up commit; see + "Validator round 1 fixes" item 4 for the full mechanism). The one-line + unregister-before-free fix from E4's `6753225` is ported to dev's non-heap code shape, + so the upstream PR built from E5 inherits it automatically. Residual notes: (a) the fix + is only reachable through the fslab-OOM path (allocation failure mid-load), so it is + untestable without an OOM-injection harness and cannot affect the orchestrator's + controlled A/B runs at sane RAM headroom — carried as correctness insurance, verified by + inspection + clean builds only; (b) the `__linux__`-gated `uring_load_add` analog is + deliberately not carried (dead code on every real build target — rationale in the fixes + section). diff --git a/c/Makefile b/c/Makefile index bb49499..4ad515f 100644 --- a/c/Makefile +++ b/c/Makefile @@ -58,9 +58,14 @@ CFLAGS = -O3 $(OMPC) -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indent # Opt-in: ARCH=native appends -mcpu=native (arm64 clang uses -mcpu, not -march), # which unlocks the i8mm SMMLA int8/int4 dot kernels in colibri.c. ARCH unset -> # no -mcpu, default build byte-identical. Apple clang knows apple-m4 / native. +# For older X86_64 Macs (for example, Mac Pro 2019) we need to use -march ifneq ($(ARCH),) +ifneq (,$(X86_64)) +CFLAGS += -march=$(ARCH) +else CFLAGS += -mcpu=$(ARCH) endif +endif LDFLAGS = -lm $(OMPL) EXE = else ifneq ($(IS_WIN),) @@ -166,7 +171,7 @@ else PYTHON ?= python3 endif CUDA_OBJ = -TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_mirror$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_kv_alloc$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE) +TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_mirror$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_tok_o200k$(EXE) tests/test_kv_alloc$(EXE) tests/test_int3$(EXE) tests/test_int3_load$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE) tests/test_pipe_block$(EXE) ifneq (,$(LINUX)) TEST_BINS += tests/test_uring$(EXE) endif @@ -273,8 +278,9 @@ olmoe$(EXE): olmoe.c st.h json.h compat.h # Use a baseline that matches the compiler target. macOS already targets a # portable baseline when ARCH is empty; forcing the x86 value there breaks # Apple Silicon. Unknown targets use native rather than an invalid x86 flag. +# Intel Macs need -march for vector instructions ifneq (,$(DARWIN)) -PORTABLE_ARCH = +PORTABLE_ARCH = $(if $(X86_64),x86-64-v3,) else ifneq (,$(AARCH64)) PORTABLE_ARCH = armv8-a else ifneq (,$(PPC64)) @@ -294,6 +300,8 @@ iobench$(EXE): iobench.c compat.h tests/test_json$(EXE): tests/test_json.c json.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) +tests/test_tok_o200k$(EXE): tests/test_tok_o200k.c tok.h tok_unicode.h tok_unicode_o200k.h json.h + $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) tests/test_st_pread$(EXE): tests/test_st_pread.c st.h json.h compat.h $(CC) $(CFLAGS) -DST_PREAD_CHUNK=7 $< -o $@ $(LDFLAGS) @@ -338,6 +346,12 @@ tests/test_sample_nan$(EXE): tests/test_sample_nan.c colibri.c st.h uring.h json tests/test_kv_alloc$(EXE): tests/test_kv_alloc.c colibri.c st.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) +tests/test_int3$(EXE): tests/test_int3.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h + $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) + +tests/test_int3_load$(EXE): tests/test_int3_load.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h + $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) + tests/test_logit_nan$(EXE): tests/test_logit_nan.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) @@ -355,9 +369,17 @@ tests/test_dsa_select$(EXE): tests/test_dsa_select.c colibri.c st.h uring.h json tests/bench_dsa_select$(EXE): tests/bench_dsa_select.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) +# bench_idot: microbenchmark (single-acc vs independent-acc AVX-VNNI idot), NOT a test gate. +# Build on demand on an AVX-VNNI CPU: make tests/bench_idot ARCH=native +tests/bench_idot$(EXE): tests/bench_idot.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h + $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) + tests/test_uring$(EXE): tests/test_uring.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) +tests/test_pipe_block$(EXE): tests/test_pipe_block.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h + $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) + test-c: $(TEST_BINS) $(PYTHON) tools/run_tests.py $(TEST_BINS) @@ -366,6 +388,29 @@ test-python: test: test-c test-python +# --- Efficiency / regression suite (issue: "test the program for inefficiencies") --- +# The tiny-model assertions live in test_inefficiency.py and run as part of +# test-python (they're discovered by the test_*.py glob). These targets are +# convenience entry points; the opt-in full-model report is NEVER in `make test`. +# +# make efficiency tiny-model asserted regression tests (CPU; part of test-python) +# make efficiency-cuda the CUDA-path tests (requires a CUDA build — see below) +# make efficiency-report opt-in full-model 🟢/🔴 diagnostic, never fails CI +# +# CUDA build (Windows): the CUDA tests need a host built with -DCOLI_CUDA plus +# the runtime DLL. Do this FIRST — note CUDA_DLL=1 on BOTH the host and the +# rule below, or `make glm.exe` will rebuild a CPU-only host and overwrite it: +# make clean && make glm.exe CUDA_DLL=1 && make cuda-dll +# The tests auto-skip with a clear message if the host is CPU-only. +efficiency: test-python + $(PYTHON) -m unittest tests.test_inefficiency -v + +efficiency-cuda: + $(PYTHON) -m unittest tests.test_inefficiency.TinyCudaEfficiencyTest -v + +efficiency-report: + $(PYTHON) tests/test_efficiency_report.py + # Local validation: one portable CPU build and dependency-free tests. check: $(MAKE) clean diff --git a/c/backend_cuda.cu b/c/backend_cuda.cu index 413e53b..327915e 100644 --- a/c/backend_cuda.cu +++ b/c/backend_cuda.cu @@ -20,6 +20,9 @@ struct ColiCudaTensor { float *scales; size_t weight_bytes; int fmt, I, O, device; + int gs; /* quant group size; 0 = per-row scales (#334) */ + int ng; /* number of scale groups per row = ceil(I/gs) for fmt=4 */ + size_t scale_count; /* floats in `scales`: O per-row, O*ng grouped */ int tracked; RaggedKVEntry ragged[512]; int ragged_count; @@ -34,15 +37,17 @@ typedef struct { size_t qx_cap, qscale_cap; float *host_x,*host_y,*host_kv; size_t host_x_cap,host_y_cap,host_kv_cap; float *aq,*al,*ar,*ac; size_t aq_cap,al_cap,ar_cap,ac_cap; - float *pipe_buf[24]; size_t pipe_cap[24]; /* scratch persistenti del resident pipeline */ + float *pipe_buf[27]; size_t pipe_cap[27]; /* scratch persistenti del resident pipeline */ cudaStream_t stream; void *group_desc; size_t group_desc_cap; size_t tensor_count, tensor_bytes; + int group_pending; size_t group_pending_bytes; /* async expert-group in flight (Inc.4) */ } DeviceContext; typedef struct { const void *g,*u,*d; const float *gs,*us,*ds; int gf,uf,df,rows,offset; + int ggs,ugs,dgs; /* per-tensor quant group size; 0 = per-row scales (#334 fmt=4) */ } GroupDesc; static DeviceContext g_ctx[COLI_CUDA_MAX_DEVICES]; @@ -54,6 +59,8 @@ static std::mutex g_group_stats_mu; static int cuda_ok(cudaError_t err, const char *what) { if (err == cudaSuccess) return 1; std::fprintf(stderr, "[CUDA] %s: %s\n", what, cudaGetErrorString(err)); + (void)cudaGetLastError(); /* consume the sticky error: a failed call must + not poison the next launch's error check */ return 0; } @@ -79,8 +86,9 @@ static int select_ctx(DeviceContext *ctx) { __host__ __device__ static size_t row_bytes(int fmt, int I) { if (fmt == 0) return (size_t)I * sizeof(float); if (fmt == 1) return (size_t)I; - if (fmt == 2) return (size_t)(I + 1) / 2; + if (fmt == 2 || fmt == 4) return (size_t)(I + 1) / 2; /* fmt=4: same packed int4 */ if (fmt == 3) return (size_t)(I + 3) / 4; + if (fmt == 4) return (size_t)(I + 1) / 2; /* grouped int4: nibbles like fmt 2 */ return 0; } @@ -89,7 +97,7 @@ __device__ static float weight_at(const void *weights, int fmt, size_t row, int if (fmt == 0) return reinterpret_cast(base)[i]; if (fmt == 1) return static_cast(reinterpret_cast(base)[i]); const uint8_t *q = base; - if (fmt == 2) { + if (fmt == 2 || fmt == 4) { /* fmt=4: same nibble layout */ uint8_t v = q[i >> 1]; int n=(i&1)?(v>>4):(v&15); return static_cast(n&8?n-16:n); } @@ -97,20 +105,45 @@ __device__ static float weight_at(const void *weights, int fmt, size_t row, int return static_cast(((v >> ((i & 3) * 2)) & 3) - 2); } +/* Scale for output `row`, input element `k`. fmt=4 (grouped int4) stores ng + * scales per row at scales[row*ng + k/gs]; every other quantized format has + * one scale per row at scales[row]. Mirrors quant_matmul's fmt==4 branch so the + * attention absorb kernels apply per-group scales instead of the per-row + * (fmt=2) semantic that crashed #298's g64 kv_b. */ +__device__ static float absorb_scale(const float *wscale, int fmt, int gs, int ng, int row, int k) { + if (!fmt) return 1.f; + if (fmt != 4) return wscale[row]; + int g = k / gs; if (g >= ng) g = ng - 1; /* tail of the last (partial) group */ + return wscale[(size_t)row * ng + g]; +} + __global__ static void offset_to_signed_s4(uint8_t *q,size_t n){ size_t i=(size_t)blockIdx.x*blockDim.x+threadIdx.x;if(i= ng) g = ng - 1; /* tail elements in the last (partial) group */ + sum += xs[i] * weight_at(weights, fmt, row, i) * scl[g]; + } + } else { + for (int i = threadIdx.x; i < I; i += blockDim.x) + sum += xs[i] * weight_at(weights, fmt, row, i); + } __shared__ float partial[256]; partial[threadIdx.x] = sum; @@ -120,7 +153,7 @@ __global__ static void quant_matmul(float *y, const float *x, const void *weight __syncthreads(); } if (!threadIdx.x) - y[(size_t)s * O + o] = partial[0] * (fmt ? scales[o] : 1.0f); + y[(size_t)s * O + o] = (fmt && fmt != 4) ? partial[0] * scales[o] : partial[0]; } __global__ static void silu_mul(float *gate, const float *up, size_t n) { @@ -296,13 +329,50 @@ __global__ static void grouped_down_w4(float *y,const float *x,const GroupDesc * if(!threadIdx.x)y[(size_t)(d.offset+s)*D+o]=p[0]*d.ds[o]; } +/* fmt=4 grouped-int4 variants (#334): identical structure to the w4 kernels, + * but the scale varies along the input dimension — sc[o*ng + i/gs], applied + * per element inside the accumulation (gs is even, so a packed byte never + * straddles a group). gs<=0 degrades to per-row (ng=1), so mixed fmt2/fmt4 + * groups run correctly through this one kernel family. */ +__global__ static void grouped_hidden_g4_dual(float *gate,float *up,const float *x, + const GroupDesc *desc,int I,int D){ + int o=blockIdx.x,s=blockIdx.y,c=blockIdx.z;GroupDesc d=desc[c];if(s>=d.rows)return; + const uint8_t *gr=(const uint8_t*)d.g+(size_t)o*((D+1)/2); + const uint8_t *ur=(const uint8_t*)d.u+(size_t)o*((D+1)/2); + int ggs=d.ggs>0?d.ggs:D, ugs=d.ugs>0?d.ugs:D; + const float *gsc=d.gs+(size_t)o*(size_t)((D+ggs-1)/ggs); + const float *usc=d.us+(size_t)o*(size_t)((D+ugs-1)/ugs); + const float *xs=x+(size_t)(d.offset+s)*D;float ga=0,ua=0; + for(int b=threadIdx.x;b<(D+1)/2;b+=blockDim.x){float g0,g1,u0,u1;unpack_s4(gr[b],&g0,&g1);unpack_s4(ur[b],&u0,&u1); + int i=b*2;float gv=gsc[i/ggs],uv=usc[i/ugs]; + ga+=xs[i]*g0*gv;ua+=xs[i]*u0*uv; + if(i+1>=1){if(threadIdx.x=d.rows)return; + const uint8_t *row=(const uint8_t*)d.d+(size_t)o*((I+1)/2); + int dgs=d.dgs>0?d.dgs:I; + const float *dsc=d.ds+(size_t)o*(size_t)((I+dgs-1)/dgs); + const float *xs=x+(size_t)(d.offset+s)*I;float sum=0; + for(int b=threadIdx.x;b<(I+1)/2;b+=blockDim.x){float a,z;unpack_s4(row[b],&a,&z); + int i=b*2;float sv=dsc[i/dgs]; + sum+=xs[i]*a*sv;if(i+1>=1){if(threadIdx.x=S||nt<1)return; extern __shared__ float sm[];float *qa=sm,*cl=qa+K,*scores=cl+K,*red=scores+T; const float *qs=q+((size_t)s*H+h)*(Q+R); for(int k=tid;kqx) cudaFree(ctx->qx); if (ctx->qscale) cudaFree(ctx->qscale); if(ctx->aq)cudaFree(ctx->aq);if(ctx->al)cudaFree(ctx->al);if(ctx->ar)cudaFree(ctx->ar);if(ctx->ac)cudaFree(ctx->ac); - for(int b=0;b<24;b++) if(ctx->pipe_buf[b]) cudaFree(ctx->pipe_buf[b]); + for(int b=0;b<27;b++) if(ctx->pipe_buf[b]) cudaFree(ctx->pipe_buf[b]); if (ctx->host_x) cudaFreeHost(ctx->host_x); if (ctx->host_y) cudaFreeHost(ctx->host_y); if (ctx->host_kv) cudaFreeHost(ctx->host_kv); @@ -503,40 +574,66 @@ extern "C" void coli_cuda_group_stats(uint64_t *calls, uint64_t *experts, uint64 if(d2h_ms) *d2h_ms=g_group_d2h_ms; } +/* group size for the NEXT upload on this thread (fmt=4): routed through a + * thread_local so the widely-wired upload signature (and the Windows DLL ABI) + * stays untouched. pin_load uploads in parallel, hence thread_local. */ +static thread_local int g_upload_gs = 0; +extern "C" int coli_cuda_tensor_upload_g(ColiCudaTensor **tensor, + const void *weights, const float *scales, + int fmt, int I, int O, int device, int gs); extern "C" int coli_cuda_tensor_upload(ColiCudaTensor **tensor, const void *weights, const float *scales, int fmt, int I, int O, int device) { + if (!tensor) return 0; + if (*tensor) { + /* Cached device copy: usable even when the caller's host pointers are + * gone. CUDA_RELEASE_HOST slots null their host pointers after upload, + * and with the old order (!weights checked first) every later matmul + * on such a slot failed here — the GPU tier silently never computed + * for host-released slab experts. */ + ColiCudaTensor *t = *tensor; + int want_gs = (fmt==4 && g_upload_gs>0) ? g_upload_gs : 0; + return t->fmt == fmt && t->I == I && t->O == O && t->device == device && t->gs == want_gs; + } DeviceContext *ctx = find_ctx(device); - if (!tensor || !weights || I < 1 || O < 1 || !select_ctx(ctx)) return 0; + if (!weights || I < 1 || O < 1 || !select_ctx(ctx)) return 0; size_t rb = row_bytes(fmt, I); if (!rb || (fmt && !scales)) return 0; - if (*tensor) { - ColiCudaTensor *t = *tensor; - return t->fmt == fmt && t->I == I && t->O == O && t->device == device; - } ColiCudaTensor *t = static_cast(std::calloc(1, sizeof(*t))); if (!t) return 0; t->fmt = fmt; t->I = I; t->O = O; t->device = device; t->weight_bytes = rb * (size_t)O; + t->gs = (fmt==4 && g_upload_gs>0) ? g_upload_gs : 0; + t->ng = t->gs ? (I + t->gs - 1) / t->gs : 1; + t->scale_count = t->gs ? (size_t)O * (size_t)t->ng : (size_t)O; if (!cuda_ok(cudaMalloc(&t->weights, t->weight_bytes), "tensor allocation") || !cuda_ok(cudaMemcpy(t->weights, weights, t->weight_bytes, cudaMemcpyHostToDevice), "tensor upload")) { coli_cuda_tensor_free(t); return 0; } - if(fmt==2){offset_to_signed_s4<<<(unsigned)((t->weight_bytes+255)/256),256>>>((uint8_t*)t->weights,t->weight_bytes); + if(fmt==2||fmt==4){ /* same nibble layout: offset-binary -> signed in place */ + offset_to_signed_s4<<<(unsigned)((t->weight_bytes+255)/256),256>>>((uint8_t*)t->weights,t->weight_bytes); if(!cuda_ok(cudaGetLastError(),"int4 weight conversion")){coli_cuda_tensor_free(t);return 0;}} if (fmt) { - if (!cuda_ok(cudaMalloc(&t->scales, (size_t)O * sizeof(float)), "scale allocation") || - !cuda_ok(cudaMemcpy(t->scales, scales, (size_t)O * sizeof(float), cudaMemcpyHostToDevice), "scale upload")) { + if (!cuda_ok(cudaMalloc(&t->scales, t->scale_count * sizeof(float)), "scale allocation") || + !cuda_ok(cudaMemcpy(t->scales, scales, t->scale_count * sizeof(float), cudaMemcpyHostToDevice), "scale upload")) { coli_cuda_tensor_free(t); return 0; } } t->tracked = 1; ctx->tensor_count++; - ctx->tensor_bytes += t->weight_bytes + (fmt ? (size_t)O * sizeof(float) : 0); + ctx->tensor_bytes += t->weight_bytes + (fmt ? t->scale_count * sizeof(float) : 0); *tensor = t; return 1; } +extern "C" int coli_cuda_tensor_upload_g(ColiCudaTensor **tensor, + const void *weights, const float *scales, + int fmt, int I, int O, int device, int gs){ + g_upload_gs = gs>0 ? gs : 0; + int r = coli_cuda_tensor_upload(tensor, weights, scales, fmt, I, O, device); + g_upload_gs = 0; + return r; +} extern "C" int coli_cuda_tensor_update(ColiCudaTensor *tensor, const void *weights, @@ -546,20 +643,35 @@ extern "C" int coli_cuda_tensor_update(ColiCudaTensor *tensor, if (!select_ctx(ctx)) return 0; if (!cuda_ok(cudaMemcpy(tensor->weights,weights,tensor->weight_bytes, cudaMemcpyHostToDevice),"tensor refresh")) return 0; - if(tensor->fmt==2){ + if(tensor->fmt==2||tensor->fmt==4){ offset_to_signed_s4<<<(unsigned)((tensor->weight_bytes+255)/256),256>>>( (uint8_t*)tensor->weights,tensor->weight_bytes); if(!cuda_ok(cudaGetLastError(),"int4 weight refresh")) return 0; } + int ng = tensor->ng > 0 ? tensor->ng : 1; return !tensor->fmt || cuda_ok(cudaMemcpy(tensor->scales,scales, - (size_t)tensor->O*sizeof(float),cudaMemcpyHostToDevice),"scale refresh"); + (tensor->scale_count?tensor->scale_count:(size_t)tensor->O)*sizeof(float), + cudaMemcpyHostToDevice),"scale refresh"); +} + +/* Test hook: COLI_GPU_FAIL_AFTER=N makes every GPU COMPUTE entry point report + * failure after N successful calls (N=0: every call fails), exercising the + * engine's CPU fallbacks and host-rematerialization end-to-end without real + * hardware faults. Uploads/queries are not gated. Unset: no effect. */ +static long g_gpu_calls; +static int fault_injected(void) { + const char *fa = std::getenv("COLI_GPU_FAIL_AFTER"); + return fa && g_gpu_calls++ >= std::atol(fa); } extern "C" int coli_cuda_matmul(ColiCudaTensor **tensor, float *y, const float *x, const void *weights, const float *scales, - int fmt, int S, int I, int O, int device) { - if (S < 1 || !coli_cuda_tensor_upload(tensor, weights, scales, fmt, I, O, device)) return 0; + int fmt, int S, int I, int O, int device, int gs) { + if (fault_injected()) return 0; + if (S < 1) return 0; + if (gs > 0) { if (!coli_cuda_tensor_upload_g(tensor, weights, scales, fmt, I, O, device, gs)) return 0; } + else { if (!coli_cuda_tensor_upload(tensor, weights, scales, fmt, I, O, device)) return 0; } ColiCudaTensor *t = *tensor; DeviceContext *ctx = find_ctx(t->device); if (!select_ctx(ctx)) return 0; @@ -568,7 +680,7 @@ extern "C" int coli_cuda_matmul(ColiCudaTensor **tensor, if (!reserve(&ctx->x, &ctx->x_cap, xb) || !reserve(&ctx->y, &ctx->y_cap, yb)) return 0; if (!cuda_ok(cudaMemcpy(ctx->x, x, xb, cudaMemcpyHostToDevice), "input upload")) return 0; dim3 grid((unsigned)O, (unsigned)S); - quant_matmul<<>>(ctx->y, ctx->x, t->weights, t->scales, fmt, S, I, O, rb); + quant_matmul<<>>(ctx->y, ctx->x, t->weights, t->scales, fmt, S, I, O, rb, t->gs, t->ng); if (!cuda_ok(cudaGetLastError(), "matmul launch") || !cuda_ok(cudaMemcpy(y, ctx->y, yb, cudaMemcpyDeviceToHost), "output download")) return 0; return 1; @@ -577,6 +689,7 @@ extern "C" int coli_cuda_matmul(ColiCudaTensor **tensor, extern "C" int coli_cuda_expert_mlp(ColiCudaTensor *gate, ColiCudaTensor *up, ColiCudaTensor *down, float *y, const float *x, int S) { + if (fault_injected()) return 0; if (!gate || !up || !down || !x || !y || S < 1 || gate->device != up->device || gate->device != down->device || gate->I != up->I || gate->O != up->O || @@ -591,13 +704,13 @@ extern "C" int coli_cuda_expert_mlp(ColiCudaTensor *gate, ColiCudaTensor *up, if (!cuda_ok(cudaMemcpy(ctx->x,x,xb,cudaMemcpyHostToDevice),"expert input upload")) return 0; dim3 hidden_grid((unsigned)I,(unsigned)S), output_grid((unsigned)D,(unsigned)S); quant_matmul<<>>(ctx->gate,ctx->x,gate->weights,gate->scales, - gate->fmt,S,D,I,row_bytes(gate->fmt,D)); + gate->fmt,S,D,I,row_bytes(gate->fmt,D),gate->gs,gate->ng); quant_matmul<<>>(ctx->up,ctx->x,up->weights,up->scales, - up->fmt,S,D,I,row_bytes(up->fmt,D)); + up->fmt,S,D,I,row_bytes(up->fmt,D),up->gs,up->ng); size_t n=(size_t)S*I; silu_mul<<<(unsigned)((n+255)/256),256>>>(ctx->gate,ctx->up,n); quant_matmul<<>>(ctx->y,ctx->gate,down->weights,down->scales, - down->fmt,S,I,D,row_bytes(down->fmt,I)); + down->fmt,S,I,D,row_bytes(down->fmt,I),down->gs,down->ng); if (!cuda_ok(cudaGetLastError(),"expert MLP launch") || !cuda_ok(cudaMemcpy(y,ctx->y,yb,cudaMemcpyDeviceToHost),"expert output download")) return 0; return 1; @@ -605,6 +718,7 @@ extern "C" int coli_cuda_expert_mlp(ColiCudaTensor *gate, ColiCudaTensor *up, extern "C" int coli_cuda_shared_mlp_w4a16(ColiCudaTensor *gate,ColiCudaTensor *up, ColiCudaTensor *down,float *y,const float *x,int S){ + if (fault_injected()) return 0; if(!gate||!up||!down||!x||!y||S<1||gate->fmt!=2||up->fmt!=2||down->fmt!=2|| gate->device!=up->device||gate->device!=down->device||gate->I!=up->I|| gate->O!=up->O||down->I!=gate->O||down->O!=gate->I)return 0; @@ -636,19 +750,24 @@ extern "C" int coli_cuda_expert_group(ColiCudaTensor *const *gates, ColiCudaTensor *const *downs, const int *rows, int count, float *y, const float *x) { + if (fault_injected()) return 0; if (!gates || !ups || !downs || !rows || !x || !y || count < 1) return 0; ColiCudaTensor *first=gates[0]; if (!first) return 0; int device=first->device,D=first->I,I=first->O,total=0,max_rows=0; GroupDesc host[64]; if(count>64) return 0; - int all_s4=1; + int all_s4=1,all_q4=1,any_g4=0; for(int c=0;cdevice!=device||u->device!=device||d->device!=device|| g->I!=D||u->I!=D||g->O!=I||u->O!=I||d->I!=I||d->O!=D) return 0; host[c]={g->weights,u->weights,d->weights,g->scales,u->scales,d->scales, - g->fmt,u->fmt,d->fmt,rows[c],total}; + g->fmt,u->fmt,d->fmt,rows[c],total, + g->gs,u->gs,d->gs}; all_s4&=g->fmt==2&&u->fmt==2&&d->fmt==2; + all_q4&=(g->fmt==2||g->fmt==4)&&(u->fmt==2||u->fmt==4)&&(d->fmt==2||d->fmt==4)&& + !(g->gs&1)&&!(u->gs&1)&&!(d->gs&1); /* even gs: a packed byte never straddles groups */ + any_g4|=g->fmt==4||u->fmt==4||d->fmt==4; total+=rows[c]; if(rows[c]>max_rows) max_rows=rows[c]; } DeviceContext *ctx=find_ctx(device); if(!select_ctx(ctx)) return 0; @@ -689,7 +808,15 @@ extern "C" int coli_cuda_expert_group(ColiCudaTensor *const *gates, quantize_s4_rows<<stream>>>(ctx->qx,ctx->qscale,ctx->gate,total,I); grouped_s4_wmma<<stream>>>(ctx->y,ctx->qx,ctx->qscale,dev,I,D,2); }else if(all_s4&&ctx->compute_major>=7&&getenv("COLI_CUDA_TC_W4A16")&& - atoi(getenv("COLI_CUDA_TC_W4A16"))){ + atoi(getenv("COLI_CUDA_TC_W4A16"))&& + [&]{ int tc16_min=getenv("COLI_CUDA_TC_W4A16_MIN")?atoi(getenv("COLI_CUDA_TC_W4A16_MIN")):16; + for(int c=0;c=tc16_min) return 1; + return 0; }()){ + /* At least one expert has enough rows for a Tensor Core tile. Groups + * where EVERY expert is below the threshold (decode: r=1) fall through + * to the grouped-W4 path below — 3 launches for the whole group instead + * of 4 per expert (#431: the launch flood measured at ~981 micro-kernels + * per token came from decode riding this branch's per-expert fallback). */ /* W4A16 Tensor Core per gruppo: attivazioni fp16 per tile (lossless al * contrario del path W4A4), un lancio per expert dentro lo stream — * l'overhead di lancio e' trascurabile rispetto ai GEMM. */ @@ -711,12 +838,12 @@ extern "C" int coli_cuda_expert_group(ColiCudaTensor *const *gates, /* piccoli batch: tile TC quasi vuoti + overhead di lancio — il * kernel naive per-elemento resta piu' veloce (misurato in decode) */ quant_matmul<<stream>>>(g16,x16, - host[c].g,host[c].gs,host[c].gf,r,D,I,row_bytes(host[c].gf,D)); + host[c].g,host[c].gs,host[c].gf,r,D,I,row_bytes(host[c].gf,D),0,1); quant_matmul<<stream>>>(u16,x16, - host[c].u,host[c].us,host[c].uf,r,D,I,row_bytes(host[c].uf,D)); + host[c].u,host[c].us,host[c].uf,r,D,I,row_bytes(host[c].uf,D),0,1); silu_mul<<<(unsigned)(((size_t)r*I+255)/256),256,0,ctx->stream>>>(g16,u16,(size_t)r*I); quant_matmul<<stream>>>(y16,g16, - host[c].d,host[c].ds,host[c].df,r,I,D,row_bytes(host[c].df,I)); + host[c].d,host[c].ds,host[c].df,r,I,D,row_bytes(host[c].df,I),0,1); } off16+=r; } @@ -730,7 +857,18 @@ extern "C" int coli_cuda_expert_group(ColiCudaTensor *const *gates, } silu_mul<<<(unsigned)(((size_t)total*I+255)/256),256,0,ctx->stream>>>(ctx->gate,ctx->up,(size_t)total*I); grouped_down_w4<<stream>>>(ctx->y,ctx->gate,dev,D,I); + }else if(all_q4&&any_g4){ + /* grouped-int4 (fmt=4) present: per-group scales (#334). fmt=2 members + * ride along as the ng=1 special case. */ + dim3 hg((unsigned)I,(unsigned)max_rows,(unsigned)count),og((unsigned)D,(unsigned)max_rows,(unsigned)count); + grouped_hidden_g4_dual<<stream>>>(ctx->gate,ctx->up,ctx->x,dev,I,D); + silu_mul<<<(unsigned)(((size_t)total*I+255)/256),256,0,ctx->stream>>>(ctx->gate,ctx->up,(size_t)total*I); + grouped_down_g4<<stream>>>(ctx->y,ctx->gate,dev,D,I); }else{ + /* generic path decodes fmt 0/1/2/3 only — a fmt=4 group that slipped the + * gates above (odd gs) must NOT be silently decoded as int2 (#334). */ + for(int c=0;cstream>>>(ctx->gate,ctx->x,dev,I,D,0); grouped_hidden<<stream>>>(ctx->up,ctx->x,dev,I,D,1); @@ -757,10 +895,79 @@ extern "C" int coli_cuda_expert_group(ColiCudaTensor *const *gates, return 1; } +/* ---- Async expert group (Inc.4): issue/take split of coli_cuda_expert_group ---- + * The measured cost of the sync call at decode is ~0.45 ms/call of HOST-side wait + * (stream sync + staging), vs ~0.18 ms of actual GPU work — 70% tax, paid ~5x per + * layer because a token's 8 experts scatter across devices. issue() stages and + * launches on the device stream and returns immediately; take() syncs and hands + * back the pinned result rows. One issue may be outstanding per device; moe() + * takes at each layer end, which also orders the next layer's reuse of the ctx + * scratch buffers. Small batches only (decode/spec): bigger totals keep the sync + * path with its TC variants. Numerics are the sync path's small-batch kernels, + * so greedy output is byte-identical by construction. */ +extern "C" int coli_cuda_expert_group_issue(ColiCudaTensor *const *gates, + ColiCudaTensor *const *ups, + ColiCudaTensor *const *downs, + const int *rows, int count, + const float *x) { + if (!gates || !ups || !downs || !rows || !x || count < 1 || count > 64) return 0; + ColiCudaTensor *first=gates[0]; + if (!first) return 0; + int device=first->device,D=first->I,I=first->O,total=0; + GroupDesc host[64]; + for(int c=0;cdevice!=device||u->device!=device||d->device!=device|| + g->I!=D||u->I!=D||g->O!=I||u->O!=I||d->I!=I||d->O!=D) return 0; + host[c]={g->weights,u->weights,d->weights,g->scales,u->scales,d->scales, + g->fmt,u->fmt,d->fmt,rows[c],total}; + total+=rows[c]; + } + if(total>8) return 0; /* decode-scale only */ + DeviceContext *ctx=find_ctx(device); if(!ctx||ctx->group_pending||!select_ctx(ctx)) return 0; + size_t xb=(size_t)total*D*sizeof(float), ib=(size_t)total*I*sizeof(float); + if(!reserve(&ctx->x,&ctx->x_cap,xb)||!reserve(&ctx->y,&ctx->y_cap,xb)|| + !reserve(&ctx->gate,&ctx->gate_cap,ib)||!reserve(&ctx->up,&ctx->up_cap,ib)|| + !reserve_pinned(&ctx->host_x,&ctx->host_x_cap,xb)|| + !reserve_pinned(&ctx->host_y,&ctx->host_y_cap,xb)) return 0; + std::memcpy(ctx->host_x,x,xb); + if(!cuda_ok(cudaMemcpyAsync(ctx->x,ctx->host_x,xb,cudaMemcpyHostToDevice,ctx->stream), + "expert group issue upload")) return 0; + for(int c=0;cgate+(size_t)host[c].offset*I,*u16=ctx->up+(size_t)host[c].offset*I; + float *x16=ctx->x+(size_t)host[c].offset*D,*y16=ctx->y+(size_t)host[c].offset*D; + quant_matmul<<stream>>>(g16,x16, + host[c].g,host[c].gs,host[c].gf,r,D,I,row_bytes(host[c].gf,D),0,1); + quant_matmul<<stream>>>(u16,x16, + host[c].u,host[c].us,host[c].uf,r,D,I,row_bytes(host[c].uf,D),0,1); + silu_mul<<<(unsigned)(((size_t)r*I+255)/256),256,0,ctx->stream>>>(g16,u16,(size_t)r*I); + quant_matmul<<stream>>>(y16,g16, + host[c].d,host[c].ds,host[c].df,r,I,D,row_bytes(host[c].df,I),0,1); + } + if(!cuda_ok(cudaGetLastError(),"expert group issue launch")|| + !cuda_ok(cudaMemcpyAsync(ctx->host_y,ctx->y,xb,cudaMemcpyDeviceToHost,ctx->stream), + "expert group issue download")) return 0; + ctx->group_pending=1; ctx->group_pending_bytes=xb; + { std::lock_guard lock(g_group_stats_mu); + g_group_calls++; g_group_experts+=(uint64_t)count; g_group_rows+=(uint64_t)total; } + return 1; +} + +extern "C" const float *coli_cuda_expert_group_take(int device) { + DeviceContext *ctx=find_ctx(device); + if(!ctx||!ctx->group_pending) return nullptr; + ctx->group_pending=0; + if(!select_ctx(ctx)) return nullptr; + if(!cuda_ok(cudaStreamSynchronize(ctx->stream),"expert group take")) return nullptr; + return ctx->host_y; +} + extern "C" int coli_cuda_attention_absorb(ColiCudaTensor *w,float *ctx,const float *q, const float *latent,const float *rope,int H,int Q, int R,int V,int K,int T,float scale){ + if (fault_injected()) return 0; if(!w||!ctx||!q||!latent||!rope||H<1||Q<1||R<1||V<1||K<1||K>512||T<1||T>4096|| w->I!=K||w->O!=H*(Q+V))return 0; DeviceContext *dc=find_ctx(w->device);if(!select_ctx(dc))return 0; @@ -773,7 +980,7 @@ extern "C" int coli_cuda_attention_absorb(ColiCudaTensor *w,float *ctx,const flo !cuda_ok(cudaMemcpyAsync(dc->ar,rope,rb,cudaMemcpyHostToDevice,dc->stream),"attention rope upload"))return 0; size_t shared=(size_t)(2*K+T)*sizeof(float); attention_absorb_kernel<<stream>>>(dc->ac,dc->aq,dc->al,dc->ar,w->weights,w->scales, - w->fmt,H,Q,R,V,K,T,scale); + w->fmt,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"attention absorb launch")|| !cuda_ok(cudaMemcpyAsync(ctx,dc->ac,cb,cudaMemcpyDeviceToHost,dc->stream),"attention context download")|| !cuda_ok(cudaStreamSynchronize(dc->stream),"attention synchronize"))return 0; @@ -796,13 +1003,13 @@ static int attention_absorb_batch_run(ColiCudaTensor *w,ColiCudaTensor *proj,flo !cuda_ok(cudaMemcpyAsync(dc->ar,rope,rb,cudaMemcpyHostToDevice,dc->stream),"attention batch rope upload"))return 0; size_t shared=(size_t)(2*K+T+256)*sizeof(float); attention_absorb_batch_kernel<<stream>>>(dc->ac,dc->aq,dc->al, - dc->ar,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale); + dc->ar,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"attention batch launch"))return 0; const float *src=dc->ac;size_t ob=cb; if(proj){ ob=(size_t)S*proj->O*sizeof(float);if(!reserve(&dc->y,&dc->y_cap,ob))return 0; quant_matmul<<O,S),256,0,dc->stream>>>(dc->y,dc->ac,proj->weights, - proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I)); + proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I),proj->gs,proj->ng); if(!cuda_ok(cudaGetLastError(),"attention o_proj launch"))return 0;src=dc->y; } if(!cuda_ok(cudaMemcpyAsync(out,src,ob,cudaMemcpyDeviceToHost,dc->stream), @@ -814,12 +1021,14 @@ static int attention_absorb_batch_run(ColiCudaTensor *w,ColiCudaTensor *proj,flo extern "C" int coli_cuda_attention_absorb_batch(ColiCudaTensor *w,float *ctx,const float *q, const float *latent,const float *rope,int S,int H,int Q,int R,int V,int K,int T, float scale){ + if (fault_injected()) return 0; return attention_absorb_batch_run(w,nullptr,ctx,q,latent,rope,S,H,Q,R,V,K,T,scale); } extern "C" int coli_cuda_attention_project_batch(ColiCudaTensor *w,ColiCudaTensor *proj, float *out,const float *q,const float *latent,const float *rope,int S,int H,int Q, int R,int V,int K,int T,float scale){ + if (fault_injected()) return 0; return attention_absorb_batch_run(w,proj,out,q,latent,rope,S,H,Q,R,V,K,T,scale); } @@ -900,7 +1109,7 @@ extern "C" int coli_cuda_attention_project_ragged(ColiCudaTensor *w,ColiCudaTens attention_absorb_ragged_kernel<<stream>>>(dc->ac,dc->aq,ddl,ddr, dn,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale); quant_matmul<<O,S),256,0,dc->stream>>>(dc->y,dc->ac,proj->weights, - proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I)); + proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I),proj->gs,proj->ng); return cuda_ok(cudaGetLastError(),"ragged attention launch")&& cuda_ok(cudaMemcpyAsync(out,dc->y,ob,cudaMemcpyDeviceToHost,dc->stream),"ragged output download")&& cuda_ok(cudaStreamSynchronize(dc->stream),"ragged attention synchronize"); @@ -911,7 +1120,8 @@ extern "C" void coli_cuda_tensor_free(ColiCudaTensor *tensor) { DeviceContext *ctx = find_ctx(tensor->device); if (ctx) select_ctx(ctx); if (tensor->tracked && ctx) { - size_t bytes = tensor->weight_bytes + (tensor->fmt ? (size_t)tensor->O * sizeof(float) : 0); + int ng = tensor->ng > 0 ? tensor->ng : 1; + size_t bytes = tensor->weight_bytes + (tensor->fmt ? (size_t)tensor->O * ng * sizeof(float) : 0); if (ctx->tensor_count) ctx->tensor_count--; if (ctx->tensor_bytes >= bytes) ctx->tensor_bytes -= bytes; } @@ -925,7 +1135,9 @@ extern "C" void coli_cuda_tensor_free(ColiCudaTensor *tensor) { } extern "C" size_t coli_cuda_tensor_bytes(const ColiCudaTensor *tensor) { - return tensor ? tensor->weight_bytes + (tensor->fmt ? (size_t)tensor->O * sizeof(float) : 0) : 0; + if (!tensor) return 0; + int ng = tensor->ng > 0 ? tensor->ng : 1; + return tensor->weight_bytes + (tensor->fmt ? (size_t)tensor->O * ng * sizeof(float) : 0); } extern "C" int coli_cuda_tensor_device(const ColiCudaTensor *tensor) { @@ -986,7 +1198,7 @@ __global__ static void pipe_rows_add(float *x,const float *partial,const int *ro * per layer (78 x ~10 alloc/richiesta erano puro churn). */ extern "C" float *coli_cuda_pipe_scratch(int device,int slot,size_t bytes){ DeviceContext *ctx=find_ctx(device); - if(slot<0||slot>=24||!select_ctx(ctx)) return NULL; + if(slot<0||slot>=27||!select_ctx(ctx)) return NULL; if(!reserve(&ctx->pipe_buf[slot],&ctx->pipe_cap[slot],bytes)) return NULL; return ctx->pipe_buf[slot]; } @@ -1010,6 +1222,7 @@ extern "C" int coli_cuda_pipe_download(int device,const void *src,void *dst,size } extern "C" int coli_cuda_pipe_rmsnorm(int device,float *y_dev,const float *x_dev, const float *w_dev,int S,int D,float eps){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(S<1||D<1||!select_ctx(ctx)) return 0; pipe_rmsnorm_rows<<>>(y_dev,x_dev,w_dev,D,eps,D,D); @@ -1018,6 +1231,7 @@ extern "C" int coli_cuda_pipe_rmsnorm(int device,float *y_dev,const float *x_dev extern "C" int coli_cuda_pipe_rmsnorm_s(int device,float *y_dev,const float *x_dev, const float *w_dev,int S,int D,float eps, int xstride,int ystride){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(S<1||D<1||xstride>>(y_dev,x_dev,w_dev,D,eps,xstride,ystride); @@ -1026,6 +1240,7 @@ extern "C" int coli_cuda_pipe_rmsnorm_s(int device,float *y_dev,const float *x_d extern "C" int coli_cuda_pipe_rope(int device,float *v_dev,const int *pos_dev, int rows,int stride,int offset,int R,int heads, float theta){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(rows<1||R<2||R>256||heads<1||!select_ctx(ctx)) return 0; pipe_rope_rows<<>>(v_dev,pos_dev,0,stride,offset,R,heads,theta); @@ -1033,11 +1248,88 @@ extern "C" int coli_cuda_pipe_rope(int device,float *v_dev,const int *pos_dev, } extern "C" int coli_cuda_pipe_rope_base(int device,float *v_dev,int pos_base,int rows, int stride,int offset,int R,int heads,float theta){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(rows<1||R<2||R>256||heads<1||!select_ctx(ctx)) return 0; pipe_rope_rows<<>>(v_dev,NULL,pos_base,stride,offset,R,heads,theta); return cuda_ok(cudaGetLastError(),"pipe rope base"); } +/* ---- device router (#431 PR-A) ------------------------------------------- + * Router for one decode row, entirely on the layer's home device: logits GEMV + * (E x D, tiny) + sigmoid, bias-augmented top-K selection, route-level TOPP + * truncation, norm_topk and routed_scale — a float-faithful clone of moe()'s + * plain routing path (colibri.c FASE A). Selection runs single-thread so the + * argmax order, tie-breaking (strict >, lowest index wins) and weight math + * match the CPU reference exactly; only the dot/expf rounding can differ, + * which is the documented kernel-family divergence class (#100/#163). + * Results are packed [idx[K] | w[K] | keff] in one scratch buffer and read + * back with a single tiny D2H. */ +__global__ void pipe_router_logits(const float *__restrict__ x, + const float *__restrict__ W, + const float *__restrict__ bias, + int D, float *logit, float *choice){ + int e = blockIdx.x; + const float *w = W + (size_t)e*D; + float acc = 0.f; + for(int i=threadIdx.x; i>1; s>0; s>>=1){ + if(threadIdx.xbv){bv=choice[e];best=e;} } + idx[kk]=best; w[kk]=logit[best]; + } + int Ke=Ksel; + if(topp>0.f && topp<1.f){ + for(int a=1;a=0 && w[b]=topp*tot){ Ke=kk+1; break; } } + } + if(norm_topk){ float sm=0.f; for(int kk=0;kk4096||Ksel<1||Ksel>64||!select_ctx(ctx)) return 0; + size_t pack=(size_t)Ksel*(sizeof(int)+sizeof(float))+sizeof(int); + float *logit=coli_cuda_pipe_scratch(device,22,(size_t)E*sizeof(float)); + float *chc =coli_cuda_pipe_scratch(device,23,(size_t)E*sizeof(float)); + char *out =(char*)coli_cuda_pipe_scratch(device,24,pack); + if(!logit||!chc||!out) return 0; + pipe_router_logits<<>>(x_dev,(const float*)rw_dev,(const float*)rb_dev,D,logit,chc); + pipe_router_select<<<1,1>>>(logit,chc,E,Ksel,topp,norm_topk,routed_scale,out); + if(!cuda_ok(cudaGetLastError(),"pipe router launch")) return 0; + char buf[64*(sizeof(int)+sizeof(float))+sizeof(int)]; + if(!cuda_ok(cudaMemcpy(buf,out,pack,cudaMemcpyDeviceToHost),"pipe router readback")) return 0; + memcpy(idx_host,buf,(size_t)Ksel*sizeof(int)); + memcpy(w_host,buf+Ksel*sizeof(int),(size_t)Ksel*sizeof(float)); + memcpy(keff_host,buf+Ksel*(sizeof(int)+sizeof(float)),sizeof(int)); + return 1; +} extern "C" int coli_cuda_pipe_copy2d(int device,float *dst,int dpitch,const float *src, int spitch,int width,int height){ DeviceContext *ctx=find_ctx(device); if(!select_ctx(ctx)) return 0; @@ -1050,6 +1342,7 @@ extern "C" int coli_cuda_pipe_copy2d(int device,float *dst,int dpitch,const floa extern "C" int coli_cuda_attention_project_batch_dev(ColiCudaTensor *w,ColiCudaTensor *proj, float *out,const float *q_dev,const float *latent_dev,const float *rope_dev, int S,int H,int Q,int R,int V,int K,int T,float scale){ + if (fault_injected()) return 0; if(!w||!proj||!out||!q_dev||!latent_dev||!rope_dev||S<1||H<1||Q<1||R<1||V<1|| K<1||K>512||T8192||w->I!=K||w->O!=H*(Q+V)|| proj->device!=w->device||proj->I!=H*V)return 0; @@ -1058,12 +1351,12 @@ extern "C" int coli_cuda_attention_project_batch_dev(ColiCudaTensor *w,ColiCudaT if(!reserve(&dc->ac,&dc->ac_cap,cb))return 0; size_t shared=(size_t)(2*K+T+256)*sizeof(float); attention_absorb_batch_kernel<<stream>>>(dc->ac,q_dev,latent_dev, - rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale); + rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"pipe attention launch"))return 0; size_t ob=(size_t)S*proj->O*sizeof(float); if(!reserve(&dc->y,&dc->y_cap,ob))return 0; quant_matmul<<O,S),256,0,dc->stream>>>(dc->y,dc->ac,proj->weights, - proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I)); + proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I),proj->gs,proj->ng); if(!cuda_ok(cudaGetLastError(),"pipe o_proj launch"))return 0; if(!cuda_ok(cudaMemcpyAsync(out,dc->y,ob,cudaMemcpyDeviceToHost,dc->stream),"pipe attention download")|| !cuda_ok(cudaStreamSynchronize(dc->stream),"pipe attention sync"))return 0; @@ -1071,17 +1364,20 @@ extern "C" int coli_cuda_attention_project_batch_dev(ColiCudaTensor *w,ColiCudaT } extern "C" int coli_cuda_pipe_silu_mul(int device,float *gate_dev,const float *up_dev, size_t n){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(!n||!select_ctx(ctx)) return 0; silu_mul<<<(unsigned)((n+255)/256),256>>>(gate_dev,up_dev,n); return cuda_ok(cudaGetLastError(),"pipe silu mul"); } extern "C" int coli_cuda_pipe_add(int device,float *x_dev,const float *t_dev,size_t n){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(!n||!select_ctx(ctx)) return 0; pipe_add_n<<<(unsigned)((n+255)/256),256>>>(x_dev,t_dev,n); return cuda_ok(cudaGetLastError(),"pipe add"); } extern "C" int coli_cuda_pipe_rows_add(int device,float *x_dev,const float *partial_dev, const int *rows_dev,int nrows,int D){ + if (fault_injected()) return 0; DeviceContext *ctx=find_ctx(device); if(nrows<1||D<1||!select_ctx(ctx)) return 0; pipe_rows_add<<>>(x_dev,partial_dev,rows_dev,D); return cuda_ok(cudaGetLastError(),"pipe rows add"); @@ -1090,11 +1386,12 @@ extern "C" int coli_cuda_pipe_rows_add(int device,float *x_dev,const float *part * coli_cuda_matmul, zero host transfers. */ extern "C" int coli_cuda_pipe_gemm(ColiCudaTensor *t,float *y_dev,const float *x_dev, int S){ + if (fault_injected()) return 0; if(!t||S<1) return 0; DeviceContext *ctx=find_ctx(t->device); if(!select_ctx(ctx)) return 0; dim3 grid((unsigned)t->O,(unsigned)S); quant_matmul<<>>(y_dev,x_dev,t->weights,t->scales,t->fmt,S,t->I,t->O, - row_bytes(t->fmt,t->I)); + row_bytes(t->fmt,t->I),t->gs,t->ng); return cuda_ok(cudaGetLastError(),"pipe gemm"); } /* copia diretta scheda->scheda (P2P se disponibile, altrimenti staging driver) */ @@ -1109,6 +1406,7 @@ extern "C" int coli_cuda_pipe_peer_copy(int dst_dev,float *dst,int src_dev, extern "C" int coli_cuda_attention_project_batch_dev_out(ColiCudaTensor *w,ColiCudaTensor *proj, float *out_dev,const float *q_dev,const float *latent_dev,const float *rope_dev, int S,int H,int Q,int R,int V,int K,int T,float scale){ + if (fault_injected()) return 0; if(!w||!proj||!out_dev||!q_dev||!latent_dev||!rope_dev||S<1||H<1||Q<1||R<1||V<1|| K<1||K>512||T8192||w->I!=K||w->O!=H*(Q+V)|| proj->device!=w->device||proj->I!=H*V)return 0; @@ -1117,10 +1415,10 @@ extern "C" int coli_cuda_attention_project_batch_dev_out(ColiCudaTensor *w,ColiC if(!reserve(&dc->ac,&dc->ac_cap,cb))return 0; size_t shared=(size_t)(2*K+T+256)*sizeof(float); attention_absorb_batch_kernel<<stream>>>(dc->ac,q_dev,latent_dev, - rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale); + rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"pipe attention launch (dev out)"))return 0; quant_matmul<<O,S),256,0,dc->stream>>>(out_dev,dc->ac,proj->weights, - proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I)); + proj->scales,proj->fmt,S,proj->I,proj->O,row_bytes(proj->fmt,proj->I),proj->gs,proj->ng); if(!cuda_ok(cudaGetLastError(),"pipe o_proj launch (dev out)"))return 0; return cuda_ok(cudaStreamSynchronize(dc->stream),"pipe attention sync (dev out)"); } @@ -1130,12 +1428,13 @@ extern "C" int coli_cuda_attention_project_batch_dev_out(ColiCudaTensor *w,ColiC extern "C" int coli_cuda_attention_absorb_batch_dev(ColiCudaTensor *w,float *ctx_dev, const float *q_dev,const float *latent_dev,const float *rope_dev, int S,int H,int Q,int R,int V,int K,int T,float scale){ + if (fault_injected()) return 0; if(!w||!ctx_dev||!q_dev||!latent_dev||!rope_dev||S<1||H<1||Q<1||R<1||V<1|| K<1||K>512||T8192||w->I!=K||w->O!=H*(Q+V))return 0; DeviceContext *dc=find_ctx(w->device);if(!select_ctx(dc))return 0; size_t shared=(size_t)(2*K+T+256)*sizeof(float); attention_absorb_batch_kernel<<stream>>>(ctx_dev,q_dev,latent_dev, - rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale); + rope_dev,w->weights,w->scales,w->fmt,S,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"pipe shard attention launch"))return 0; return cuda_ok(cudaStreamSynchronize(dc->stream),"pipe shard attention sync"); } @@ -1144,6 +1443,7 @@ extern "C" int coli_cuda_attention_absorb_batch_dev(ColiCudaTensor *w,float *ctx extern "C" int coli_cuda_attention_absorb_kvdev(ColiCudaTensor *w,float *ctx,const float *q, const float *latent_dev,const float *rope_dev,int H,int Q,int R,int V,int K,int T, float scale){ + if (fault_injected()) return 0; if(!w||!ctx||!q||!latent_dev||!rope_dev||H<1||Q<1||R<1||V<1||K<1||K>512||T<1||T>8192|| w->I!=K||w->O!=H*(Q+V))return 0; DeviceContext *dc=find_ctx(w->device);if(!select_ctx(dc))return 0; @@ -1152,7 +1452,7 @@ extern "C" int coli_cuda_attention_absorb_kvdev(ColiCudaTensor *w,float *ctx,con if(!cuda_ok(cudaMemcpyAsync(dc->aq,q,qb,cudaMemcpyHostToDevice,dc->stream),"kvdev q upload"))return 0; size_t shared=(size_t)(2*K+T+256)*sizeof(float); attention_absorb_batch_kernel<<stream>>>(dc->ac,dc->aq,latent_dev, - rope_dev,w->weights,w->scales,w->fmt,1,H,Q,R,V,K,T,scale); + rope_dev,w->weights,w->scales,w->fmt,1,H,Q,R,V,K,T,scale,w->gs,w->ng); if(!cuda_ok(cudaGetLastError(),"kvdev absorb launch")|| !cuda_ok(cudaMemcpyAsync(ctx,dc->ac,cb,cudaMemcpyDeviceToHost,dc->stream),"kvdev ctx download")|| !cuda_ok(cudaStreamSynchronize(dc->stream),"kvdev absorb sync"))return 0; diff --git a/c/backend_cuda.h b/c/backend_cuda.h index 63c1f11..5196a13 100644 --- a/c/backend_cuda.h +++ b/c/backend_cuda.h @@ -36,20 +36,24 @@ COLI_CUDA_DLLEXPORT void coli_cuda_group_stats(uint64_t *calls, uint64_t *expert double *h2d_ms, double *kernel_ms, double *d2h_ms); /* Upload without executing, so capacity failures happen during model startup. */ +COLI_CUDA_DLLEXPORT int coli_cuda_tensor_upload_g(ColiCudaTensor **tensor, + const void *weights, const float *scales, + int fmt, int I, int O, int device, int gs); COLI_CUDA_DLLEXPORT int coli_cuda_tensor_upload(ColiCudaTensor **tensor, const void *weights, const float *scales, int fmt, int I, int O, int device); /* * y[S,O] = x[S,I] @ W[O,I]^T. - * fmt matches QT in glm.c: 0=f32, 1=int8, 2=int4, 3=int2. - * The first successful call uploads W and its row scales; later calls reuse it. + * fmt matches QT in glm.c: 0=f32, 1=int8, 2=int4, 3=int2, 4=grouped int4. + * gs is the group size for fmt=4 (0 for all other formats). + * The first successful call uploads W and its scales; later calls reuse it. * Returns 1 on success and 0 when CUDA is not initialized or the format is invalid. */ COLI_CUDA_DLLEXPORT int coli_cuda_matmul(ColiCudaTensor **tensor, float *y, const float *x, const void *weights, const float *scales, - int fmt, int S, int I, int O, int device); + int fmt, int S, int I, int O, int device, int gs); /* Fused expert pipeline: y = down(silu(gate(x)) * up(x)). All three tensors * must already be resident on one device. Activations cross PCIe once in @@ -67,6 +71,16 @@ COLI_CUDA_DLLEXPORT int coli_cuda_shared_mlp_w4a16(ColiCudaTensor *gate, ColiCud /* Packed group of same-shaped experts. Inputs and outputs contain sum(rows) * consecutive [D] rows in call order. */ +/* Async issue/take split of the group call below (Inc.4): issue launches on the + * device stream and returns; take syncs and returns the pinned result rows (valid + * until the next issue on that device). Small totals only (<=8 rows); one + * outstanding issue per device. */ +COLI_CUDA_DLLEXPORT int coli_cuda_expert_group_issue(ColiCudaTensor *const *gates, + ColiCudaTensor *const *ups, + ColiCudaTensor *const *downs, + const int *rows, int count, const float *x); +COLI_CUDA_DLLEXPORT const float *coli_cuda_expert_group_take(int device); + COLI_CUDA_DLLEXPORT int coli_cuda_expert_group(ColiCudaTensor *const *gates, ColiCudaTensor *const *ups, ColiCudaTensor *const *downs, @@ -126,6 +140,10 @@ COLI_CUDA_DLLEXPORT int coli_cuda_pipe_rmsnorm_s(int device,float *y_dev,const f int xstride,int ystride); COLI_CUDA_DLLEXPORT int coli_cuda_pipe_rope_base(int device,float *v_dev,int pos_base,int rows, int stride,int offset,int R,int heads,float theta); +COLI_CUDA_DLLEXPORT int coli_cuda_pipe_router(int device,const float *x_dev, + const void *rw_dev,const void *rb_dev,int D,int E,int Ksel, + float topp,int norm_topk,float routed_scale, + int *idx_host,float *w_host,int *keff_host); COLI_CUDA_DLLEXPORT int coli_cuda_pipe_copy2d(int device,float *dst,int dpitch,const float *src, int spitch,int width,int height); COLI_CUDA_DLLEXPORT int coli_cuda_attention_project_batch_dev(ColiCudaTensor *kv_b,ColiCudaTensor *o_proj, diff --git a/c/backend_loader.c b/c/backend_loader.c index eedbd50..e65f641 100644 --- a/c/backend_loader.c +++ b/c/backend_loader.c @@ -41,14 +41,20 @@ typedef int (*fn_expert_mlp)(ColiCudaTensor *gate, ColiCudaTensor *up typedef int (*fn_expert_group)(ColiCudaTensor *const *gates, ColiCudaTensor *const *ups, ColiCudaTensor *const *downs, const int *rows, int count, float *y, const float *x); +typedef int (*fn_expert_group_issue)(ColiCudaTensor *const *gates, + ColiCudaTensor *const *ups, + ColiCudaTensor *const *downs, + const int *rows, int count, const float *x); +typedef const float * (*fn_expert_group_take)(int device); typedef int (*fn_attention_absorb)(ColiCudaTensor *kv_b, float *ctx, const float *q, const float *latent, const float *rope, int H, int Q, int R, int V, int K, int T, float attention_scale); typedef int (*fn_tensor_upload)(ColiCudaTensor **tensor, const void *weights, const float *scales, int fmt, int I, int O, int device); +typedef int (*fn_tensor_upload_g)(ColiCudaTensor **tensor, const void *weights, const float *scales, int fmt, int I, int O, int device, int gs); typedef int (*fn_matmul)(ColiCudaTensor **tensor, float *y, const float *x, const void *weights, const float *scales, - int fmt, int S, int I, int O, int device); + int fmt, int S, int I, int O, int device, int gs); typedef void (*fn_tensor_free)(ColiCudaTensor *tensor); typedef size_t (*fn_tensor_bytes)(const ColiCudaTensor *tensor); typedef int (*fn_tensor_device)(const ColiCudaTensor *tensor); @@ -76,6 +82,7 @@ typedef int (*fn_pipe_gemm)(ColiCudaTensor *t,float *y_dev,const float *x_dev,in typedef int (*fn_pipe_peer_copy)(int dst_dev,float *dst,int src_dev, const float *src,size_t bytes); typedef int (*fn_pipe_rmsnorm)(int device,float *y_dev,const float *x_dev, const float *w_dev,int S,int D,float eps); typedef int (*fn_pipe_rmsnorm_s)(int device,float *y_dev,const float *x_dev, const float *w_dev,int S,int D,float eps, int xstride,int ystride); +typedef int (*fn_pipe_router)(int device,const float *x_dev,const void *rw_dev,const void *rb_dev,int D,int E,int Ksel,float topp,int norm_topk,float routed_scale,int *idx_host,float *w_host,int *keff_host); typedef int (*fn_pipe_rope)(int device,float *v_dev,const int *pos_dev,int rows, int stride,int offset,int R,int heads,float theta); typedef int (*fn_pipe_rope_base)(int device,float *v_dev,int pos_base,int rows, int stride,int offset,int R,int heads,float theta); typedef int (*fn_pipe_rows_add)(int device,float *x_dev,const float *partial_dev, const int *rows_dev,int nrows,int D); @@ -100,8 +107,11 @@ static struct { fn_group_stats group_stats; fn_expert_mlp expert_mlp; fn_expert_group expert_group; + fn_expert_group_issue expert_group_issue; + fn_expert_group_take expert_group_take; fn_attention_absorb attention_absorb; fn_tensor_upload tensor_upload; + fn_tensor_upload_g tensor_upload_g; fn_matmul matmul; fn_tensor_free tensor_free; fn_tensor_bytes tensor_bytes; @@ -123,6 +133,7 @@ static struct { fn_pipe_peer_copy pipe_peer_copy; fn_pipe_rmsnorm pipe_rmsnorm; fn_pipe_rmsnorm_s pipe_rmsnorm_s; + fn_pipe_router pipe_router; fn_pipe_rope pipe_rope; fn_pipe_rope_base pipe_rope_base; fn_pipe_rows_add pipe_rows_add; @@ -194,8 +205,11 @@ static int coli_cuda_load(void){ RESOLVE(group_stats, fn_group_stats) RESOLVE(expert_mlp, fn_expert_mlp) RESOLVE(expert_group, fn_expert_group) + RESOLVE(expert_group_issue, fn_expert_group_issue) + RESOLVE(expert_group_take, fn_expert_group_take) RESOLVE(attention_absorb, fn_attention_absorb) RESOLVE(tensor_upload, fn_tensor_upload) + RESOLVE(tensor_upload_g, fn_tensor_upload_g) RESOLVE(matmul, fn_matmul) RESOLVE(tensor_free, fn_tensor_free) RESOLVE(tensor_bytes, fn_tensor_bytes) @@ -217,6 +231,7 @@ static int coli_cuda_load(void){ RESOLVE(pipe_peer_copy, fn_pipe_peer_copy) RESOLVE(pipe_rmsnorm, fn_pipe_rmsnorm) RESOLVE(pipe_rmsnorm_s, fn_pipe_rmsnorm_s) + RESOLVE(pipe_router, fn_pipe_router) RESOLVE(pipe_rope, fn_pipe_rope) RESOLVE(pipe_rope_base, fn_pipe_rope_base) RESOLVE(pipe_rows_add, fn_pipe_rows_add) @@ -289,6 +304,19 @@ int coli_cuda_expert_group(ColiCudaTensor *const *gates, ColiCudaTensor *const * return g_cuda.expert_group(gates, ups, downs, rows, count, y, x); } +int coli_cuda_expert_group_issue(ColiCudaTensor *const *gates, + ColiCudaTensor *const *ups, + ColiCudaTensor *const *downs, + const int *rows, int count, const float *x){ + if(!g_cuda.available) return 0; + return g_cuda.expert_group_issue(gates, ups, downs, rows, count, x); +} + +const float *coli_cuda_expert_group_take(int device){ + if(!g_cuda.available) return NULL; + return g_cuda.expert_group_take(device); +} + int coli_cuda_attention_absorb(ColiCudaTensor *kv_b, float *ctx, const float *q, const float *latent, const float *rope, int H, int Q, int R, int V, int K, int T, float attention_scale){ @@ -302,11 +330,16 @@ int coli_cuda_tensor_upload(ColiCudaTensor **tensor, const void *weights, return g_cuda.tensor_upload(tensor, weights, scales, fmt, I, O, device); } +int coli_cuda_tensor_upload_g(ColiCudaTensor **tensor, const void *weights, const float *scales, int fmt, int I, int O, int device, int gs){ + if(!g_cuda.available || !g_cuda.tensor_upload_g){ return 0; } + return g_cuda.tensor_upload_g(tensor, weights, scales, fmt, I, O, device, gs); +} + int coli_cuda_matmul(ColiCudaTensor **tensor, float *y, const float *x, const void *weights, const float *scales, - int fmt, int S, int I, int O, int device){ + int fmt, int S, int I, int O, int device, int gs){ if(!g_cuda.available) return 0; - return g_cuda.matmul(tensor, y, x, weights, scales, fmt, S, I, O, device); + return g_cuda.matmul(tensor, y, x, weights, scales, fmt, S, I, O, device, gs); } void coli_cuda_tensor_free(ColiCudaTensor *tensor){ @@ -407,6 +440,11 @@ int coli_cuda_pipe_rmsnorm(int device,float *y_dev,const float *x_dev, const flo return g_cuda.pipe_rmsnorm(device, y_dev, x_dev, w_dev, S, D, eps); } +int coli_cuda_pipe_router(int device,const float *x_dev,const void *rw_dev,const void *rb_dev,int D,int E,int Ksel,float topp,int norm_topk,float routed_scale,int *idx_host,float *w_host,int *keff_host){ + if(!g_cuda.available || !g_cuda.pipe_router){ return 0; } + return g_cuda.pipe_router(device, x_dev, rw_dev, rb_dev, D, E, Ksel, topp, norm_topk, routed_scale, idx_host, w_host, keff_host); +} + int coli_cuda_pipe_rmsnorm_s(int device,float *y_dev,const float *x_dev, const float *w_dev,int S,int D,float eps, int xstride,int ystride){ if(!g_cuda.available){ return 0; } return g_cuda.pipe_rmsnorm_s(device, y_dev, x_dev, w_dev, S, D, eps, xstride, ystride); diff --git a/c/backend_metal.h b/c/backend_metal.h index 4383b3b..35c4931 100644 --- a/c/backend_metal.h +++ b/c/backend_metal.h @@ -84,6 +84,23 @@ int coli_metal_layer_decode(float *x, int coli_metal_gemm(float *y, const float *x, const void *weights, const float *scales, int fmt, int S, int I, int O); /* large-batch sync GEMM; 0 -> CPU */ +/* Parallel top-8 expert selection (r_top8_par): run ONE top-8 selection kernel standalone + * on host arrays — par=0 the serial r_top8, par=1 the parallel exact-match replica gated + * in the engine by COLI_RTOP8 (default ON; COLI_RTOP8=0 opts out to the serial kernel). + * Exists so the metal-test suite (and any battery probe) can prove serial/parallel + * equivalence on the ENGINE build's own compiled shaders, not just in the bench tool. + * sig[S*E], bias[E], idx[S*K], w[S*K], keff[S]. + * Expert-count generality: the parallel kernel's blocked-lane design (ch[8]/32-lane + * threadgroup) is validated correct for arbitrary E<=256, including non-multiples of the + * 32-lane width and small E (see metal-test's E=24/E=168/E=256 cases — 168 is the REAP + * expert-pruned package width from #428/#426). For E>256 (out of contract) this function + * transparently falls back to the serial kernel even when par=1 is requested, and the + * same automatic fallback is wired into the engine dispatch site — "par" is a request, + * never a guarantee, so no caller can reach the unguarded parallel path out of contract. + * Returns 1 on success, 0 if Metal is unavailable. */ +int coli_metal_rtop8(int par, const float *sig, const float *bias, int S, int E, int K, + int Ksel, float topp, int normk, float rscale, + int *idx, float *w, int *keff); void coli_metal_attn_counts(uint64_t *ok, double *wall, double *kernel); void coli_metal_attn_lat(double *ksched, double *gsched); int coli_metal_attn_decode(const float *x, @@ -98,6 +115,10 @@ int coli_metal_attn_decode(const float *x, void coli_metal_moe_counts(uint64_t *ok, uint64_t *fb, uint64_t *experts); void coli_metal_moe_times(double *setup, double *gpu, double *scatter); double coli_metal_moe_kernel_time(void); +/* E5 (COLI_METAL_RESSET=1): returns 1 when the queue-attached residency set is active and + * writes the cumulative seconds moe_submit spent committing pending set adds -- a cost that + * sits OUTSIDE the setup/gpu/scatter breakdown above. Returns 0 (and writes 0) when off. */ +int coli_metal_resset_stats(double *flush_s); /* * Batched routed-expert SwiGLU for one MoE block, in ONE command buffer. diff --git a/c/backend_metal.mm b/c/backend_metal.mm index 51ed8a0..6ec3fb4 100644 --- a/c/backend_metal.mm +++ b/c/backend_metal.mm @@ -219,6 +219,63 @@ kernel void r_top8(device const float* sig [[buffer(0)]], device const float* bi if(normk){ float sm=0; for(int kk=0;kk' ascending max (lowest index wins +// within a lane, matching the serial ascending scan), then a shuffle-down argmax +// reduction where ties prefer the LOWER index — together exactly the serial kernel's +// first-max-wins order. The topp/normk/rscale tail is the serial code verbatim on lane 0 +// (same ops, same order => bitwise-identical results; metal-test enforces this with +// memcmp). Contract: E<=256 (ch[8]/taken mask sizing: ceil(E/32)<=8) — the defensive +// return below makes an out-of-contract dispatch a visible no-op (idx/w/keff untouched), +// never an OOB write; both call sites (coli_metal_layer_decode's dispatch and the +// standalone coli_metal_rtop8 runner) additionally gate on E<=256 in host code before +// selecting this pipeline at all, so the return here is defense-in-depth, not the only +// guard. Sentinel-per-lane design (ch[j]=-1e30f for e>=E) makes non-multiple-of-32 E +// and small E correct without special-casing — validated for E=24, E=168 (REAP +// expert-pruned packages, see the upstream feature-request thread) and E=256 by metal-test. +// ASSUMES SIMD width 32 (shuffle offsets 16..1, 32-thread threadgroup per row): enforced +// at init — coli_metal_init clears g_rtop8_width_ok (and therefore both call sites' use +// of this pipeline) if threadExecutionWidth != 32. +kernel void r_top8_par(device const float* sig [[buffer(0)]], device const float* bias [[buffer(1)]], + device int* idx [[buffer(2)]], device float* w [[buffer(3)]], + device int* keff [[buffer(4)]], constant int& E [[buffer(5)]], + constant int& K [[buffer(6)]], constant int& Ksel [[buffer(7)]], + constant float& topp [[buffer(8)]], constant int& normk [[buffer(9)]], + constant float& rscale [[buffer(10)]], + uint s [[threadgroup_position_in_grid]], + uint slane [[thread_index_in_simdgroup]]) { + if(E>256) return; + device const float* sg=sig+(long)s*E; + device int* id_=idx+(long)s*K; device float* ww=w+(long)s*K; + int per=(E+31)/32, base=(int)slane*per; + float ch[8]; uint taken=0u; + for(int j=0;jbv){ bv=ch[j]; bi=base+j; } + for(uint off=16;off>0;off>>=1){ + float ov=simd_shuffle_down(bv,off); int oi=simd_shuffle_down(bi,off); + if(ov>bv || (ov==bv && oi=base && bi0.0f && topp<1.0f){ + for(int a=1;a=0 && ww[b]=topp*tot){ Ke=kk+1; break; } } + } + keff[s]=Ke; + if(normk){ float sm=0; for(int kk=0;kk g_dev; static id g_queue; static id g_gemv, g_moe_gemv, g_moe_silu; static id g_a_rms, g_a_rope, g_a_copy, g_a_qabs, g_a_score, g_a_smax, g_a_clat, g_a_ctx; -static id g_a_add, g_r_router, g_r_top8; +static id g_a_add, g_r_router, g_r_top8, g_r_top8p; +static int g_rtop8_par = 1; // COLI_RTOP8 (default ON); COLI_RTOP8=0 opts out to the + // serial kernel — see coli_metal_init. +static int g_rtop8_width_ok = 1; // hardware fact, independent of the policy gate above: + // false if this device's threadExecutionWidth != 32. + // Consulted by BOTH the engine dispatch site and the + // standalone coli_metal_rtop8 runner, so no caller can + // reach r_top8_par's 32-lane reduction on an unsafe + // device even by explicitly requesting par=1. static size_t g_tensor_count, g_tensor_bytes; static uint64_t g_moe_ok, g_moe_fb, g_moe_experts; // GPU blocks / CPU-fallback blocks / experts on GPU static double g_t_setup, g_t_gpu, g_t_scatter, g_t_kernel; // per-block time breakdown (seconds) @@ -258,6 +323,73 @@ extern "C" void coli_metal_attn_lat(double *ksched, double *gsched){ struct Slab { void *base; size_t len; id buf; }; static std::vector g_slabs; static std::mutex g_slab_mtx; // expert_load registers slabs from parallel OpenMP threads + +// ---- E5 experiment: COLI_METAL_RESSET=1 -- one persistent MTLResidencySet attached to +// g_queue (macOS 15+) replaces moe_submit's per-command-buffer useResource: loop over +// resolved expert weight/scale slabs. Allocation is untouched (same newBufferWithBytesNoCopy +// wrap as stock); only residency bookkeeping moves off the dispatch hot path -- see +// SUMMARY.md for why skipping useResource: there is safe (read-only, indirectly-referenced +// buffers only; residency sets don't do hazard tracking, but nothing here relied on it). +// g_resset_obj is a bare `id` (holds id) so the global's declared type +// carries no availability annotation -- the protocol name only appears inside +// @available(macOS 15.0, *) guards below, keeping -Wunguarded-availability clean. +static id g_resset_obj; +static bool g_resset_enabled; // COLI_METAL_RESSET=1, macOS 15+, and creation succeeded +static bool g_resset_dirty; // addAllocation: calls pending commit; g_resset_mtx-guarded +// Set mutations + dirty flag get their OWN mutex, never held together with g_slab_mtx: no +// live Metal call may run under the slab lock the parallel OMP loader threads contend on +// (E4's audit round 2 found exactly that shape -- mutex over a live Metal call -- as the +// leading suspect for its +12s expert-disk regression). g_slab_mtx keeps guarding g_slabs +// bookkeeping only, exactly as on stock. +static std::mutex g_resset_mtx; +static double g_t_resset_flush; // sec committing pending adds in moe_submit (gate on only) + +// Add a just-wrapped buffer to the set; commit deferred (an OMP loader burst batches into +// one commit at the next moe_submit instead of one per slab). Called by coli_metal_register +// after it drops g_slab_mtx but before it returns -- and the engine cannot dispatch an +// expert before the load that registers its slab returns, so any slab a given moe_submit +// can resolve() was added (and marked dirty) under g_resset_mtx strictly before that +// moe_submit's resset_flush() acquired the same mutex: the flush covers it. The slab-table +// ordering itself (register-before-resolve) is unchanged and stays under g_slab_mtx. +// Cost lands in the caller's existing expert-load accounting (t_ewait window in colibri.c); +// no separate counter for the add/remove side. +static void resset_add(id b) { + if (!g_resset_enabled) return; + std::lock_guard lk(g_resset_mtx); + if (@available(macOS 15.0, *)) { [(id)g_resset_obj addAllocation:b]; g_resset_dirty = true; } +} +// Remove + commit immediately, NOT deferred: the caller frees the underlying host memory +// right after coli_metal_unregister returns, so the removal must be applied before that -- +// an uncommitted-but-still-resident allocation pointing at freed memory is a use-after-free +// risk the GPU could act on. Also runs outside g_slab_mtx (see g_resset_mtx above). +static void resset_remove(id b) { + if (!g_resset_enabled) return; + std::lock_guard lk(g_resset_mtx); + if (@available(macOS 15.0, *)) { + id rs = (id)g_resset_obj; + [rs removeAllocation:b]; [rs commit]; + } + g_resset_dirty = false; // commit above also flushes any pending adds +} +// Flush pending adds before moe_submit relies on the set alone for residency -- the only +// caller that skips per-buffer useResource: (see moe_submit below). Takes g_resset_mtx +// only, never g_slab_mtx; the happens-before argument lives at resset_add above. +static void resset_flush() { + if (!g_resset_enabled) return; + std::lock_guard lk(g_resset_mtx); + if (!g_resset_dirty) return; + if (@available(macOS 15.0, *)) { [(id)g_resset_obj commit]; } + g_resset_dirty = false; +} +// Harness visibility for the flush cost, which sits OUTSIDE the moe_times setup/gpu +// breakdown (timed around resset_flush in moe_submit, before ts_start). Returns whether +// the set is active so colibri.c prints the METAL-RESSET line only when the gate is on -- +// stock output stays byte-identical. +extern "C" int coli_metal_resset_stats(double *flush_s) { + if (flush_s) *flush_s = g_t_resset_flush; + return g_resset_enabled ? 1 : 0; +} + // Persistent scratch buffers (grow-only) for the MoE pipeline. static id g_gg, g_uu, g_hh, g_xg; static size_t g_gg_cap, g_uu_cap, g_hh_cap, g_xg_cap; static id ensure(id b, size_t *cap, size_t need) { @@ -284,6 +416,8 @@ extern "C" int coli_metal_init(void) { if (g_dev) return 1; if (getenv("COLI_METAL_UNTRACKED") && atoi(getenv("COLI_METAL_UNTRACKED"))) g_res_opts = MTLResourceStorageModeShared | MTLResourceHazardTrackingModeUntracked; + { const char *e = getenv("COLI_RTOP8"); // default ON; COLI_RTOP8=0 opts out + if (e && atoi(e) == 0) g_rtop8_par = 0; } @autoreleasepool { g_dev = MTLCreateSystemDefaultDevice(); if (!g_dev) return 0; @@ -299,11 +433,45 @@ extern "C" int coli_metal_init(void) { auto P=[&](const char*n){ return [g_dev newComputePipelineStateWithFunction:[lib newFunctionWithName:@(n)] error:&err]; }; g_a_rms=P("a_rmsnorm"); g_a_rope=P("a_rope"); g_a_copy=P("a_copy"); g_a_qabs=P("a_qabs"); g_a_score=P("a_score"); g_a_smax=P("a_smax"); g_a_clat=P("a_clat"); g_a_ctx=P("a_ctx"); - g_a_add=P("a_add"); g_r_router=P("r_router"); g_r_top8=P("r_top8"); - if(!g_a_add||!g_r_router||!g_r_top8){ fprintf(stderr,"[metal] tail pipelines failed\n"); g_dev=nil; return 0; } + g_a_add=P("a_add"); g_r_router=P("r_router"); g_r_top8=P("r_top8"); g_r_top8p=P("r_top8_par"); + if(!g_a_add||!g_r_router||!g_r_top8||!g_r_top8p){ fprintf(stderr,"[metal] tail pipelines failed\n"); g_dev=nil; return 0; } + // r_top8_par's reduction hardcodes SIMD width 32 (shuffle-down offsets 16..1, one + // 32-thread threadgroup per row). True on all Apple Silicon shipped to date, but a + // non-32-width device would reduce wrongly AND race multiple lane-0 writers, so this + // is a hard safety fact (g_rtop8_width_ok), not just a policy default: it gates BOTH + // the engine dispatch site and the standalone coli_metal_rtop8 runner (degrade-to-safe, + // same pattern as the pool/ring fallbacks elsewhere) — no caller can opt back into an + // unsafe reduction on such a device, even by explicitly requesting par=1. + if ([g_r_top8p threadExecutionWidth] != 32) { + g_rtop8_width_ok = 0; + if (g_rtop8_par) + fprintf(stderr, "[metal] COLI_RTOP8 parallel top-8 disabled: threadExecutionWidth=%lu " + "!= 32 (r_top8_par's reduction assumes 32-lane simdgroups) — serial " + "r_top8 in use\n", (unsigned long)[g_r_top8p threadExecutionWidth]); + g_rtop8_par = 0; + } if (!g_gemv || !g_moe_gemv || !g_moe_silu || !g_a_rms || !g_a_rope || !g_a_copy || !g_a_qabs || !g_a_score || !g_a_smax || !g_a_clat || !g_a_ctx) { fprintf(stderr, "[metal] pipeline failed\n"); g_dev = nil; return 0; } + // E5 experiment: COLI_METAL_RESSET=1 -- see g_resset_obj comment above. + if (getenv("COLI_METAL_RESSET") && atoi(getenv("COLI_METAL_RESSET"))) { + if (@available(macOS 15.0, *)) { + MTLResidencySetDescriptor *rd = [MTLResidencySetDescriptor new]; + rd.initialCapacity = 4096; // hint only (internal array presize), not a hard limit + NSError *rerr = nil; + id rs = [g_dev newResidencySetWithDescriptor:rd error:&rerr]; + if (rs) { + [g_queue addResidencySet:rs]; + g_resset_obj = rs; g_resset_enabled = true; + fprintf(stderr, "[METAL] residency-set: on (macOS 15+, moe_submit skips per-buffer useResource:)\n"); + } else { + fprintf(stderr, "[METAL] residency-set create failed: %s -- stock per-CB residency path\n", + rerr ? [[rerr localizedDescription] UTF8String] : "?"); + } + } else { + fprintf(stderr, "[METAL] COLI_METAL_RESSET=1 requested but OS < macOS 15 -- stock per-CB residency path\n"); + } + } } return 1; } @@ -313,13 +481,29 @@ extern "C" void coli_metal_register(void *base, size_t len) { id b = [g_dev newBufferWithBytesNoCopy:base length:len options:g_res_opts deallocator:nil]; if (!b) return; - std::lock_guard lk(g_slab_mtx); // called from parallel expert_load threads - for (auto &s : g_slabs) if (s.base == base) { s.len = len; s.buf = b; return; } - g_slabs.push_back({base, len, b}); + id old = nil; // E5: replaced wrapper on re-register of a live base (defensive) + { + std::lock_guard lk(g_slab_mtx); // called from parallel expert_load threads + bool found = false; + for (auto &s : g_slabs) if (s.base == base) { old = s.buf; s.len = len; s.buf = b; found = true; break; } + if (!found) g_slabs.push_back({base, len, b}); + } + // E5, outside g_slab_mtx (no Metal call under the slab lock), before returning. Invariant + // defended: set membership mirrors g_slabs exactly -- a re-register of a live base must + // drop the replaced wrapper from the set (ARC releases our reference, but the set retains + // it and keeps its pages resident forever) before adding the new one. No in-tree caller + // re-registers a live base today; defensive. + if (old && old != b) resset_remove(old); + if (old != b) resset_add(b); } extern "C" void coli_metal_unregister(void *base) { - std::lock_guard lk(g_slab_mtx); - for (size_t i=0;i b = nil; + { + std::lock_guard lk(g_slab_mtx); + for (size_t i=0;i resolve(const void *p, uint64_t *addr) { @@ -357,7 +541,14 @@ extern "C" void coli_metal_spin_start(void) { } extern "C" void coli_metal_spin_stop(void) { g_spin_run.store(false); } -extern "C" void coli_metal_shutdown(void) { coli_metal_spin_stop(); g_gemv=nil; g_queue=nil; g_dev=nil; g_tensor_count=g_tensor_bytes=0; } +extern "C" void coli_metal_shutdown(void) { + coli_metal_spin_stop(); + if (g_resset_enabled) { + if (@available(macOS 15.0, *)) { [g_queue removeResidencySet:(id)g_resset_obj]; } + } + g_resset_obj=nil; g_resset_enabled=false; g_resset_dirty=false; + g_gemv=nil; g_queue=nil; g_dev=nil; g_tensor_count=g_tensor_bytes=0; +} extern "C" int coli_metal_available(void) { return g_dev != nil; } extern "C" void coli_metal_stats(size_t *c, size_t *b) { if(c)*c=g_tensor_count; if(b)*b=g_tensor_bytes; } extern "C" int coli_metal_mem_info(size_t *used, size_t *total) { @@ -597,12 +788,22 @@ extern "C" int coli_metal_layer_decode(float *x, // 5) silu(gate)*up + exact top-K select [e setComputePipelineState:g_moe_silu]; [e setBuffer:ash1_ offset:0 atIndex:0]; [e setBuffer:ash2_ offset:0 atIndex:1]; [e dispatchThreads:MTLSizeMake((size_t)S*SI,1,1) threadsPerThreadgroup:MTLSizeMake(256,1,1)]; - { [e setComputePipelineState:g_r_top8]; + { // COLI_RTOP8 (default ON) swaps the serial 1-thread-per-row select for the exact- + // match 1-simdgroup-per-row replica (same buffers/args; only pipeline+grid change). + // E<=256 is required by r_top8_par's ch[8]/32-lane blocking contract; this call + // site's E is always 256 today (layer_forward_rows' own architecture-shape gate in + // colibri.c requires c->n_experts==256 to reach coli_metal_layer_decode at all — + // see PR body "Scope statement") but the check is kept here too, defense-in-depth, + // so a future relaxation of that gate (e.g. to admit REAP-pruned E=168 models into + // the fused path) degrades safely to the serial kernel instead of mis-dispatching. + int use_par = g_rtop8_par && g_rtop8_width_ok && E<=256; + [e setComputePipelineState:use_par?g_r_top8p:g_r_top8]; [e setBuffer:asig_ offset:0 atIndex:0]; [e setBuffer:rbB offset:rboff atIndex:1]; [e setBuffer:aidx_ offset:0 atIndex:2]; [e setBuffer:aw_ offset:0 atIndex:3]; [e setBuffer:akeff_ offset:0 atIndex:4]; [e setBytes:&E length:4 atIndex:5]; [e setBytes:&K length:4 atIndex:6]; [e setBytes:&Ksel length:4 atIndex:7]; [e setBytes:&topp length:4 atIndex:8]; [e setBytes:&normk length:4 atIndex:9]; [e setBytes:&rscale length:4 atIndex:10]; - [e dispatchThreads:MTLSizeMake(S,1,1) threadsPerThreadgroup:MTLSizeMake(S,1,1)]; } + if(use_par) [e dispatchThreadgroups:MTLSizeMake(S,1,1) threadsPerThreadgroup:MTLSizeMake(32,1,1)]; + else [e dispatchThreads:MTLSizeMake(S,1,1) threadsPerThreadgroup:MTLSizeMake(S,1,1)]; } BAR(); // 6) shared down bind_gemv(e,shd_w,shd_s,shd_fmt,SI,AH,ash1_,ashout_,S); @@ -651,6 +852,44 @@ extern "C" int coli_metal_gemm(float *y, const float *x, const void *wp, const f return 1; } +// Standalone single-kernel runner for the top-8 select (see backend_metal.h). Fresh +// shared buffers per call (a test/probe path, not a hot path); grids exactly as the +// engine dispatch site: serial = S threads of one S-wide threadgroup, parallel = S +// threadgroups x 32 (one simdgroup per row). "par" is a REQUEST, not a guarantee: same +// E<=256 and SIMD-width-32 host-side checks as the engine dispatch site gate the actual +// pipeline choice, so a caller (including metal-test itself) can never reach the parallel +// kernel out of contract by asking for it — par=1 with E>256, or on a non-32-wide device, +// transparently runs the serial kernel instead and still returns 1 (success). +extern "C" int coli_metal_rtop8(int par, const float *sig, const float *bias, int S, int E, int K, + int Ksel, float topp, int normk, float rscale, + int *idx, float *w, int *keff) { + if (!g_dev || S < 1 || E < 1 || K < 1 || Ksel < 1 || Ksel > K) return 0; + int use_par = par && g_r_top8p && g_rtop8_width_ok && E<=256; + @autoreleasepool { + id bs=[g_dev newBufferWithBytes:sig length:(size_t)S*E*4 options:MTLResourceStorageModeShared]; + id bb=[g_dev newBufferWithBytes:bias length:(size_t)E*4 options:MTLResourceStorageModeShared]; + id bi=[g_dev newBufferWithLength:(size_t)S*K*4 options:MTLResourceStorageModeShared]; + id bw=[g_dev newBufferWithLength:(size_t)S*K*4 options:MTLResourceStorageModeShared]; + id bk=[g_dev newBufferWithLength:(size_t)S*4 options:MTLResourceStorageModeShared]; + if(!bs||!bb||!bi||!bw||!bk) return 0; + memset(bi.contents,0xFF,(size_t)S*K*4); // poison: untouched slots stay visible + id cb=[g_queue commandBuffer]; id e=[cb computeCommandEncoder]; + [e setComputePipelineState:use_par?g_r_top8p:g_r_top8]; + [e setBuffer:bs offset:0 atIndex:0]; [e setBuffer:bb offset:0 atIndex:1]; + [e setBuffer:bi offset:0 atIndex:2]; [e setBuffer:bw offset:0 atIndex:3]; [e setBuffer:bk offset:0 atIndex:4]; + [e setBytes:&E length:4 atIndex:5]; [e setBytes:&K length:4 atIndex:6]; [e setBytes:&Ksel length:4 atIndex:7]; + [e setBytes:&topp length:4 atIndex:8]; [e setBytes:&normk length:4 atIndex:9]; [e setBytes:&rscale length:4 atIndex:10]; + if(use_par) [e dispatchThreadgroups:MTLSizeMake((NSUInteger)S,1,1) threadsPerThreadgroup:MTLSizeMake(32,1,1)]; + else [e dispatchThreads:MTLSizeMake((NSUInteger)S,1,1) threadsPerThreadgroup:MTLSizeMake((NSUInteger)S,1,1)]; + [e endEncoding]; [cb commit]; [cb waitUntilCompleted]; + if(cb.status==MTLCommandBufferStatusError){ fprintf(stderr,"[metal] rtop8 cmdbuf error\n"); return 0; } + memcpy(idx,bi.contents,(size_t)S*K*4); + memcpy(w,bw.contents,(size_t)S*K*4); + memcpy(keff,bk.contents,(size_t)S*4); + } + return 1; +} + extern "C" void coli_metal_tensor_free(ColiMetalTensor *t) { if (!t) return; g_tensor_count--; g_tensor_bytes -= t->wbytes; @@ -668,6 +907,9 @@ static id moe_submit(int nb, int D, int Iinter, int fmt, const float *xg, const int *xoff, const int *nr, int R, id xg_buf, id gg_buf, id uu_buf, id hh_buf) { if (!g_dev || (fmt != 1 && fmt != 2)) return nil; + if (g_resset_enabled) { // E5: commit any pending slab adds before we may skip useResource: + double t0 = mnow(); resset_flush(); g_t_resset_flush += mnow() - t0; // METAL-RESSET line + } double ts_start = mnow(); std::vector ag(nb),au(nb),ad(nb),sgv(nb),suv(nb),sdv(nb); std::vector> use; use.reserve(nb*2); @@ -689,7 +931,18 @@ static id moe_submit(int nb, int D, int Iinter, int fmt, memcpy([xg_buf contents], xg, (size_t)R*D*4); id cb=[g_queue commandBuffer]; id e=[cb computeCommandEncoder]; - for(auto&b:use) [e useResource:b usage:MTLResourceUsageRead]; + // E5 (COLI_METAL_RESSET=1): the queue-attached MTLResidencySet already guarantees these + // buffers are resident, so skip the per-buffer declaration whose count scales with LRU + // cache size (mechanism history v5). Residency sets don't do hazard tracking (Apple docs), + // but none was load-bearing here: every buffer in `use` is MTLResourceUsageRead-only and + // referenced only indirectly (moe_gemv dereferences waddr[]/saddr[] baked into bag/bsg's + // contents), so there's no GPU-side write to serialize against; the one real hazard -- a + // slab unregistered+freed+reused while an async in-flight CB still reads it -- is a + // CPU-write race outside Metal's hazard tracking either way, held by the engine's own slot + // lifecycle, not by useResource:. See SUMMARY.md UNCERTAINTIES. + if (!g_resset_enabled) { + for(auto&b:use) [e useResource:b usage:MTLResourceUsageRead]; + } auto gemv=[&](id wa,id sa,id xin,id y,int O,int K,int Kin){ int NT=R*O; [e setComputePipelineState:g_moe_gemv]; diff --git a/c/coli b/c/coli index 6378ce5..5cef2fc 100755 --- a/c/coli +++ b/c/coli @@ -240,6 +240,43 @@ def env_for(a): gpu=f" · VRAM {format_bytes(vt['budget_bytes'])}" if has_cuda and vt["devices"] else " · CPU" print(f" {C.dim}[PLAN] RAM {format_bytes(rt['budget_bytes'])} · cap {rt['cache_slots_per_layer']}/layer{gpu}{C.r}",file=sys.stderr) else: + # Windows: a bare `coli chat` (no --gpu/--vram/--auto-tier) used to ALWAYS + # run CPU-only, even on a CUDA build with a GPU present — cuda_binary() + # returned False on Windows (see above), and nothing set COLI_CUDA without + # an explicit flag. Now that detection works, auto-enable the GPU when one + # is detected so `coli chat` Just Works. Scoped to Windows: Linux already + # has working detection + the explicit-flag UX, and changing bare-chat + # semantics there is out of scope. Falls back to CPU with a warning if + # nvidia-smi is missing (discover_gpus can't size VRAM without it). + if (sys.platform == "win32" and a.gpu is None and not a.vram): + if cuda_binary(): + from resource_plan import discover_gpus, build_plan, environment_for_plan, format_bytes + gpus = discover_gpus() + if gpus: + e["COLI_CUDA"]="1" + e.setdefault("COLI_GPUS", ",".join(str(g["index"]) for g in gpus)) + # Reuse the planner so the expert-tier VRAM budget is the real + # free VRAM minus the 2 GB reserve — not a guess. Same machinery + # as --auto-tier, just without requiring the user to pass it. + ram,ctx,devices,vram_req = resource_request(a, e) + try: + plan=build_plan(a.model,ram,ctx,devices,vram_req,policy=a.policy) + e.update(environment_for_plan(plan,e,cuda_enabled=True)) + vt=plan["tiers"]["vram"] + names=",".join(g["name"].strip() for g in gpus) + print(f" {C.dim}[GPU] auto-enabled CUDA · {names} · " + f"{format_bytes(vt['budget_bytes'])} expert tier{C.r}", file=sys.stderr) + except (OSError,ValueError,json.JSONDecodeError) as error: + # Plan failed (e.g. model dir unreadable): don't block the + # run, just leave the unsized COLI_CUDA=1 and let the engine + # pick its own budget. Engine handles a missing budget. + print(f" {C.yel}[GPU] auto-enable: could not size VRAM ({error}); " + f"using engine default{C.r}", file=sys.stderr) + else: + print(f" {C.yel}[GPU] coli_cuda.dll present but nvidia-smi not found on PATH " + f"(cannot size VRAM); running CPU-only. Add nvidia-smi to PATH or pass " + f"--vram N to enable CUDA.{C.r}", file=sys.stderr) + # else: CPU build (no coli_cuda.dll) — stay silent, CPU is correct. # --gpu/--vram SENZA --auto-tier: prima venivano ignorati in silenzio e il run # partiva CPU-only senza alcun avviso — benchmark "GPU" pubblicati per errore (#121). if a.gpu is not None: diff --git a/c/colibri.c b/c/colibri.c index e5e08f3..4a4b2e0 100644 --- a/c/colibri.c +++ b/c/colibri.c @@ -69,10 +69,12 @@ static inline int omp_get_thread_num(void){ return 0; } #include static int g_metal_enabled; static int g_metal_gemm_min=16; /* COLI_METAL_GEMM_MIN: min rows to send a matmul_qt GEMM to GPU */ -/* routing precalcolata dalla GPU (layer CB): moe() la usa e salta la FASE A */ -static const int *g_pre_idx; static const float *g_pre_w; static const int *g_pre_keff; -static const float *g_pre_sh; /* output dello shared expert gia' calcolato su GPU */ +/* output dello shared expert gia' calcolato su GPU (solo Metal layer-CB) */ +static const float *g_pre_sh; #endif +/* routing precalcolata dalla GPU (Metal layer CB o device router CUDA, #431): + * moe() la usa e salta la FASE A. NULL = router su CPU. */ +static const int *g_pre_idx; static const float *g_pre_w; static const int *g_pre_keff; #ifdef __APPLE__ #include /* host_statistics64: MemAvailable di macOS */ #endif @@ -92,7 +94,13 @@ typedef struct { * fmt=1 INT8 -> q8 (1 byte/param) + scala per riga * fmt=2 INT4 -> q4 (2 valori per byte, impacchettati) + scala per riga * INT4 e' cio' che fa stare la densa residente nei 15 GB (0.5 byte/param). */ -/* fmt: 0 F32, 1 INT8, 2 INT4 (2/byte), 3 INT2 (4/byte). q4 ospita sia int4 che int2 packed. */ +/* fmt: 0 F32, 1 INT8, 2 INT4 (2/byte), 3 INT2 (4/byte), 4 INT4-GROUPED, 5 INT3-G64. + * q4 ospita int4/int2/int3 packed. fmt=4 (grouped int4, #242): per-row nibbles + one f32 + * scale per group of `gs` inputs (s has O*ceil(I/gs) entries). + * fmt=5 (int3, per-GROUP scales, group=64, see quant.h I3_*): values in [-4,3] stored per + * 64-input group as 24 bytes = 16B low plane (2 bits/val, int2 layout) + 8B high plane + * (1 bit/val), plus ONE f32 scale PER GROUP (s has O*ceil(I/64) entries, not O). 3.5 + * bits/weight effective — the quality/size sweet spot measured in the #132 ablation. */ typedef struct { int fmt; float *qf; int8_t *q8; uint8_t *q4; float *s; int O, I, gs; /* gs=group size (0=per-row, 128=grouped) */ #ifdef COLI_CUDA @@ -108,6 +116,10 @@ static int64_t qt_bytes(const QT *t){ /* byte residenti del tensore */ if(t->fmt==4){ /* int4 grouped: packed nibbles + O*ceil(I/gs) scales */ int ng=(t->I+t->gs-1)/t->gs; return (int64_t)t->O*((t->I+1)/2) + (int64_t)t->O*ng*4; } + if(t->fmt==5){ /* int3-g64: 24B/group weights + one f32 scale per group (I3_* in quant.h, + * included below — keep the arithmetic literal here) */ + int64_t ng=((int64_t)t->I+63)/64; + return (int64_t)t->O*ng*24 + (int64_t)t->O*ng*4; } return (int64_t)t->O*((t->I+1)/2) + (int64_t)t->O*4; /* fmt=2 int4 per-row */ } @@ -125,6 +137,10 @@ typedef struct { QT gate_proj, up_proj, down_proj; /* moe (sparse==1) */ float *router, *router_bias; /* router f32 (sensibile) */ +#ifdef COLI_CUDA + void *router_cuda, *router_bias_cuda; /* device router (#431 PR-A), lazy-uploaded */ + int router_cuda_bad; /* upload failed once: stay on the CPU router */ +#endif QT sh_gate, sh_up, sh_down; /* shared expert */ } Layer; @@ -133,7 +149,12 @@ typedef struct { * slab_cap/fslab_cap: capienza allocata — gli slot ws[] sono riusati TRA layer e gli * expert non hanno tutti la stessa taglia (layer MTP int8 = 2x i layer int4). */ typedef struct { int eid; QT g,u,d; uint8_t *slab; float *fslab; - int64_t slab_cap, fslab_cap; uint64_t used; } ESlot; + int64_t slab_cap, fslab_cap; uint64_t used; + /* pin-arena backing (#419): when set, slab/fslab are interior + * slices of a per-layer arena and must never be free()d — + * expert_host_release detaches them, expert_host_ensure + * re-attaches. NULL for every individually-allocated slot. */ + uint8_t *aslab; float *afslab; } ESlot; typedef struct { float **Lc, **Rc, **Ic; @@ -163,11 +184,34 @@ typedef struct { KVState *kv; ESlot **ecache; int *ecn; int ecap; /* LRU expert per-layer */ float **kv_dev_L, **kv_dev_R; int *kv_dev_valid; /* ombra KV su device (decode) */ + float **ln_dev; /* in_ln/post_ln cached on device: [layer*2+{0,1}] (Inc.4) */ ESlot ws[64]; /* working set del layer corrente (load paralleli) */ ESlot **pin; int *npin; /* HOT-STORE: expert pinnati in RAM (mai evicted) */ uint32_t **eusage; /* contatori persistenti (per STATS/PIN) */ uint32_t **eheat; /* calore recente per promotion/demotion live */ uint32_t **elast, eaccess_clock; /* recency per LFRU session-local */ + /* DISK-CLASS: PRIVATE recency state, read only by expert_classify(). Private -- + * not the real elast/eaccess_clock -- kept fully separate so DISK-CLASS's bookkeeping + * can never read from or write into stock eviction state: every DISK-CLASS write lives + * inside its own need_classify/dc_on gate, so "byte-identical with PROF=0" is provable + * by construction instead of by argument. (Historical note: when this was first written, + * the Metal pre-routed FASE A path (g_pre_idx) never bumped the real elast/eaccess_clock + * -- on Metal decode the real clock froze at end of prefill, so REPIN's LRU tie-breaker + * ran on stale recency for the rest of the run. That was an upstream defect; it has since + * been reported and fixed (#417, cfcc742) -- FASE A now bumps the real clock too. The + * private clock is retained anyway: separation from stock state is the stronger property, + * independent of whether the real clock is correct.) elast_dc/eaccess_clock_dc tick in + * BOTH FASE A paths, under the same need_classify gate, at the same rate the real clock + * ticks on the CPU path (one per selected (position,expert)) -- so the + * COLI_DISKCLASS_WINDOW window keeps its meaning in every mode. elast_pre snapshots + * elast_dc just BEFORE this call's own bump (see the touched[] guard in FASE A) -- + * classifying against the live array would read the bump routing just made a few lines + * above the load that needed it, so a giant cold prefill burst would score every expert + * "just accessed" and get called warm. Recency alone (not eheat's access COUNT): a count + * never decays, so an expert hot early in a long session would keep reading "warm" long + * after it dropped out of the working set. Same shape/allocation as elast; NULL for dense + * layers. */ + uint32_t **elast_dc, **elast_pre, eaccess_clock_dc; /* DSA lightning indexer (attivo solo se i pesi out-idx-* sono presenti) */ int has_dsa; QT *ix_wq, *ix_wk, *ix_wp; /* per layer FULL: wq_b, wk, weights_proj */ @@ -212,9 +256,17 @@ static inline int spec_pinned(void){ return g_spec_pin && g_spec_live; } static void matmul_qt_ex(float *y, const float *x, QT *w, int S, int allow_idot); static void matmul_qt(float *y, const float *x, QT *w, int S){ matmul_qt_ex(y,x,w,S,1); } +/* fmt=4 fused gate+up (defined later, after the quant kernels) */ +static void matmul_i4_grouped_pair(float *yg, float *yu, const float *x, + const uint8_t *qg, const float *sg, + const uint8_t *qu, const float *su, + int S, int I, int O, int gs); + static void expert_gate_up(float *g,float *u,const float *x,QT *wg,QT *wu,int S){ if(!g_no_fused_pair&&!spec_pinned()&&S==1&&wg->fmt==2&&wu->fmt==2&&wg->I==wu->I&&wg->O==wu->O) matmul_i4_pair(g,u,x,wg->q4,wg->s,wu->q4,wu->s,wg->I,wg->O); + else if(!g_no_fused_pair&&S==1&&wg->fmt==4&&wu->fmt==4&&wg->I==wu->I&&wg->O==wu->O&&wg->gs==wu->gs) + matmul_i4_grouped_pair(g,u,x,wg->q4,wg->s,wu->q4,wu->s,S,wg->I,wg->O,wg->gs); else { matmul_qt(g,x,wg,S); matmul_qt(u,x,wu,S); } } @@ -233,8 +285,14 @@ static void qt_cuda_reset(QT *t){ t->cuda_failed=0; } static int qt_cuda_upload(QT *t){ + if(t->fmt==5) return 0; /* int3-g64: no CUDA kernel yet — tensor stays CPU-side */ const void *weights = t->fmt==0 ? (const void*)t->qf : t->fmt==1 ? (const void*)t->q8 : (const void*)t->q4; + if(t->fmt==4) /* grouped int4 (#334): scales are [O, ceil(I/gs)] — the plain + * upload would truncate them to O floats and the group kernels + * would read garbage. An old DLL without the _g symbol returns 0 + * and the tensor simply stays CPU-side. */ + return coli_cuda_tensor_upload_g(&t->cuda,weights,t->s,t->fmt,t->I,t->O,t->cuda_device,t->gs); return coli_cuda_tensor_upload(&t->cuda,weights,t->s,t->fmt,t->I,t->O,t->cuda_device); } static int qt_cuda_update(QT *t){ @@ -242,6 +300,7 @@ static int qt_cuda_update(QT *t){ t->fmt==1?(const void*)t->q8:(const void*)t->q4; return coli_cuda_tensor_update(t->cuda,weights,t->s); } +static double g_ovl_issue,g_ovl_cpu,g_ovl_take,g_ovl_mark; /* Inc.4 overlap-window split (OVL report) */ static void cuda_stats_print(void){ size_t n=0,b=0; coli_cuda_stats(-1,&n,&b); fprintf(stderr,"[CUDA] resident set: %zu tensors, %.2f GB VRAM\n",n,b/1e9); @@ -257,6 +316,9 @@ static void cuda_stats_print(void){ getenv("COLI_CUDA_PROFILE")?"; timing sotto":""); if(calls&&getenv("COLI_CUDA_PROFILE")) fprintf(stderr, "[CUDA] expert groups timing: H2D %.1f ms | kernel %.1f ms | D2H %.1f ms\n",h2d,kernel,d2h); + if(g_ovl_issue+g_ovl_cpu+g_ovl_take>0) fprintf(stderr, + "[CUDA] overlap window: pack+issue %.2fs | cpu-rows %.2fs | take(sync+acc) %.2fs\n", + g_ovl_issue,g_ovl_cpu,g_ovl_take); } static int parse_cuda_devices(const char *list, int *out){ if(!list||!*list) return 0; @@ -298,6 +360,65 @@ static _Atomic int64_t g_prof_io; /* bytes pread()/faulted from e * wait ~ service means the loads block the compute thread. */ static _Atomic int64_t g_edisk_ns; static double edisk_s(void){ return atomic_load_explicit(&g_edisk_ns,memory_order_relaxed)*1e-9; } +/* DISK-CLASS (PROF=1): per-load cold/warm classification against the engine's own + * recency state. Instrumentation only -- it never changes which fd serves a read (see + * expert_classify() and its call site in expert_load_impl; the fd choice expression is + * untouched by this feature). COLI_DISKCLASS_WINDOW is the recency window in ticks of the + * PRIVATE clock (m->eaccess_clock_dc -- NOT the real eaccess_clock; see elast_dc in Model + * for why DISK-CLASS keeps its own clock instead of reading the real one): one tick per + * selected (position,expert) in FASE A while classification is active, the same per-token + * rate the real clock has on the CPU path, so the window's meaning is unchanged. At or + * under the window = warm; 0 (default, unset) derives it from topk*n_layers*8 once the + * model config is known (main(), right after model_init) -- roughly "seen in the last ~8 + * tokens", generous on purpose (conservative-toward-warm: a load the page cache could have + * served that gets labeled cold overstates the cold class, the bucket this line exists to + * size). */ +static uint32_t g_direct_heat_ticks=0; +static int g_direct_heat_explicit=0; /* 1 if COLI_DISKCLASS_WINDOW was set (skip the auto-derive) */ +#define DC_COLD 0 +#define DC_WARM 1 +static _Atomic uint64_t g_dc_n[2]; /* [DC_COLD]/[DC_WARM]: loads classified */ +static _Atomic int64_t g_dc_bytes[2]; /* bytes read (weights + scales, matches g_prof_io) */ +static _Atomic int64_t g_dc_ns[2]; /* wall ns spent reading (thread-seconds, like g_edisk_ns) */ +static _Atomic uint64_t g_dc_direct_n[2]; /* subset of the above ACTUALLY served by the uncached fd */ +/* Busy-wall per class + combined: how much WALL time had >=1 classified load of the + * class in flight (thread-seconds / busy-wall = average concurrency; bytes / busy-wall + * = aggregate GB/s the disk actually delivered for that class -- the quantity the + * thread-second numbers alone can't answer: N slow overlapped reads can beat N fast + * serial ones in aggregate, and only wall-denominated rates see it). Transition scheme: + * 0->1 records a start, 1->0 accumulates (now - start). One dedicated mutex serializes + * the transition bookkeeping -- two short lock/unlock pairs per load against ms-scale + * reads; a CAS scheme would save nothing measurable and be harder to audit + * (correctness over cleverness). Only COMPLETED intervals are in the accumulators: an + * interval still open at report time is not counted (bounded by one read's duration -- + * noise at report granularity). */ +static pthread_mutex_t g_dc_wall_mx=PTHREAD_MUTEX_INITIALIZER; +static int g_dc_inflight[2], g_dc_inflight_all; /* guarded by g_dc_wall_mx */ +static double g_dc_wall_open[2], g_dc_wall_open_all; /* start of the open interval */ +static int64_t g_dc_wall_ns[2], g_dc_wall_all_ns; /* completed busy-wall ns */ +static void dc_wall_enter(int cls, double now){ + pthread_mutex_lock(&g_dc_wall_mx); + if(g_dc_inflight[cls]++==0) g_dc_wall_open[cls]=now; + if(g_dc_inflight_all++==0) g_dc_wall_open_all=now; + pthread_mutex_unlock(&g_dc_wall_mx); +} +static void dc_wall_exit(int cls, double now){ + pthread_mutex_lock(&g_dc_wall_mx); + if(--g_dc_inflight[cls]==0) g_dc_wall_ns[cls]+=(int64_t)((now-g_dc_wall_open[cls])*1e9); + if(--g_dc_inflight_all==0) g_dc_wall_all_ns +=(int64_t)((now-g_dc_wall_open_all)*1e9); + pthread_mutex_unlock(&g_dc_wall_mx); +} +static int dc_needed(void); /* fwd: defined with the classifier (needs g_prof) */ +static void dc_wall_read(int64_t out[2], int64_t *all){ /* mutex-consistent snapshot for the report */ + if(!dc_needed()){ out[0]=out[1]=0; *all=0; return; } /* off for the whole process => accumulators are + * provably zero (only dc_wall_exit writes them, + * only under dc_on): skip the lock, zero work. + * Needed because prof_base runs unconditionally + * at some call sites ("cheap enough to always"). */ + pthread_mutex_lock(&g_dc_wall_mx); + out[0]=g_dc_wall_ns[0]; out[1]=g_dc_wall_ns[1]; *all=g_dc_wall_all_ns; + pthread_mutex_unlock(&g_dc_wall_mx); +} #define PROF_LAT_CAP 32768 static double g_prof_lat[PROF_LAT_CAP]; /* per-forward decode wall clock (ring) */ static uint64_t g_prof_nlat; /* forwards recorded (monotonic) */ @@ -307,6 +428,8 @@ typedef struct { double edisk,ewait,emm,ecpu,egpu,route,p2p,attn,head; int64_t io,cpu_bytes; uint64_t hits,miss,ereq,n_fw,n_emit,nlat,n_p2p,cpu_rows; uint64_t hit_pin,hit_ecache; + uint64_t dc_n[2], dc_direct_n[2]; int64_t dc_bytes[2], dc_ns[2]; /* DISK-CLASS */ + int64_t dc_wall_ns[2], dc_wall_all_ns; /* busy-wall (per class + combined) */ } ProfBase; static void prof_base(Model *m, ProfBase *b){ b->edisk=edisk_s(); b->ewait=m->t_ewait; b->emm=m->t_emm; @@ -317,6 +440,13 @@ static void prof_base(Model *m, ProfBase *b){ b->hit_pin=m->hit_pin; b->hit_ecache=m->hit_ecache; b->n_fw=m->n_fw; b->n_emit=m->n_emit; b->nlat=g_prof_nlat; b->n_p2p=m->n_p2p; b->cpu_bytes=m->cpu_expert_bytes;b->cpu_rows=m->cpu_expert_rows; + for(int i=0;i<2;i++){ + b->dc_n[i]=atomic_load_explicit(&g_dc_n[i],memory_order_relaxed); + b->dc_bytes[i]=atomic_load_explicit(&g_dc_bytes[i],memory_order_relaxed); + b->dc_ns[i]=atomic_load_explicit(&g_dc_ns[i],memory_order_relaxed); + b->dc_direct_n[i]=atomic_load_explicit(&g_dc_direct_n[i],memory_order_relaxed); + } + dc_wall_read(b->dc_wall_ns,&b->dc_wall_all_ns); } static float *falloc(int64_t n){ @@ -326,6 +456,70 @@ static float *falloc(int64_t n){ float *p=malloc((size_t)n*sizeof(float)); if(!p){fprintf(stderr,"OOM\n");exit(1);} return p; } + +/* Fused gate+up for grouped int4 (fmt=4): computes both yg[S,O] and yu[S,O] from + * the same x[S,I], reading x once instead of twice — saves ~33% of expert-matmul time at decode. + * The per-group scale logic matches matmul_i4_grouped exactly. */ +static void matmul_i4_grouped_pair(float *yg, float *yu, const float *x, + const uint8_t *qg, const float *sg, + const uint8_t *qu, const float *su, + int S, int I, int O, int gs){ + int rb=(I+1)/2; int ng=(I+gs-1)/gs; + #pragma omp parallel for schedule(static) + for(int o=0;oI) glen=I-base; + float scg=sgl[g], scu=sul[g]; + int i=base; +#ifdef __AVX2__ + const __m128i m4=_mm_set1_epi8(0x0F); const __m256i b8=_mm256_set1_epi32(8); + __m256 accg=_mm256_setzero_ps(), accu=_mm256_setzero_ps(); + for(; i+16<=base+glen; i+=16){ + __m128i byg=_mm_loadl_epi64((const __m128i*)(wg+(i>>1))); + __m128i log=_mm_and_si128(byg,m4),hig=_mm_and_si128(_mm_srli_epi16(byg,4),m4); + __m128i nibg=_mm_unpacklo_epi8(log,hig); + __m256 w0g=_mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_cvtepu8_epi32(nibg),b8)); + __m256 w1g=_mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_srli_si128(nibg,8)),b8)); + accg=_mm256_fmadd_ps(_mm256_loadu_ps(xs+i), w0g, accg); + accg=_mm256_fmadd_ps(_mm256_loadu_ps(xs+i+8), w1g, accg); + __m128i byu=_mm_loadl_epi64((const __m128i*)(wu2+(i>>1))); + __m128i lou=_mm_and_si128(byu,m4),hiu=_mm_and_si128(_mm_srli_epi16(byu,4),m4); + __m128i nibu=_mm_unpacklo_epi8(lou,hiu); + __m256 w0u=_mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_cvtepu8_epi32(nibu),b8)); + __m256 w1u=_mm256_cvtepi32_ps(_mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_srli_si128(nibu,8)),b8)); + accu=_mm256_fmadd_ps(_mm256_loadu_ps(xs+i), w0u, accu); + accu=_mm256_fmadd_ps(_mm256_loadu_ps(xs+i+8), w1u, accu); + } + ag+=hsum256(accg)*scg; au+=hsum256(accu)*scu; +#endif + for(; i+1>1], bu=wu2[i>>1]; + ag+=(xs[i]*(float)((int)(bg&0xF)-8)+xs[i+1]*(float)((int)(bg>>4)-8))*scg; + au+=(xs[i]*(float)((int)(bu&0xF)-8)+xs[i+1]*(float)((int)(bu>>4)-8))*scu; + } + if(i>1], bu=wu2[i>>1]; + ag+=xs[i]*(float)((int)(bg&0xF)-8)*scg; + au+=xs[i]*(float)((int)(bu&0xF)-8)*scu; + } + } + yg[(int64_t)s*O+o]=ag; yu[(int64_t)s*O+o]=au; + } + } +} + +/* allow_idot=0: forza il kernel int4/int8 ESATTO (attivazioni f32). Serve alle proiezioni di + * attenzione: sono sensibili alla quantizzazione int8 delle attivazioni dell'IDOT. Misurato su + * GLM-5.2 int4, 1023 token, log-lik -5040.33 (esatto) -> -5160.47 (IDOT) = +0.117 nat/token, + * ~+12% perplexity. Gli altri matmul del prefill (o_proj, kv_b, expert) tengono l'IDOT. + * EN: allow_idot=0 forces the EXACT int4/int8 kernel (f32 activations). The attention + * projections need it: IDOT's int8 activation quantization costs +0.117 nats/token there + * (~+12% perplexity), measured. Every other prefill matmul keeps IDOT as before. */ static void matmul_qt_ex(float *y, const float *x, QT *w, int S, int allow_idot){ #ifdef COLI_METAL if(g_metal_enabled && S>=g_metal_gemm_min && !spec_pinned() && (w->fmt==1||w->fmt==2) && !omp_in_parallel()){ @@ -334,10 +528,10 @@ static void matmul_qt_ex(float *y, const float *x, QT *w, int S, int allow_idot) } #endif #ifdef COLI_CUDA - if(g_cuda_enabled && w->cuda_eligible && !w->cuda_failed && !omp_in_parallel()){ + if(g_cuda_enabled && w->cuda_eligible && !w->cuda_failed && w->fmt!=5 && !omp_in_parallel()){ const void *weights = w->fmt==0 ? (const void*)w->qf : w->fmt==1 ? (const void*)w->q8 : (const void*)w->q4; - if(coli_cuda_matmul(&w->cuda,y,x,weights,w->s,w->fmt,S,w->I,w->O,w->cuda_device)) return; + if(coli_cuda_matmul(&w->cuda,y,x,weights,w->s,w->fmt,S,w->I,w->O,w->cuda_device,w->gs)) return; w->cuda_failed=1; fprintf(stderr,"[CUDA] tensor [%d,%d] on device %d disabled after an error; falling back to CPU\n", w->O,w->I,w->cuda_device); @@ -356,6 +550,7 @@ static void matmul_qt_ex(float *y, const float *x, QT *w, int S, int allow_idot) } if(w->fmt==1) matmul_q(y,x,w->q8,w->s,S,w->I,w->O); else if(w->fmt==3) matmul_i2(y,x,w->q4,w->s,S,w->I,w->O); + else if(w->fmt==5) matmul_i3(y,x,w->q4,w->s,S,w->I,w->O); else matmul_i4(y,x,w->q4,w->s,S,w->I,w->O); } @@ -401,12 +596,6 @@ static int g_draft=0; /* metodo E: DRAFT=n token auto-speculati per forward v * parte (#8). MAI un vincolo sul sampling: solo proposte, la verifica batch-union * decide — grammatica sbagliata = draft rifiutati, output identico. * GRAMMAR_DRAFT=n (default 24) limita i token forzati per forward. */ -static Grammar g_gram; static GrState g_gst; -static Tok *g_gr_T=NULL; -static int g_gr_on=0; /* grammatica caricata e walker vivo */ -static int g_gr_armed=0; /* lazy: parte dal primo byte ammesso dalla radice (salta i preamboli) */ -static int g_gr_max=24; -static uint64_t g_gr_prop=0, g_gr_acc=0; static FILE *g_route_fp=NULL; /* ROUTE_TRACE=: dump per-position top-K routing (ids:gates) * per layer — offline co-activation / coupling analysis. Zero * effect on computation; measurement only. */ @@ -426,6 +615,19 @@ static int g_couple=0, g_couple_k=8, g_couple_d=1; static int16_t *cp_pred=NULL; /* [(L*2+(dL-1))*E + e]*CP_M + j -> target id (-1 none) */ static float *cp_cnt=NULL; static long g_cp_enq=0; +/* All grammar-forced-draft state in one struct so it can become per-request + * in the multiplexed server. Fields (same semantics as the former globals): + * on = grammar loaded and walker alive; armed = lazy start from the first byte + * accepted at the root (skips preambles); max = forced-span cap per forward; + * prop/acc = proposed/accepted forced-draft counters. */ +typedef struct { + Grammar gram; + GrState st; + Tok *T; + int on, armed, max; + uint64_t prop, acc; +} GrDraft; +static GrDraft g_grd={.max=24}; /* process-level instance: PROMPT mode + run_serve keep using this */ static void couple_prefetch(Model *m, int layer, const int *idx, int Ke); static int g_looka=0; /* LOOKA=1: misura (solo contatori, zero effetti) quanto il routing MoE * e' predicibile IN ANTICIPO — la domanda che decide se un prefetch @@ -459,19 +661,31 @@ static int g_disk_split=0; /* DISK_SPLIT=1: contatori che spezzano i DISK LOAD ( * +40% on a 4-socket (#82). Blanket `numactl --interleave=all` is NOT equivalent: * it also interleaves the CUDA pinned staging buffers and cost a 4-socket GPU host * 10x (#82) — hence per-region mbind here and nothing else. Raw syscall, no libnuma - * dependency; MPOL_MF_MOVE migrates pages of reused heap chunks too. Linux-only, - * silent no-op elsewhere or on single-node hosts. */ + * dependency. Linux-only, silent no-op elsewhere or on single-node hosts. + * + * VMA discipline (#419): every mbind carries its own memory policy, so a bound + * region cannot merge with its neighbours — measured ~2 VMAs per slab, with or + * without MPOL_MF_MOVE. Per-slab binds on a PIN_GB=all load (19,456 experts x + * slab+fslab) cross the default vm.max_map_count=65530 and posix_memalign dies + * with terabytes free. So the bulk (the pinned hot-store) is bound as ONE arena + * per layer (see pin_load), and per-slab mbind remains only for the bounded + * allocations: dense qalloc, the LRU ecache, and GPU-tier staging. No flag: + * every bind here lands before the pread that first-touches the pages, so + * there is nothing to migrate. */ #ifdef __linux__ static int g_numa_nodes=0; /* only touched under __linux__; off-Linux NUMA is a no-op */ +static int g_numa_skip_bind=0; /* raised around the GPU-prefix pin load: those slabs are + * upload staging, freed right after — binding them buys + * nothing and costs ~2 transient VMAs each (#419) */ #endif static void numa_slab_bind(void *p, size_t n){ #ifdef __linux__ - if(g_numa_nodes<2 || !p || !n) return; + if(g_numa_nodes<2 || g_numa_skip_bind || !p || !n) return; unsigned long mask=(1UL<=2) fprintf(stderr,"[NUMA] expert slabs interleaved across %d nodes\n",g_numa_nodes); - else fprintf(stderr,"[NUMA] single node: COLI_NUMA ignored\n"); + if(g_numa_nodes<2){ fprintf(stderr,"[NUMA] single node: COLI_NUMA ignored\n"); return; } + /* Probe mbind once so a constrained container degrades with a message + * instead of silently losing the interleave. The probe page must be + * page-aligned (mbind rejects unaligned addresses with EINVAL) and only + * errno==EPERM disables — any other failure keeps NUMA on. */ + { void *pg=NULL; + if(!posix_memalign(&pg,4096,4096)){ + unsigned long mask=(1UL<=16 f32, 5..8 int8, <=4 int4-packed */ +/* format from `bits`: >=16 f32, 5..8 int8, 4 int4-packed, 3 int3-g64 (group scales), <=2 int2 */ static void qt_alloc(QT *t, int O, int I, int bits){ t->O=O; t->I=I; t->qf=NULL; t->q8=NULL; t->q4=NULL; t->s=NULL; if(bits>=16){ t->fmt=0; t->qf=falloc((int64_t)O*I); } else if(bits>=5 || g_nopack){ t->fmt=1; t->q8=qalloc((int64_t)O*I); t->s=qsalloc(O); } - else if(bits>=3){ t->fmt=2; t->q4=qalloc((int64_t)O*((I+1)/2)); t->s=qsalloc(O); } + else if(bits>=4){ t->fmt=2; t->q4=qalloc((int64_t)O*((I+1)/2)); t->s=qsalloc(O); } + else if(bits==3){ t->fmt=5; t->q4=qalloc((int64_t)O*i3_rowbytes(I)); + t->s=(float*)qalloc((size_t)O*i3_groups(I)*sizeof(float)); } else { t->fmt=3; t->q4=qalloc((int64_t)O*((I+3)/4)); t->s=qsalloc(O); } } static void qt_fill(QT *t, const float *w, int bits){ if(t->fmt==0) memcpy(t->qf, w, (int64_t)t->O*t->I*sizeof(float)); else if(t->fmt==1) quantize_rows(w, t->q8, t->s, t->O, t->I, bits); else if(t->fmt==3) pack_int2(w, t->q4, t->s, t->O, t->I, bits); + else if(t->fmt==5) pack_int3_g64(w, t->q4, t->s, t->O, t->I); else pack_int4(w, t->q4, t->s, t->O, t->I, bits); } @@ -700,16 +934,22 @@ static int detect_group_size(int O, int I, int64_t ns){ * diventava un int2 valido e il matmul leggeva oltre il buffer (O*I nibble a * 4/byte). Qui i byte del peso devono corrispondere a un layout noto e i byte * della scala alla cardinalita' attesa (O per-row, O*ng per-gruppo) — altrimenti - * si termina invece di sforare. Ritorna fmt (1/2/3/4) e scrive *gs. */ + * si termina invece di sforare. Ritorna fmt (1/2/3/4/5) e scrive *gs. */ static int qt_resolve_fmt(const char *name, int O, int I, int64_t nb, int64_t ns, int *gs){ int64_t exp_i8=(int64_t)O*I, exp_i4=(int64_t)O*((I+1)/2), exp_i2=(int64_t)O*((I+3)/4); - int fmt = (nb==exp_i8)?1 : (nb==exp_i4)?2 : (nb==exp_i2)?3 : 0; + int64_t exp_i3=(int64_t)O*i3_rowbytes(I); /* int3-g64 (fmt=5): 24B per 64-input group */ + /* Row formats take precedence: for tiny I the int3-g64 byte count can coincide with + * a row layout (e.g. [O,48]: ceil(48/2)=24=1*24). For real tensor shapes the counts + * are distinct, and the weight bytes — not the scale size — are the int3 tag, because + * int3-g64 and grouped-int4-at-gs=64 carry the SAME scale cardinality O*ceil(I/64). */ + int fmt = (nb==exp_i8)?1 : (nb==exp_i4)?2 : (nb==exp_i2)?3 : (nb==exp_i3)?5 : 0; if(!fmt){ - fprintf(stderr,"%s: quantized weight is %lld bytes — no int8/int4/int2 layout for [%d,%d], refusing (untrusted container)\n", + fprintf(stderr,"%s: quantized weight is %lld bytes — no int8/int4/int2/int3-g64 layout for [%d,%d], refusing (untrusted container)\n", name,(long long)nb,O,I); exit(1); } *gs=0; if(fmt==2){ int g=detect_group_size(O,I,ns); if(g>0){ fmt=4; *gs=g; } } - int64_t exp_scale = (fmt==4)? (int64_t)O*((I+*gs-1)/(*gs)) : (int64_t)O; /* in FLOAT */ + int64_t exp_scale = (fmt==4)? (int64_t)O*((I+*gs-1)/(*gs)) + : (fmt==5)? (int64_t)O*i3_groups(I) : (int64_t)O; /* in FLOAT */ if(ns != exp_scale*4){ fprintf(stderr,"%s: scale array is %lld bytes — expected %lld for [%d,%d] fmt=%d, refusing (untrusted container)\n", name,(long long)ns,(long long)(exp_scale*4),O,I,fmt); exit(1); } @@ -733,6 +973,9 @@ static void qt_from_disk(Model *m, const char *name, int O, int I, int bits, int else if(fmt==4){ int ng=(I+gs-1)/gs; if(t->fmt!=4||!t->q4){ t->fmt=4; t->O=O; t->I=I; t->gs=gs; t->q4=qalloc(nb); t->s=falloc((int64_t)O*ng); } st_read_raw(&m->S,name,t->q4,drop); } + else if(fmt==5){ int64_t ng=i3_groups(I); /* int3-g64: 24B/group weights + O*ng group scales */ + if(t->fmt!=5||!t->q4){ t->fmt=5; t->O=O; t->I=I; t->gs=0; t->q4=qalloc(nb); t->s=falloc((int64_t)O*ng); } + st_read_raw(&m->S,name,t->q4,drop); } else { if(t->fmt!=fmt||!t->q4){ t->fmt=fmt; t->O=O; t->I=I; t->gs=0; t->q4=qalloc(nb); t->s=qsalloc(O); } st_read_raw(&m->S,name,t->q4,drop); } st_read_f32(&m->S,sn,t->s,drop); } else { @@ -770,13 +1013,14 @@ static void qt_cuda_colocate(QT *dst,const QT *src){ } static void layer_cuda_shard_kvb(Layer *l,int H,int Q,int V){ if(!g_cuda_enabled||!g_cuda_dense||g_cuda_ndev<2||l->kv_b.fmt==0)return; - int rb=l->kv_b.fmt==1?l->kv_b.I:(l->kv_b.fmt==2?(l->kv_b.I+1)/2:(l->kv_b.I+3)/4); + int rb=l->kv_b.fmt==1?l->kv_b.I: + (l->kv_b.fmt==2||l->kv_b.fmt==4)?(l->kv_b.I+1)/2:(l->kv_b.I+3)/4; const uint8_t *weights=l->kv_b.fmt==1?(const uint8_t*)l->kv_b.q8:l->kv_b.q4; for(int d=0,h0=0;dkv_b.s+(int64_t)h0*(Q+V); - if(!coli_cuda_tensor_upload(&l->kv_b_shard[d],part,scale,l->kv_b.fmt,l->kv_b.I,rows,g_cuda_devices[d]))return; + const float *scale=l->kv_b.s+(int64_t)h0*(Q+V)*(l->kv_b.gs>0?(l->kv_b.I+l->kv_b.gs-1)/l->kv_b.gs:1); + if(!coli_cuda_tensor_upload_g(&l->kv_b_shard[d],part,scale,l->kv_b.fmt,l->kv_b.I,rows,g_cuda_devices[d],l->kv_b.gs))return; l->shard_h0[d]=h0;l->shard_hn[d]=hn;l->n_kv_b_shard++;h0+=hn; } int old=-1;for(int i=0;ikv_b.cuda_device)old=i; @@ -804,6 +1048,7 @@ static void model_init(Model *m, const char *snap, int cap, int ebits, int dbits m->pin=calloc(NR,sizeof(ESlot*)); m->npin=calloc(NR,sizeof(int)); m->eusage=calloc(NR,sizeof(uint32_t*)); m->eheat=calloc(NR,sizeof(uint32_t*)); m->elast=calloc(NR,sizeof(uint32_t*)); + m->elast_dc=calloc(NR,sizeof(uint32_t*)); m->elast_pre=calloc(NR,sizeof(uint32_t*)); m->kv=calloc(1,sizeof(KVState)); m->kv_start=m->kv->kv_start=calloc(NR,sizeof(int)); for(int i=0;in_layers;i++){ @@ -848,6 +1093,8 @@ static void model_init(Model *m, const char *snap, int cap, int ebits, int dbits m->eusage[i]=calloc(c->n_experts,sizeof(uint32_t)); m->eheat[i]=calloc(c->n_experts,sizeof(uint32_t)); m->elast[i]=calloc(c->n_experts,sizeof(uint32_t)); + m->elast_dc[i]=calloc(c->n_experts,sizeof(uint32_t)); + m->elast_pre[i]=calloc(c->n_experts,sizeof(uint32_t)); } #undef P } @@ -860,12 +1107,17 @@ static void model_init(Model *m, const char *snap, int cap, int ebits, int dbits "self_attn.q_a_proj.weight","self_attn.q_b_proj.weight","self_attn.kv_a_proj_with_mqa.weight", "self_attn.kv_b_proj.weight","self_attn.o_proj.weight","mlp.gate.weight", "mlp.shared_experts.gate_proj.weight","mlp.shared_experts.down_proj.weight", - "mlp.experts.0.gate_proj.weight","mlp.experts.255.down_proj.weight"}; + "mlp.experts.0.gate_proj.weight"}; char mn[256]; m->has_mtp=1; for(unsigned q=0;qn_layers,req[q]); if(!st_has(&m->S,mn)){ m->has_mtp=0; break; } } + /* probe the LAST expert by index, not a fixed 255: REAP-pruned + * checkpoints have n_routed_experts < 256 and the MTP set stays complete, + * so a hardcoded expert.255 would spuriously report has_mtp=0 on them. */ + snprintf(mn,sizeof(mn),"model.layers.%d.mlp.experts.%d.down_proj.weight",c->n_layers,c->n_experts-1); + if(!st_has(&m->S,mn)) m->has_mtp=0; if(getenv("MTP") && atoi(getenv("MTP"))==0) m->has_mtp=0; if(m->has_mtp){ int i=c->n_layers; Layer *l=&m->mtpL; @@ -894,6 +1146,8 @@ static void model_init(Model *m, const char *snap, int cap, int ebits, int dbits m->eusage[i]=calloc(c->n_experts,sizeof(uint32_t)); m->eheat[i]=calloc(c->n_experts,sizeof(uint32_t)); m->elast[i]=calloc(c->n_experts,sizeof(uint32_t)); + m->elast_dc[i]=calloc(c->n_experts,sizeof(uint32_t)); + m->elast_pre[i]=calloc(c->n_experts,sizeof(uint32_t)); m->kv_start[i]=-1; /* KV MTP: parte dalla prima posizione di decode */ #undef PM } @@ -948,12 +1202,27 @@ static void model_init(Model *m, const char *snap, int cap, int ebits, int dbits static void embed_row(Model *m, int tok, float *x){ int D=m->c.hidden; QT *e=&m->embed; if(e->fmt==0){ memcpy(x, e->qf+(int64_t)tok*D, D*sizeof(float)); return; } + if(e->fmt==4){ /* grouped int4: per-group scale (embed/lm_head at io_bits, usually fmt 0/1) */ + const uint8_t *q=e->q4+(int64_t)tok*((D+1)/2); int gs=e->gs,ng=(D+gs-1)/gs; + const float *scl=e->s+(int64_t)tok*ng; + for(int g=0;g*gsD)glen=D-base; float s=scl[g]; + for(int i=base;i+1>1]; x[i]=(float)((int)(byte&0xF)-8)*s; + x[i+1]=(float)((int)(byte>>4)-8)*s; } + if(glen&1){ uint8_t byte=q[(base+glen-1)>>1]; x[base+glen-1]=(float)((int)(byte&0xF)-8)*s; } } + return; } if(e->fmt==1){ const int8_t *q=e->q8+(int64_t)tok*D; float s=e->s[tok]; for(int i=0;ifmt==2){ const uint8_t *q=e->q4+(int64_t)tok*((D+1)/2); float s=e->s[tok]; /* int4 */ for(int i=0;i>1]; x[i]=(float)((int)(byte&0xF)-8)*s; if(i+1>4)-8)*s; } return; } + if(e->fmt==5){ const uint8_t *q=e->q4+(int64_t)tok*i3_rowbytes(D); /* int3-g64 */ + const float *sr=e->s+(int64_t)tok*i3_groups(D); int64_t ng=i3_groups(D); + for(int64_t g=0; g>2]>>((k&3)*2))&3)|(((hi[k>>3]>>(k&7))&1)<<2); + x[base+k]=(float)((int)u-4)*sr[g]; } } + return; } const uint8_t *q=e->q4+(int64_t)tok*((D+3)/4); float s=e->s[tok]; /* int2 */ for(int i=0;i>2]; int sh=(i&3)*2; x[i]=(float)((int)((byte>>sh)&3)-2)*s; } } @@ -1079,7 +1348,35 @@ static int pread_full(int fd, void *buf, int64_t n, int64_t off, const char *tag } return 0; } -static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal){ +/* DISK-CLASS: is classification active right now? Single point of truth shared by + * moe()'s pre-bump snapshot (FASE A, below) and expert_load_impl's classify-and-read -- + * they must agree, or expert_load_impl reads a snapshot moe() never bothered to write. + * PROF=1 is the only trigger: this feature is measurement-only, the verdict never picks + * an fd (kept as one function anyway so a future policy consumer cannot drift out of + * sync with the snapshot writer by construction). */ +static int dc_needed(void){ return g_prof; } +/* DISK-CLASS: cold/warm verdict for ONE demand load, read from the pre-bump snapshot + * (Model's elast_pre) so routing's OWN bump for THIS call can't contaminate the read -- + * prefill is one giant moe() call where every newly-seen expert gets its `last` bumped + * a few lines above the load that made it "new"; classifying off the live, post-bump + * array would call the whole cold burst warm. Ages against the PRIVATE clock + * (eaccess_clock_dc, see its declaration in Model), NEVER the real eaccess_clock: kept + * separate by design (see elast_dc in Model for why DISK-CLASS keeps its own clock + * instead of reading the real one -- before #417/cfcc742 the real one also froze on the + * Metal pre-routed decode path, which would have made every prefill-touched expert warm + * forever and everything else cold forever; that's fixed upstream now, but DISK-CLASS + * still doesn't read the real clock, for the isolation property, not the freeze). + * Conservative-toward-warm on the one genuinely ambiguous input (missing snapshot -- + * defensive only, elast_pre is allocated everywhere elast is); a demonstrable first-ever + * access (last_pre==0) is not a judgment call, it stays cold regardless of that bias. */ +static int expert_classify(Model *m, int layer, int eid){ + if(!m->elast_pre || !m->elast_pre[layer]) return DC_WARM; /* no snapshot: label as the safe class */ + uint32_t last_pre=m->elast_pre[layer][eid]; + if(last_pre==0) return DC_COLD; /* never touched before this call: certain cold */ + uint32_t age=m->eaccess_clock_dc-last_pre; /* ticks since last access, PRE this call's bump */ + return age>g_direct_heat_ticks ? DC_COLD : DC_WARM; /* '>' not '>=': ties lean warm */ +} +static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal, int demand){ #ifdef COLI_CUDA /* A live REPIN may reuse a GPU-enabled pinned slot for a different expert. * Keep its tier assignment, but invalidate the old device weights. */ @@ -1189,6 +1486,11 @@ static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal){ if(ftot<0 || (uint64_t)ftot > SIZE_MAX/sizeof(float) || posix_memalign((void**)&s->fslab,16384,fb)){ fprintf(stderr,"OOM fslab\n"); if(fatal) exit(1); + /* unregister BEFORE freeing -- a stale g_slabs entry would let resolve() hand + * the GPU a pointer into freed memory (and under COLI_METAL_RESSET=1, leave the + * buffer a permanent residency-set member over it). Ported from e4/metal-heap + * validator fix 6753225; pre-existing gap on main/dev. */ + if(s->slab && g_metal_enabled) coli_metal_unregister(s->slab); compat_aligned_free(s->slab); s->slab=NULL; s->slab_cap=0; /* clean, hidden slot (eid stays -1) */ s->fslab=NULL; s->fslab_cap=0; return -1; } @@ -1210,12 +1512,22 @@ static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal){ numa_slab_bind(s->fslab,(size_t)ftot*sizeof(float)); #endif } + /* DISK-CLASS: classify before the reads; computed unconditionally at dc_on sites so + * the timer (dc_t0) brackets exactly the read work, matching what the GB/s in the + * DISK-CLASS line describes. dc_on gates ALL of it off demand=0 call sites + * (pilot/repin/pin -- never classified, see the call sites) and off PROF=0 runs + * (dc_needed()) -- zero cost, zero behavior change there. The fd choice below is + * NOT influenced by the verdict: this is measurement only. */ + int dc_on = demand && dc_needed(); + int dc_cls = dc_on ? expert_classify(m,layer,eid) : DC_WARM; + double dc_t0 = dc_on ? now_s() : 0; + if(dc_on) dc_wall_enter(dc_cls,dc_t0); /* busy-wall open; EVERY exit path below must pair it */ int ord[3]={0,1,2}; /* ordina per offset nel file */ for(int a=0;a<3;a++) for(int bb=a+1;bb<3;bb++) if(tw[ord[bb]]->offoff){ int t=ord[a]; ord[a]=ord[bb]; ord[bb]=t; } int contig = tw[ord[0]]->fd==tw[ord[1]]->fd && tw[ord[1]]->fd==tw[ord[2]]->fd && tw[ord[0]]->off+tw[ord[0]]->nbytes==tw[ord[1]]->off && tw[ord[1]]->off+tw[ord[1]]->nbytes==tw[ord[2]]->off; - int64_t pos[3]; int done=0; + int64_t pos[3]; int done=0, dc_direct=0; if(contig){ int64_t off0=tw[ord[0]]->off; int dfd = g_direct ? st_direct_fd_rep(&m->S, tw[ord[0]]->fd, rep) : -1; @@ -1225,27 +1537,43 @@ static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal){ ssize_t r=pread(dfd, s->slab, len, base); if(r>=need){ pos[ord[0]]=off0-base; pos[ord[1]]=pos[ord[0]]+tw[ord[0]]->nbytes; - pos[ord[2]]=pos[ord[1]]+tw[ord[1]]->nbytes; done=1; + pos[ord[2]]=pos[ord[1]]+tw[ord[1]]->nbytes; done=1; dc_direct=1; atomic_fetch_add_explicit(&g_mir_bytes[rep],(int64_t)r,memory_order_relaxed); atomic_fetch_add_explicit(&g_mir_nread[rep],1,memory_order_relaxed); } } if(!done){ /* fallback bufferizzato */ - if(mir_pread(&m->S, tw[ord[0]]->fd, rep, s->slab, wtot, off0, "pread expert")){ if(fatal) exit(1); return -1; } + if(mir_pread(&m->S, tw[ord[0]]->fd, rep, s->slab, wtot, off0, "pread expert")){ if(fatal) exit(1); + if(dc_on) dc_wall_exit(dc_cls,now_s()); /* pair the enter on the non-fatal unwind */ + return -1; } pos[ord[0]]=0; pos[ord[1]]=tw[ord[0]]->nbytes; pos[ord[2]]=tw[ord[0]]->nbytes+tw[ord[1]]->nbytes; done=1; } } if(!done){ /* non contigui: 3 pread bufferizzate */ int64_t o=0; for(int a=0;a<3;a++){ int k=ord[a]; - if(mir_pread(&m->S, tw[k]->fd, rep, s->slab+o, tw[k]->nbytes, tw[k]->off, "pread expert")){ if(fatal) exit(1); return -1; } + if(mir_pread(&m->S, tw[k]->fd, rep, s->slab+o, tw[k]->nbytes, tw[k]->off, "pread expert")){ if(fatal) exit(1); + if(dc_on) dc_wall_exit(dc_cls,now_s()); /* pair the enter on the non-fatal unwind */ + return -1; } pos[k]=o; o+=tw[k]->nbytes; } } float *fp[3]; int64_t fo=0; /* scale (piccole) */ for(int k=0;k<3;k++){ - if(mir_pread(&m->S, tq[k]->fd, rep, (char*)(s->fslab+fo), tq[k]->nbytes, tq[k]->off, "pread qs")){ if(fatal) exit(1); return -1; } + if(mir_pread(&m->S, tq[k]->fd, rep, (char*)(s->fslab+fo), tq[k]->nbytes, tq[k]->off, "pread qs")){ if(fatal) exit(1); + if(dc_on) dc_wall_exit(dc_cls,now_s()); /* pair the enter on the non-fatal unwind */ + return -1; } fp[k]=s->fslab+fo; fo+=tq[k]->nbytes/4; } atomic_fetch_add_explicit(&g_prof_io,wtot+fo*4,memory_order_relaxed); + if(dc_on){ /* DISK-CLASS accounting, see dc_needed() */ + double dc_t1=now_s(); /* one clock read for thread-ns AND the wall exit */ + int64_t bytes=wtot+fo*4; + atomic_fetch_add_explicit(&g_dc_n[dc_cls],1,memory_order_relaxed); + atomic_fetch_add_explicit(&g_dc_bytes[dc_cls],bytes,memory_order_relaxed); + atomic_fetch_add_explicit(&g_dc_ns[dc_cls],(int64_t)((dc_t1-dc_t0)*1e9),memory_order_relaxed); + dc_wall_exit(dc_cls,dc_t1); + if(dc_direct) /* which fd ACTUALLY served this class */ + atomic_fetch_add_explicit(&g_dc_direct_n[dc_cls],1,memory_order_relaxed); + } if(g_drop){ /* scarta subito le pagine: evita che la page * cache in pressione strangoli il throughput. * The drop targets the fd of the replica READ. */ @@ -1264,9 +1592,15 @@ static int expert_load_impl(Model *m, int layer, int eid, ESlot *s, int fatal){ } /* Every expert read goes through here: time the whole load (pread/fault + * bookkeeping) on the thread that runs it, into the disk-service counter. */ -static int expert_load(Model *m, int layer, int eid, ESlot *s, int fatal){ +static int expert_load(Model *m, int layer, int eid, ESlot *s, int fatal, int demand){ + /* `demand` marks a routing-driven demand-load (moe()'s PIPE/OMP miss path, where the + * pre-bump elast_pre snapshot moe() just wrote is valid) -- pass 0 from anywhere else + * (pilot speculative loads, repin, startup PIN loading): those never run through THIS + * call's own FASE A, so the snapshot either doesn't apply or was never written for + * them, and DISK-CLASS deliberately leaves them unclassified -- see expert_classify()'s + * call site. */ double t0=now_s(); - int rc=expert_load_impl(m,layer,eid,s,fatal); + int rc=expert_load_impl(m,layer,eid,s,fatal,demand); atomic_fetch_add_explicit(&g_edisk_ns,(int64_t)((now_s()-t0)*1e9),memory_order_relaxed); return rc; } @@ -1441,8 +1775,11 @@ static int uring_finalize_load(UringBatch *b,int li,int publish_eid){ for(int k=0;k<3;k++){ fp[k]=s->fslab+fo; fo+=l->tq[k]->nbytes/4; int64_t nb=l->tw[k]->nbytes; - int fmt=(nb==(int64_t)OO[k]*II[k])?1:(nb==(int64_t)OO[k]*((II[k]+1)/2))?2:3; - qt[k]->fmt=fmt; qt[k]->O=OO[k]; qt[k]->I=II[k]; qt[k]->qf=NULL; + /* qt_resolve_fmt like the other two expert paths: the raw ?1:?2:3 inference here + * missed grouped int4 (fmt=4, gs never set) and would mis-tag int3-g64 as int2. */ + int gs=0; + int fmt=qt_resolve_fmt(l->tw[k]->name,OO[k],II[k],nb,l->tq[k]->nbytes,&gs); + qt[k]->fmt=fmt; qt[k]->O=OO[k]; qt[k]->I=II[k]; qt[k]->gs=gs; qt[k]->qf=NULL; qt[k]->q8=(int8_t*)(s->slab+l->pos[k]); qt[k]->q4=s->slab+l->pos[k]; qt[k]->s=fp[k]; } if(publish_eid) s->eid=l->eid; @@ -1487,6 +1824,22 @@ static int g_pipe=0; /* PIPE=1: async expert-load pipeline. Default ON for * the matmul. PIPE=0 opts back into the blocking serial path. */ static int g_pipe_nw=8; /* PIPE_WORKERS=n: I/O worker threads (disk-parallel reads) */ static int g_uring=0; /* URING=1: Linux io_uring load/completion backend; implies PIPE */ +static int g_pipe_block=0;/* COLI_PIPE_BLOCK=1: pipe_wait blocca su una condvar invece dello + * spin sched_yield (default OFF = spin byte-identico). EN: a yield + * storm on the main thread fights the OpenMP team for cycles during + * multi-ms loads; the condvar wake costs ~5us against reads that + * cost 0.5-3ms (#159). Pthread pool only: the URING backend has no + * waiter spin to replace. */ +/* PIPE_WORKERS>0 esplicito nell'env implica PIPE=1: dimensionare il pool + * dichiara l'intento di usarlo (una campagna intera l'ha impostato con la + * pipe spenta senza accorgersene). EN: fires ONLY when PIPE is unset in the + * env AND the platform default left the pipe off (on _WIN32 it already + * defaults to 1) AND PIPE_WORKERS parses positive — the internal default of + * 8 does not count, PIPE_WORKERS=0/empty does not, and an explicit PIPE=0 + * always wins. */ +static int pipe_workers_imply_pipe(const char *pipe_env, const char *pw_env, int pipe_now){ + return !pipe_now && !pipe_env && pw_env && atoi(pw_env)>0; +} typedef struct { _Atomic uint64_t cur; /* (gen<<8)|index; gen main-only, index 0..njobs (≤64) */ _Atomic int njobs; /* current batch job count */ @@ -1494,6 +1847,7 @@ typedef struct { _Atomic int layer; /* current batch layer */ _Atomic int ready[64]; /* per-slot load-done flag */ pthread_mutex_t mx; pthread_cond_t cv; /* ONLY for parking/waking idle workers */ + pthread_cond_t cv_done; /* COLI_PIPE_BLOCK: signals ready[] transitions */ Model *m; pthread_t th[16]; int nw; int started; } PipePool; @@ -1516,8 +1870,13 @@ static void *pipe_worker(void *arg){ memory_order_acq_rel,memory_order_relaxed)){ int L =atomic_load_explicit(&p->layer,memory_order_relaxed); int eid=atomic_load_explicit(&p->eids[i],memory_order_relaxed); /* AFTER winning CAS */ - expert_load(p->m,L,eid,&p->m->ws[i],1); /* needed-now load: fatal on I/O error (matches serial path) */ + expert_load(p->m,L,eid,&p->m->ws[i],1,1); /* needed-now load: fatal on I/O error (matches serial path); demand=1: this IS moe()'s own miss path */ atomic_store_explicit(&p->ready[i],1,memory_order_release); + if(g_pipe_block){ /* wake a main thread parked in pipe_wait */ + pthread_mutex_lock(&p->mx); + pthread_cond_broadcast(&p->cv_done); + pthread_mutex_unlock(&p->mx); + } } /* CAS failed → another worker advanced index (or gen advanced): re-loop */ } @@ -1535,6 +1894,7 @@ static void pipe_init(Model *m){ g_pp.m=m; g_pp.nw=g_pipe_nw; if(g_pp.nw>16) g_pp.nw=16; if(g_pp.nw<1) g_pp.nw=1; atomic_store(&g_pp.cur,0); atomic_store(&g_pp.njobs,0); pthread_mutex_init(&g_pp.mx,NULL); pthread_cond_init(&g_pp.cv,NULL); + pthread_cond_init(&g_pp.cv_done,NULL); for(int i=0;islab); free(s->fslab); s->slab=NULL; s->fslab=NULL; s->slab_cap=s->fslab_cap=0; + if(s->aslab){ s->slab=NULL; s->fslab=NULL; } /* arena slice (#419): detach, keep caps, never free */ + else { compat_aligned_free(s->slab); free(s->fslab); s->slab=NULL; s->fslab=NULL; s->slab_cap=s->fslab_cap=0; } QT *q[3]={&s->g,&s->u,&s->d}; for(int k=0;k<3;k++){ q[k]->qf=NULL; q[k]->q8=NULL; q[k]->q4=NULL; q[k]->s=NULL; } m->resident_bytes-=bytes; if(m->resident_bytes<0) m->resident_bytes=0; } static void expert_host_ensure(Model *m, int layer, ESlot *s){ - if(!s->slab) expert_load(m,layer,s->eid,s,1); + if(s->slab) return; + if(s->aslab){ s->slab=s->aslab; s->fslab=s->afslab; } /* re-attach the arena slice; caps survived release */ + /* re-materializing a GPU-resident expert's host copy, not a routing miss: demand=0 */ + expert_load(m,layer,s->eid,s,1,0); /* rebuild the QT views (release NULLed them) + reload */ } #endif /* prefetch asincrono dei pesi di un expert (e delle sue scale .qs): avvia il readahead - * cosi' le letture sincrone successive trovano la page-cache calda. The readahead - * targets the SAME replica that will serve the pread (expert_route is deterministic). */ + * cosi' le letture sincrone successive trovano la page-cache calda. + * Sotto g_direct i PESI vengono letti con O_DIRECT (bypassa la page-cache, vedi + * expert_load): il WILLNEED su di essi scalda pagine che la lettura di domanda non + * consuma -> readahead sprecato sul disco, la risorsa piu' scarsa nello streaming. + * Le scale .qs restano SEMPRE bufferizzate (pread sul fd normale), quindi il loro + * WILLNEED resta utile anche con DIRECT=1. fadvise e' solo consultivo: saltarlo non + * cambia mai l'output (bit-identico), riduce solo I/O sprecato. + * The readahead targets the SAME replica that will serve the pread (expert_route + * is deterministic). + * EN: under O_DIRECT the weights bypass the page cache, so their WILLNEED is wasted; + * the .qs scales are always buffered, so keep theirs. Advisory hint -> output-preserving. */ static void expert_prefetch(Model *m, int layer, int eid){ char nm[300]; int rep=expert_route(layer,eid); const char *suf[3]={"gate_proj.weight","up_proj.weight","down_proj.weight"}; for(int k=0;k<3;k++){ - snprintf(nm,sizeof(nm),"model.layers.%d.mlp.experts.%d.%s",layer,eid,suf[k]); st_prefetch_rep(&m->S,nm,rep); + snprintf(nm,sizeof(nm),"model.layers.%d.mlp.experts.%d.%s",layer,eid,suf[k]); + if(!g_direct) st_prefetch_rep(&m->S,nm,rep); char qs[320]; snprintf(qs,sizeof(qs),"%s.qs",nm); st_prefetch_rep(&m->S,qs,rep); } } @@ -1627,6 +2012,14 @@ static void qt_addrow(const QT *t, int row, float coef, float *acc){ acc[i] +=coef*scl[i/gs] *((int)(b&0xF)-8); acc[i+1]+=coef*scl[(i+1)/gs]*((int)(b>>4)-8); } if(I&1){ uint8_t b=w[I>>1]; acc[I-1]+=coef*scl[(I-1)/gs]*((int)(b&0xF)-8); } return; } + /* fmt=5 likewise before c: int3-g64 scales are per-GROUP [O,ng], not s[row] */ + if(t->fmt==5){ const uint8_t *w=t->q4+(int64_t)row*i3_rowbytes(I); + const float *sr=t->s+(int64_t)row*i3_groups(I); int64_t ng=i3_groups(I); + for(int64_t g=0; g>2]>>((k&3)*2))&3)|(((hi[k>>3]>>(k&7))&1)<<2); + acc[base+k]+=cg*(float)((int)u-4); } } + return; } float c=coef*t->s[row]; if(t->fmt==1){ const int8_t *w=t->q8+(int64_t)row*I; for(int i=0;ifmt==2){ const uint8_t *w=t->q4+(int64_t)row*((I+1)/2); @@ -1640,6 +2033,13 @@ static void qt_matvec_rows(const QT *t, int r0, int n, const float *x, float *y) int I=t->I; for(int j=0;jfmt==0){ const float *w=t->qf+(int64_t)row*I; for(int i=0;ifmt==4){ /* grouped int4: per-group scale */ + const uint8_t *w=t->q4+(int64_t)row*((I+1)/2); int gs=t->gs,ng=(I+gs-1)/gs; + const float *scl=t->s+(int64_t)row*ng; + for(int g=0;g*gsI)glen=I-base; float sc=scl[g]; float acc=0; + for(int i=base;i+1>1]; acc+=((int)(b&0xF)-8)*x[i]+((int)(b>>4)-8)*x[i+1]; } + if(glen&1){ uint8_t b=w[(base+glen-1)>>1]; acc+=((int)(b&0xF)-8)*x[base+glen-1]; } + a+=acc*sc; } } else if(t->fmt==1){ const int8_t *w=t->q8+(int64_t)row*I; float s=t->s[row]; float acc=0; for(int i=0;ifmt==2){ const uint8_t *w=t->q4+(int64_t)row*((I+1)/2); float s=t->s[row]; float acc=0; @@ -1652,6 +2052,13 @@ static void qt_matvec_rows(const QT *t, int r0, int n, const float *x, float *y) for(int i=base;i>1]; acc+=(float)((i&1)?((int)(b>>4)-8):((int)(b&0xF)-8))*x[i]; } a+=(double)acc*scl[g]; } } + else if(t->fmt==5){ const uint8_t *w=t->q4+(int64_t)row*i3_rowbytes(I); + const float *sr=t->s+(int64_t)row*i3_groups(I); int64_t ng=i3_groups(I); + for(int64_t g=0; g>2]>>((k&3)*2))&3)|(((hi[k>>3]>>(k&7))&1)<<2); + acc+=(float)((int)u-4)*x[base+k]; } + a+=(double)(acc*sr[g]); } } else { const uint8_t *w=t->q4+(int64_t)row*((I+3)/4); float s=t->s[row]; float acc=0; for(int i=0;i>2]; acc+=((int)((b>>((i&3)*2))&3)-2)*x[i]; } a=acc*s; } y[j]=(float)a; @@ -1660,6 +2067,7 @@ static void qt_matvec_rows(const QT *t, int r0, int n, const float *x, float *y) static int g_absorb=-1; #ifdef COLI_CUDA static int g_cuda_pipe=0; /* COLI_CUDA_PIPE=1: prefill attention chain resident on the layer home device */ +static int g_cuda_router=0; /* COLI_CUDA_ROUTER=1 (#431 PR-A): router on the layer home device at decode */ #endif /* ABSORB: -1 auto (decode S<=4), 0 mai, 1 sempre (test) */ static int g_dsa_force=0; /* DSA_FORCE=1: selezione sempre attiva (test: top-min(k,T)=denso) */ static int cmp_fdesc(const void *a,const void *b){ @@ -2203,6 +2611,15 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int pthread_mutex_unlock(&g_pilot_mx); } Cfg *c=&m->c; int D=c->hidden, E=c->n_experts, K=c->topk, I=c->moe_inter; + /* DISK-CLASS: does THIS call need the pre-bump recency snapshot? Must agree with + * dc_needed() in expert_load_impl -- that's what reads what this writes. touched[] + * makes the write once-per-call: an expert routed by more than one position in a big + * batch (prefill's S) must snapshot the state from BEFORE this call started, not from + * an earlier position's bump within the SAME call (which would reintroduce the + * same-call contamination one position later). Unconditional VLA like the FASE B + * `seen[E]` below -- E is small, cost is noise. */ + int need_classify = dc_needed(); + unsigned char touched[E]; if(need_classify) memset(touched,0,(size_t)E); float *choice=falloc(E); int sI=c->moe_inter*c->n_shared; /* Rank buffer for CACHE_ROUTE max-rank selection (up to all E experts). */ @@ -2222,7 +2639,8 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int /* router in UN matmul batch: stessa matematica, via le S chiamate S=1 */ float *logits_all=falloc((int64_t)S*E); int pre_routed=0; (void)pre_routed; -#ifdef COLI_METAL + /* pre-routed shortcut: Metal layer-CB o device router CUDA (#431) — stessa + * contabilita' (#417: recency clock incluso), la selezione arriva dalla GPU */ if(g_pre_idx){ /* routing gia' calcolata dal layer CB (GPU) */ memcpy(idxs,g_pre_idx,(size_t)S*K*sizeof(int)); memcpy(ws,g_pre_w,(size_t)S*K*sizeof(float)); @@ -2232,6 +2650,15 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int for(int kk=0;kkeusage[layer][idxs[(int64_t)s*K+kk]]++; ehit_mark(m,layer,idxs[(int64_t)s*K+kk]); + if(need_classify){ /* DISK-CLASS private recency -- snapshot BEFORE this call's own bump, + * then tick. This path also bumps the REAL elast/eaccess_clock a few + * lines below (#417/cfcc742 fixed the once-missing bump on Metal + * decode) -- DISK-CLASS's own clock stays independent regardless of + * that fix, see elast_dc in Model. */ + int e=idxs[(int64_t)s*K+kk]; + if(!touched[e]){ m->elast_pre[layer][e]=m->elast_dc[layer][e]; touched[e]=1; } + m->elast_dc[layer][e]=++m->eaccess_clock_dc; + } if(m->eheat[layer][idxs[(int64_t)s*K+kk]]eheat[layer][idxs[(int64_t)s*K+kk]]++; /* #417: la scorciatoia GPU-prerouted deve far avanzare l'orologio di recency * come il percorso router completo (riga ~3055), altrimenti elast/eaccess_clock @@ -2243,7 +2670,6 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int } pre_routed=1; } -#endif if(!pre_routed) matmul(logits_all, x, l->router, S, D, E); if(!pre_routed) for(int s=0;seusage[layer][idx[kk]]++; ehit_mark(m,layer,idx[kk]); + if(need_classify){ /* DISK-CLASS private recency -- snapshot BEFORE this call's own bump, + * then tick (same rate as the real clock below: one per (s,kk)) */ + if(!touched[idx[kk]]){ m->elast_pre[layer][idx[kk]]=m->elast_dc[layer][idx[kk]]; touched[idx[kk]]=1; } + m->elast_dc[layer][idx[kk]]=++m->eaccess_clock_dc; + } if(m->eheat[layer][idx[kk]]eheat[layer][idx[kk]]++; m->elast[layer][idx[kk]]=++m->eaccess_clock; } @@ -2558,7 +2989,7 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int * are timed as service inside expert_load */ } else { double t0=now_s(); /* ORIGINALE: blocking parallel load */ #pragma omp parallel for schedule(dynamic,1) - for(int q=0;qws[q],1); + for(int q=0;qws[q],1,1); /* demand=1: this IS the miss path */ m->t_ewait += now_s()-t0; } /* compute thread blocked for the whole load */ } /* I/O ASINCRONO: readahead (WILLNEED) del blocco SUCCESSIVO mentre calcoliamo @@ -2575,6 +3006,11 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int } #ifdef COLI_CUDA ESlot *group_e[64]; int group_n[64]; int ngroup=0; + /* Inc.4 overlap stash: pass-1 packing kept for the take phase after the CPU loop */ + ESlot *eg_e[64]; int eg_n[64], eg_row[64][4], eg_npg=0; float eg_w[64][4]; + int dev_nc0[COLI_CUDA_MAX_DEVICES], dev_off0[COLI_CUDA_MAX_DEVICES], + dev_total0[COLI_CUDA_MAX_DEVICES], dev_which0[COLI_CUDA_MAX_DEVICES][64]; + memset(dev_nc0,0,sizeof(dev_nc0)); (void)eg_npg; (void)dev_total0; (void)dev_off0; #endif #ifdef COLI_METAL if(g_metal_enabled){ @@ -2603,9 +3039,83 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int free(mxg); free(mrows); free(mrw); } #undef MB_BUILD +#endif +#ifdef COLI_CUDA + /* Inc.4 pass 1: collect the VRAM-resident experts' groups and ISSUE them async + * BEFORE the CPU loop below, so the GPU computes its share while the CPU works + * through the RAM-tier/miss rows — t_emm becomes max(cpu, gpu) instead of the + * sum. Only resident experts are collected (misses are never cuda_eligible), so + * no pipe_wait is needed here; the CPU loop keeps its own waits. Any issue + * failure drops the layer back to the collect-in-loop + sync-group path. */ + int early_issued=0, done_j[64]={0}; + { + static int g_group_async2=-1; + if(g_group_async2<0) g_group_async2=getenv("COLI_GROUP_ASYNC")?atoi(getenv("COLI_GROUP_ASYNC")):0; + if(!metal_done && g_group_async2 && group_enabled && S<=4 && g_cuda_enabled && + g_cuda_ndev>0 && !omp_in_parallel()){ + ESlot *pg_e[64]; int pg_n[64], pg_j[64], npg=0; + int prow[64][4]; float pw[64][4]; + for(int j=0;jg.cuda_eligible&&e->u.cuda_eligible&&e->d.cuda_eligible)) continue; + int nr=0; + for(int s=0;sg.cuda_device==g_cuda_devices[di]) pd_total[di]+=pg_n[q]; + for(int di=1;dig.cuda_device==device){ + int nc=pd_nc[di]++; ESlot *e=pg_e[q]; + pd_g[di][nc]=e->g.cuda; pd_u[di][nc]=e->u.cuda; pd_d[di][nc]=e->d.cuda; + pd_rows[di][nc]=pg_n[q]; pd_which[di][nc]=q; + for(int r=0;rt_emm+=now_s()-tg0; + for(int q=0;qgpu_expert_calls++; + } + } else { + for(int di=0;di=1 routing invariant. @@ -2652,6 +3162,36 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int m->cpu_expert_rows+=(uint64_t)nr;} } #ifdef COLI_CUDA + /* Inc.4 take phase: the CPU loop above ran while the GPU computed the issued + * groups — collect them now. A failed device recomputes its experts on the CPU + * (expert_host_ensure reloads slabs released by CUDA_RELEASE_HOST). */ + if(early_issued){ + double tg1=now_s(); + g_ovl_cpu+=tg1-g_ovl_mark; /* CPU-row window between issue and take */ + for(int di=0;dig,&e->u,nr); + for(int64_t z=0;z<(int64_t)nr*I;z++) gg[z]=siluf(gg[z])*uu[z]; + matmul_qt(hh,gg,&e->d,nr); + for(int r=0;rt_emm+=now_s()-tg1; g_ovl_take+=now_s()-tg1; + } ColiCudaTensor *dev_g[COLI_CUDA_MAX_DEVICES][64],*dev_u[COLI_CUDA_MAX_DEVICES][64]; ColiCudaTensor *dev_d[COLI_CUDA_MAX_DEVICES][64]; int dev_rows[COLI_CUDA_MAX_DEVICES][64],dev_which[COLI_CUDA_MAX_DEVICES][64]; @@ -2673,6 +3213,33 @@ static void moe(Model *m, Layer *l, int layer, float *x, int S, float *out, int } } double tg=now_s(); + /* Inc.4: at decode scale, issue every device's group WITHOUT syncing, then take + * them all — one stream sync per device per layer instead of a full staged + * round-trip per call (measured: ~70% of the sync call is host-side wait). + * Any issue failure drains what was issued and the whole layer falls back to + * the sync path below, which recomputes from group_x (idempotent). */ + int async_done=0; + static int g_group_async=-1; + if(g_group_async<0) g_group_async=getenv("COLI_GROUP_ASYNC")?atoi(getenv("COLI_GROUP_ASYNC")):0; + if(g_group_async && S<=4 && g_cuda_ndev>0){ + int issued[COLI_CUDA_MAX_DEVICES]={0}, all=1; + for(int di=0;di1) schedule(static) for(int di=0;dish_gate.cuda_eligible||!l->sh_up.cuda_eligible||!l->sh_down.cuda_eligible|| !qt_cuda_upload(&l->sh_gate)||!qt_cuda_upload(&l->sh_up)||!qt_cuda_upload(&l->sh_down)|| l->sh_gate.cuda_device!=dev||l->sh_up.cuda_device!=dev||l->sh_down.cuda_device!=dev) return 0; - float *w_in =coli_cuda_pipe_scratch(dev,8,(size_t)D*4); - float *w_post=coli_cuda_pipe_scratch(dev,9,(size_t)D*4); + /* Inc.4: the layernorm weights are constants — upload once per layer and keep them + * on the layer's device, instead of two synchronous 24 KB uploads per layer per + * token (152 sync H2D/token measured on the profile). */ + if(!m->ln_dev) m->ln_dev=calloc((size_t)(c->n_layers+1)*2,sizeof(float*)); + float *w_in=m->ln_dev[(size_t)li*2], *w_post=m->ln_dev[(size_t)li*2+1]; + if(!w_in){ + w_in=coli_cuda_pipe_alloc(dev,(size_t)D*4); + if(!w_in||!coli_cuda_pipe_upload(dev,w_in,l->in_ln,(size_t)D*4)) return 0; + m->ln_dev[(size_t)li*2]=w_in; + } + if(!w_post){ + w_post=coli_cuda_pipe_alloc(dev,(size_t)D*4); + if(!w_post||!coli_cuda_pipe_upload(dev,w_post,l->post_ln,(size_t)D*4)) return 0; + m->ln_dev[(size_t)li*2+1]=w_post; + } float *nrm_d=coli_cuda_pipe_scratch(dev,10,xb); float *y_d =coli_cuda_pipe_scratch(dev,11,xb); float *sg_d =coli_cuda_pipe_scratch(dev,12,(size_t)S*sI*4); float *su_d =coli_cuda_pipe_scratch(dev,13,(size_t)S*sI*4); float *snap =coli_cuda_pipe_scratch(dev,14,xb); - if(!w_in||!w_post||!nrm_d||!y_d||!sg_d||!su_d||!snap) return 0; + if(!nrm_d||!y_d||!sg_d||!su_d||!snap) return 0; if(!coli_cuda_pipe_peer_copy(dev,snap,dev,x_dev,xb)) return 0; /* snapshot per il fallback */ - if(!coli_cuda_pipe_upload(dev,w_in,l->in_ln,(size_t)D*4)|| - !coli_cuda_pipe_upload(dev,w_post,l->post_ln,(size_t)D*4)) return 0; double ta=now_s(); if(!coli_cuda_pipe_rmsnorm(dev,nrm_d,x_dev,w_in,S,D,c->eps)) return 0; /* DSA: i layer con indexer FULL cachano k_idx dalla nrm pre-attention (CPU, piccolo) */ @@ -3147,6 +3725,36 @@ static int pipe_layer_sparse(Model *m, Layer *l, int li, float *x_dev, int S, in if(!attn_pipe_prefill(m,l,li,nrm_d,1,S,pos_base,NULL,y_d)) return 0; if(!coli_cuda_pipe_add(dev,x_dev,y_d,(size_t)S*D)) return 0; /* prima mutazione */ if(!coli_cuda_pipe_rmsnorm(dev,nrm_d,x_dev,w_post,S,D,c->eps)) return 0; + /* device router (#431 PR-A): route THIS row on the home device while the + * stream is still hot, then hand the selection to moe() through the same + * pre-routed shortcut the Metal layer-CB uses. Any failure (upload, launch, + * feature gate) falls back to the CPU router inside moe() — byte-identical + * behaviour, just slower. Gated to the plain routing path: CACHE_ROUTE / + * ROUTE_P / ROUTE_TRACE keep the CPU ranking they need. */ + static int lr_idx[64]; static float lr_w[64]; static int lr_keff[1]; + int dev_routed=0; + if(g_cuda_router && S==1 && !g_cache_route && g_route_p<=0.f && !g_route_fp + && c->n_experts<=4096 && c->topk<=64 && !l->router_cuda_bad){ + int E=c->n_experts, K=c->topk; + int Ksel = g_topk>0 ? (g_topk0 && g_topp<1.f) ? g_topp : 0.f; + if(!l->router_cuda){ + void *rw=coli_cuda_pipe_alloc(dev,(size_t)E*D*4); + void *rb=coli_cuda_pipe_alloc(dev,(size_t)E*4); + if(rw&&rb&&coli_cuda_pipe_upload(dev,rw,l->router,(size_t)E*D*4) + &&coli_cuda_pipe_upload(dev,rb,l->router_bias,(size_t)E*4)){ + l->router_cuda=rw; l->router_bias_cuda=rb; + } else { + if(rw)coli_cuda_pipe_free(dev,rw); if(rb)coli_cuda_pipe_free(dev,rb); + l->router_cuda_bad=1; + } + } + if(l->router_cuda && + coli_cuda_pipe_router(dev,nrm_d,l->router_cuda,l->router_bias_cuda, + D,E,Ksel,tp,c->norm_topk,c->routed_scale, + lr_idx,lr_w,lr_keff)) + dev_routed=1; + } if(!coli_cuda_pipe_download(dev,nrm_d,nrm_host,xb)) return 0; m->t_attn+=now_s()-ta; /* OVERLAP: issue the shared expert on the GPU BEFORE moe() runs on the CPU. @@ -3177,7 +3785,9 @@ static int pipe_layer_sparse(Model *m, Layer *l, int li, float *x_dev, int S, in if(!coli_cuda_pipe_add(dev,x_dev,y_d,(size_t)S*D)) return 0; /* shared residual (async) */ m->t_emm += now_s()-te; /* shared-expert GPU dispatch only */ /* expert routed su CPU/gruppi GPU come oggi (shared saltata: la fa il device) */ + if(dev_routed){ g_pre_idx=lr_idx; g_pre_w=lr_w; g_pre_keff=lr_keff; } moe(m,l,li,nrm_host,S,out_host,0); /* self-times its own t_emm */ + if(dev_routed){ g_pre_idx=NULL; g_pre_w=NULL; g_pre_keff=NULL; } te=now_s(); if(!coli_cuda_pipe_upload(dev,y_d,out_host,xb)) return 0; /* sync: waits for moe */ if(!coli_cuda_pipe_add(dev,x_dev,y_d,(size_t)S*D)) return 0; /* routed residual (async) */ @@ -3549,10 +4159,36 @@ static void repin_pass(Model *m){ repin_pass_limit(m,16); } * grammar_draft propone lo span FORZATO successivo (un solo byte legale per posizione) * gia' tokenizzato. Il confine di tokenizzazione non e' garantito coincidere con quello * del modello: la verifica assorbe la differenza (al peggio l'ultimo draft e' rifiutato). */ -static void grammar_setup(Tok *T){ +/* Compile grammar TEXT into g: raw GBNF, or — if the first non-space byte is '{' + * — a JSON-Schema compiled via schema_gbnf.h. Takes ownership of txt (always + * freed). Fail-soft: returns -1 with g->on=0, engine runs without a grammar. */ +static int grammar_setup_text(GrDraft *g, Tok *T, char *txt, const char *label){ + const char *p=txt; while(*p==' '||*p=='\t'||*p=='\n'||*p=='\r') p++; + if(*p=='{'){ + char serr[160]; + char *gbnf=schema_to_gbnf(txt,serr,sizeof serr); + free(txt); + if(!gbnf){ fprintf(stderr,"[SCHEMA] %s: %s (running without grammar)\n",label,serr); return -1; } + txt=gbnf; + } + if(gr_parse(&g->gram,txt)){ fprintf(stderr,"[GRAMMAR] %s: %s\n",label,g->gram.err); free(txt); return -1; } + free(txt); + gr_state_init(&g->st,&g->gram); + if(!g->st.alive){ fprintf(stderr,"[GRAMMAR] %s: grammar cannot be evaluated (left recursion?)\n",label); return -1; } + if(g->max<1) g->max=24; + if(g->max>48) g->max=48; + g->T=T; g->on=1; g->armed=0; + fprintf(stderr,"[GRAMMAR] %s: %d rules, forced span capped at %d tokens/forward\n",label,g->gram.n,g->max); + return 0; +} +/* Release a per-request grammar so the slot can host the next request (keeps max). */ +static void grammar_teardown(GrDraft *g){ + if(g->gram.n) gr_free(&g->gram); + int max=g->max; memset(g,0,sizeof(*g)); g->max=max; +} +static void grammar_setup(GrDraft *g, Tok *T){ /* GRAMMAR= takes precedence; SCHEMA= compiles a JSON-Schema - * to GBNF (schema_gbnf.h) for the same draft source. Both fail soft: the engine - * runs without a grammar and output is unchanged. */ + * to GBNF. Both fail soft: the engine runs without a grammar, output unchanged. */ const char *gf=getenv("GRAMMAR"); const char *sf=(gf&&*gf)?NULL:getenv("SCHEMA"); if((!gf||!*gf)&&(!sf||!*sf)) return; @@ -3564,59 +4200,45 @@ static void grammar_setup(Tok *T){ if(!txt || fread(txt,1,(size_t)n,f)!=(size_t)n){ fprintf(stderr,"[GRAMMAR] failed to read %s\n",path); fclose(f); free(txt); return; } fclose(f); txt[n]=0; - if(sf){ /* schema -> GBNF, then the same gr_parse as the GRAMMAR path */ - char serr[160]; - char *gbnf=schema_to_gbnf(txt,serr,sizeof serr); - free(txt); - if(!gbnf){ fprintf(stderr,"[SCHEMA] %s: %s (running without grammar)\n",sf,serr); return; } - txt=gbnf; - } - if(gr_parse(&g_gram,txt)){ fprintf(stderr,"[GRAMMAR] %s: %s\n",path,g_gram.err); free(txt); return; } - free(txt); - gr_state_init(&g_gst,&g_gram); - if(!g_gst.alive){ fprintf(stderr,"[GRAMMAR] %s: grammar cannot be evaluated (left recursion?)\n",path); return; } - if(getenv("GRAMMAR_DRAFT")) g_gr_max=atoi(getenv("GRAMMAR_DRAFT")); - if(g_gr_max<1) g_gr_max=1; - if(g_gr_max>48) g_gr_max=48; - g_gr_T=T; g_gr_on=1; - fprintf(stderr,"[GRAMMAR] %s: %d rules, forced span capped at %d tokens/forward\n",path,g_gram.n,g_gr_max); + if(getenv("GRAMMAR_DRAFT")) g->max=atoi(getenv("GRAMMAR_DRAFT")); + grammar_setup_text(g,T,txt,path); } /* stato pulito all'inizio di ogni RISPOSTA (non tra i \x02MORE, che continuano) */ -static void grammar_reset(void){ - if(!g_gr_on) return; - gr_state_init(&g_gst,&g_gram); g_gr_armed=0; - if(!g_gst.alive) g_gr_on=0; +static void grammar_reset(GrDraft *g){ + if(!g->on) return; + gr_state_init(&g->st,&g->gram); g->armed=0; + if(!g->st.alive) g->on=0; } /* consuma i byte di un token emesso. Preambolo (prima dell'arming): ignorato. * Desync dopo l'arming: si riarma in attesa del prossimo inizio valido — al peggio * i draft vengono rifiutati dalla verifica, l'output non cambia MAI. */ -static void gr_feed(int t){ - if(!g_gr_on||!g_gr_T) return; - char b[64]; int n=tok_decode(g_gr_T,&t,1,b,63); +static void gr_feed(GrDraft *g, int t){ + if(!g->on||!g->T) return; + char b[64]; int n=tok_decode(g->T,&t,1,b,63); for(int i=0;ist,(unsigned char)b[i]); + if(r==1){ g->armed=1; continue; } + if(r<0){ g->on=0; return; } /* walker spento: fine dei draft */ + if(!g->armed) continue; /* preambolo: aspetta l'inizio */ + gr_state_init(&g->st,&g->gram); g->armed=0; /* desync: riparti dalla radice */ + if(!g->st.alive){ g->on=0; return; } + if(gr_accept(&g->st,(unsigned char)b[i])==1) g->armed=1; } } /* propone lo span forzato come token (max cap); 0 se la grammatica dirama qui */ -static int grammar_draft(int *draft, int cap){ - if(!g_gr_on||!g_gr_armed||!g_gr_T||cap<1) return 0; - if(g_gr_prop>=32 && g_gr_acc*2on||!g->armed||!g->T||cap<1) return 0; + if(g->prop>=32 && g->acc*2prop){ /* guardia adattiva, come per MTP: acceptance sotto il 50% = tokenizzazione fuori asse, meglio spegnersi */ - g_gr_on=0; + g->on=0; fprintf(stderr,"[GRAMMAR] %.0f%% acceptance after %llu proposals: grammar drafts disabled\n", - 100.0*g_gr_acc/g_gr_prop,(unsigned long long)g_gr_prop); + 100.0*g->acc/g->prop,(unsigned long long)g->prop); return 0; } - char fb[512]; int nb=gr_forced(&g_gst,fb,(int)sizeof fb-1); + char fb[512]; int nb=gr_forced(&g->st,fb,(int)sizeof fb-1); if(nb<=0) return 0; - int g=tok_encode(g_gr_T,fb,nb,draft,cap); - return g>0?g:0; + int nt=tok_encode(g->T,fb,nb,draft,cap); /* renamed local: 'g' is now the state param */ + return nt>0?nt:0; } /* ---- SAMPLING (temperatura + nucleus) con verifica speculativa LOSSLESS ---- @@ -3674,12 +4296,12 @@ static int spec_decode(Model *m, int *all, int kv, int n_new, int eos, float *lo int next=pick_tok(logit,V,carry_ban); carry_ban=-1; free(logit); logit=NULL; if((eos>=0 && next==eos) || is_stop(next)) break; emit(next,ud); all[kv]=next; emitted++; m->n_emit++; - gr_feed(next); /* il walker segue l'output emesso */ + gr_feed(&g_grd,next); /* il walker segue l'output emesso */ if(emitted>=n_new) break; /* l'ultimo token non serve forwardarlo */ int g = 0, gsrc = 0; /* sorgente: 1=grammatica 2=MTP/n-gram */ - if(g_gr_on){ /* metodo F: prima la grammatica — dove + if(g_grd.on){ /* metodo F: prima la grammatica — dove * forza, l'acceptance e' ~1 (#48) */ - g=grammar_draft(draft,g_gr_max); + g=grammar_draft(&g_grd,draft,g_grd.max); if(g>0) gsrc=1; } if(!g && g_draft>0 && m->has_mtp){ @@ -3701,7 +4323,7 @@ static int spec_decode(Model *m, int *all, int kv, int n_new, int eos, float *lo if(g>n_new-emitted) g=n_new-emitted; if(kv+1+g+1>m->max_t) g=m->max_t-kv-2; if(g<0) g=0; - if(gsrc==1) g_gr_prop+=(uint64_t)g; + if(gsrc==1) g_grd.prop+=(uint64_t)g; int S=1+g; int batch[64]; batch[0]=next; memcpy(batch+1,draft,g*sizeof(int)); double tf0=g_prof?now_s():0; float *lo=step_all(m,batch,S,kv); m->n_fw++; @@ -3717,9 +4339,9 @@ static int spec_decode(Model *m, int *all, int kv, int n_new, int eos, float *lo if(!accept){ if(g_temp>0) carry_ban=draft[k]; break; } if((eos>=0 && draft[k]==eos) || is_stop(draft[k])){ done=1; break; } emit(draft[k],ud); all[kv+1+k]=draft[k]; emitted++; m->n_emit++; - gr_feed(draft[k]); k++; + gr_feed(&g_grd,draft[k]); k++; } - if(gsrc==1) g_gr_acc+=(uint64_t)k; + if(gsrc==1) g_grd.acc+=(uint64_t)k; else if(gsrc==2 && m->has_mtp) m->mtp_acc+=k; if(m->has_mtp && k>=1) mtp_absorb(m, all+kv+1, m->h_all, k, kv); /* KV MTP in sync coi verificati */ /* hlast deve corrispondere all'ultima posizione ACCETTATA (kv+k), non a fine batch */ @@ -3868,7 +4490,9 @@ static void profile_print(Model *m, double elapsed){ if(aok){ double ks=0,gs=0; coli_metal_attn_lat(&ks,&gs); printf("METAL-ATTN: layer GPU %llu | gpu-wall %.2fs (kernel %.2fs | cpu-sched %.2fs gpu-sched %.2fs)\n",(unsigned long long)aok,aw,ak,ks,gs); } } printf("METAL: blocchi GPU %llu | fallback CPU %llu | expert su GPU %llu | setup %.2fs gpu-wall %.2fs (kernel %.2fs) scatter %.2fs\n", - (unsigned long long)ok,(unsigned long long)fb,(unsigned long long)ex,su,gp,coli_metal_moe_kernel_time(),sc); } + (unsigned long long)ok,(unsigned long long)fb,(unsigned long long)ex,su,gp,coli_metal_moe_kernel_time(),sc); + { double rsf=0; if(coli_metal_resset_stats(&rsf)) /* E5: printed only when the gate is on */ + printf("METAL-RESSET: flush %.2fs (residency-set commit in moe_submit, outside setup/gpu-wall)\n",rsf); } } #endif } @@ -3919,6 +4543,42 @@ static void prof_report(Model *m, const ProfBase *b, double elapsed, int tokens, g_mmap?"; COLI_MMAP=1: page cache may serve part":"", hitp,(unsigned long long)dhp,(unsigned long long)dhe,(unsigned long long)dm, tokens>0?(double)dq/tokens:0.0, io_svc,io_w); + /* DISK-CLASS: per-load cold/warm classification vs. which fd ACTUALLY served it. + * Three per-class rates, labeled to keep the units unambiguous (ambiguous units + * mislead -- measured lesson): GB/s-thread = bytes / thread-seconds (per-read + * service rate, same convention as the read-service line above); GB/s-wall = + * bytes / busy-wall (aggregate rate the disk actually delivered while >=1 load of + * the class was in flight); avg-conc = thread-seconds / busy-wall (mean overlap + * depth). disk-busy = combined busy-wall (either class in flight) as a share of + * the profile window. */ + { + uint64_t dcn[2], dcdn[2]; int64_t dcbytes[2], dcns[2], dcwall[2], wnow[2], dcwall_all, wall_now; + dc_wall_read(wnow,&wall_now); + for(int i=0;i<2;i++){ + dcn[i]=atomic_load_explicit(&g_dc_n[i],memory_order_relaxed)-b->dc_n[i]; + dcbytes[i]=atomic_load_explicit(&g_dc_bytes[i],memory_order_relaxed)-b->dc_bytes[i]; + dcns[i]=atomic_load_explicit(&g_dc_ns[i],memory_order_relaxed)-b->dc_ns[i]; + dcdn[i]=atomic_load_explicit(&g_dc_direct_n[i],memory_order_relaxed)-b->dc_direct_n[i]; + dcwall[i]=wnow[i]-b->dc_wall_ns[i]; + } + dcwall_all=wall_now-b->dc_wall_all_ns; + if(dcn[DC_COLD]+dcn[DC_WARM]){ + double ct=dcns[DC_COLD]>0?(double)dcbytes[DC_COLD]/dcns[DC_COLD]:0.0; /* bytes/ns = GB/s */ + double wt=dcns[DC_WARM]>0?(double)dcbytes[DC_WARM]/dcns[DC_WARM]:0.0; + double cw=dcwall[DC_COLD]>0?(double)dcbytes[DC_COLD]/dcwall[DC_COLD]:0.0; + double ww=dcwall[DC_WARM]>0?(double)dcbytes[DC_WARM]/dcwall[DC_WARM]:0.0; + double cc=dcwall[DC_COLD]>0?(double)dcns[DC_COLD]/dcwall[DC_COLD]:0.0; + double wc=dcwall[DC_WARM]>0?(double)dcns[DC_WARM]/dcwall[DC_WARM]:0.0; + fprintf(f,"[PROF] DISK-CLASS (recency split): cold %llu x %.2f GB @ %.2f GB/s-thread, wall %.1fs @ %.2f GB/s-wall, avg-conc %.1f (direct %llu/%llu) | " + "warm %llu x %.2f GB @ %.2f GB/s-thread, wall %.1fs @ %.2f GB/s-wall, avg-conc %.1f (direct %llu/%llu) | " + "disk-busy %.1fs (%.0f%% of window)\n", + (unsigned long long)dcn[DC_COLD],dcbytes[DC_COLD]/1e9,ct,dcwall[DC_COLD]/1e9,cw,cc, + (unsigned long long)dcdn[DC_COLD],(unsigned long long)dcn[DC_COLD], + (unsigned long long)dcn[DC_WARM],dcbytes[DC_WARM]/1e9,wt,dcwall[DC_WARM]/1e9,ww,wc, + (unsigned long long)dcdn[DC_WARM],(unsigned long long)dcn[DC_WARM], + dcwall_all/1e9,100.0*(dcwall_all/1e9)/elapsed); + } + } fprintf(f,"[PROF] resident experts: %d pinned (%.1f GB) + %d in LRU (%.1f GB, cap %d/layer)\n", pinned,pinned*eb/1e9,lru,lru*eb/1e9,m->ecap); double emm=m->t_emm-b->emm, ecpu=m->t_ecpu-b->ecpu, egpu=m->t_egpu-b->egpu; @@ -3993,12 +4653,24 @@ static void run_text(Model *m, const char *snap, const char *prompt, int ngen){ Tok T; tok_load(&T,tkp); int eos=tok_id_of(&T,"<|endoftext|>"); stops_arm_tok(&m->c, eos, &T); - grammar_setup(&T); /* metodo F: GRAMMAR=file.gbnf (#48) */ + grammar_setup(&g_grd,&T); /* metodo F: GRAMMAR=file.gbnf (#48) */ if(g_temp<0) g_temp=0.7f; /* auto: 0.7, NON l'1.0 ufficiale — la coda della * distribuzione int4 e' rumore di quantizzazione */ - int cap=(int)strlen(prompt)+16; int *pids=malloc(cap*sizeof(int)); + int cap=(int)strlen(prompt)+16; int *pids=malloc((cap+2)*sizeof(int)); int np=tok_encode(&T,prompt,(int)strlen(prompt),pids,cap); if(np<1){ fprintf(stderr,"prompt is empty after tokenization\n"); return; } + /* GLM prefix (#108): every GLM training sequence starts [gMASK]; without it the + * model is out-of-distribution and generates garbage. Serve/SCORE modes prepend it; + * run_text must too. CHAT_TEMPLATE=0 disables (raw prompt, for debugging/non-GLM). + * A prompt that already starts with the prefix passes through intact. */ + int templ=getenv("CHAT_TEMPLATE")?atoi(getenv("CHAT_TEMPLATE")):1; + if(templ){ + int gmask=tok_id_of(&T,"[gMASK]"), sop=tok_id_of(&T,""); + if(gmask>=0 && sop>=0 && (np<2 || pids[0]!=gmask || pids[1]!=sop)){ + memmove(pids+2,pids,np*sizeof(int)); + pids[0]=gmask; pids[1]=sop; np+=2; + } + } printf("prompt: %d tokens | generating up to %d (EOS stop=%d) | n-gram draft=%d\n", np, ngen, eos, g_draft); fputs(prompt,stdout); fflush(stdout); kv_alloc(m, np+ngen+g_draft+2); @@ -4022,7 +4694,7 @@ static void run_text(Model *m, const char *snap, const char *prompt, int ngen){ ProfBase pb; prof_base(m,&pb); double t=now_s(); EmitStream es={&T,m,t,0,0}; - grammar_reset(); + grammar_reset(&g_grd); int produced=spec_decode(m,all,np,ngen,eos,logit,emit_stream,&es,NULL); double dt=now_s()-t; double tot=m->hits+m->miss; @@ -4049,8 +4721,8 @@ static void run_text(Model *m, const char *snap, const char *prompt, int ngen){ m->n_fw?(double)m->n_emit/m->n_fw:1.0, (unsigned long long)m->n_fw, (unsigned long long)m->n_emit, m->mtp_prop?100.0*m->mtp_acc/m->mtp_prop:0.0, (unsigned long long)m->mtp_acc, (unsigned long long)m->mtp_prop); if(g_cp_enq) printf("couple: %ld cross-layer prefetch hints enqueued\n", g_cp_enq); - if(g_gr_prop) printf("grammar: %.0f%% acceptance (%llu/%llu forced drafts)\n", - 100.0*g_gr_acc/g_gr_prop, (unsigned long long)g_gr_acc, (unsigned long long)g_gr_prop); + if(g_grd.prop) printf("grammar: %.0f%% acceptance (%llu/%llu forced drafts)\n", + 100.0*g_grd.acc/g_grd.prop, (unsigned long long)g_grd.acc, (unsigned long long)g_grd.prop); if(g_disk_split) printf("disk-load split: draft %llu + absorb %llu + verify/main %llu misses | " "MTP-layer %llu loads %.2f GB | main-layers %llu loads %.2f GB (MTP %.1f%% of bytes)\n", (unsigned long long)m->miss_draft, (unsigned long long)m->miss_absorb, @@ -4181,16 +4853,21 @@ static void rss_guard(Model *m){ if(lru<0){ pthread_mutex_unlock(&g_pilot_mx); continue; } ESlot *s=&m->ecache[l][lru]; s->eid=-1; /* nascosto: nessun hit/evict altrui */ - pthread_mutex_unlock(&g_pilot_mx); int64_t sb=s->slab_cap + s->fslab_cap*4; #ifdef COLI_METAL if(s->slab && g_metal_enabled) coli_metal_unregister(s->slab); #endif + /* La free resta SOTTO g_pilot_mx. Sbloccando prima, lo slot e' visibile come + * {eid=-1, slab ancora valido}: il pilota (pilot_realload) riusa per primo gli + * slot eid==-1, quindi puo' prenderlo e fare pread dentro lo slab MENTRE lo + * liberiamo -> use-after-free / double-free. Slab valido e slot riusabile + * devono restare mutuamente esclusivi finche' il puntatore non e' NULL. */ compat_aligned_free(s->slab); free(s->fslab); s->slab=NULL; s->fslab=NULL; s->slab_cap=s->fslab_cap=0; QT *q[3]={&s->g,&s->u,&s->d}; for(int k=0;k<3;k++){ q[k]->qf=NULL; q[k]->q8=NULL; q[k]->q4=NULL; q[k]->s=NULL; } s->used=0; /* primo candidato al riuso */ + pthread_mutex_unlock(&g_pilot_mx); freed += sb; dropped++; } if(m->ecap>2) m->ecap--; /* il tetto scende: niente ricrescita */ @@ -4258,7 +4935,7 @@ static void repin_pass_limit(Model *m,int limit){ +(int64_t)coli_cuda_tensor_bytes(s->d.cuda) : 0; #endif double t0=now_s(); - expert_load(m,cd[b].l,cd[b].eid,s,1); /* disk -> RAM, same resident slot */ + expert_load(m,cd[b].l,cd[b].eid,s,1,0); /* disk -> RAM, same resident slot; demand=0: repin, never classified */ const char *tier="RAM"; #ifdef COLI_CUDA if(gpu){ /* refresh the same VRAM slot now, not lazily */ @@ -4365,8 +5042,8 @@ static void mux_done(Model *m, ServeCtx *sc, ServeReq *r){ /* Read and prefill one request. Returns -1 on EOF, 0 for a rejected frame and * 1 for an accepted request. Prefill deliberately remains serial: continuous * batching starts at decode, where every active slot contributes one row. */ -static int mux_submit(Model *m, Tok *T, ServeCtx *ctx, ServeReq *req, int nctx, - int maxctx, int eos){ +static int mux_submit(Model *m, Tok *T, ServeCtx *ctx, ServeReq *req, GrDraft *grd, + int nctx, int maxctx, int eos){ char *line=NULL; size_t cap=0; ssize_t nr=getline(&line,&cap,stdin); if(nr<0){ free(line); return -1; } if(nr && line[nr-1]=='\n') line[--nr]=0; @@ -4387,21 +5064,31 @@ static int mux_submit(Model *m, Tok *T, ServeCtx *ctx, ServeReq *req, int nctx, char *raw=malloc((size_t)sub.bytes+1); if(!raw){ fprintf(stderr,"OOM multiplex payload\n"); exit(1); } if(fread(raw,1,(size_t)sub.bytes,stdin)!=(size_t)sub.bytes){ free(raw); free(line); return -1; } + char *gtxt=NULL; /* optional per-request grammar/schema text */ + if(sub.gbytes){ + gtxt=malloc((size_t)sub.gbytes+1); + if(!gtxt){ fprintf(stderr,"OOM multiplex payload\n"); exit(1); } + if(fread(gtxt,1,(size_t)sub.gbytes,stdin)!=(size_t)sub.gbytes){ + free(gtxt); free(raw); free(line); return -1; } + gtxt[sub.gbytes]=0; + } int delim=fgetc(stdin); if(delim!='\n'){ printf("ERROR %llu BAD_FRAME\n",sub.id); fflush(stdout); - free(raw); free(line); return -1; + free(gtxt); free(raw); free(line); return -1; } raw[sub.bytes]=0; if(sub.slot>=nctx || memchr(raw,0,(size_t)sub.bytes)){ - printf("ERROR %llu BAD_REQUEST\n",sub.id); fflush(stdout); free(raw); free(line); return 0; + printf("ERROR %llu BAD_REQUEST\n",sub.id); fflush(stdout); free(gtxt); free(raw); free(line); return 0; } if(req[sub.slot].active){ - printf("ERROR %llu SLOT_BUSY\n",sub.id); fflush(stdout); free(raw); free(line); return 0; + printf("ERROR %llu SLOT_BUSY\n",sub.id); fflush(stdout); free(gtxt); free(raw); free(line); return 0; } for(int i=0;ikv); int *tmp=malloc(maxctx*sizeof(int)); if(!tmp){ fprintf(stderr,"OOM mux_submit tmp\n"); free(raw); free(line); exit(1); } @@ -4427,6 +5114,7 @@ static int mux_submit(Model *m, Tok *T, ServeCtx *ctx, ServeReq *req, int nctx, int next=pick_tok(logit,m->c.vocab,-1); free(logit); if(r->maximum<=0 || next==eos || is_stop(next)){ mux_done(m,sc,r); return 1; } r->pending=next; r->emitted=1; r->active=1; sc->hist[sc->len]=next; m->n_emit++; + if(grd[sub.slot].on){ grammar_reset(&grd[sub.slot]); gr_feed(&grd[sub.slot],next); } mux_data(T,r->id,next); if(r->emitted>=r->maximum) mux_done(m,sc,r); return 1; @@ -4435,14 +5123,18 @@ static int mux_submit(Model *m, Tok *T, ServeCtx *ctx, ServeReq *req, int nctx, static void run_serve_mux(Model *m, const char *snap){ char tkp[2048]; snprintf(tkp,sizeof(tkp),"%s/tokenizer.json",snap); Tok T; tok_load(&T,tkp); int eos=tok_id_of(&T,"<|endoftext|>"); stops_arm_tok(&m->c,eos,&T); - g_draft=0; /* one scheduler owns every forward; MTP/speculation is not ragged-safe */ + g_draft=0; /* one scheduler owns every forward; MTP/n-gram speculation is not ragged-safe. + * Grammar-forced drafts ARE mux-safe (below): a drafting slot leaves the shared + * batch for one forward and runs the proven single-sequence verify path + * (kv_bind + step_all), exactly like prefill already does per submission. */ int maxctx=getenv("CTX")?atoi(getenv("CTX")):4096; int nctx=getenv("KV_SLOTS")?atoi(getenv("KV_SLOTS")):1; if(nctx<1||nctx>512){fprintf(stderr,"KV_SLOTS must be between 1 and 512\n");exit(2);} g_kvsave=getenv("KVSAVE")?atoi(getenv("KVSAVE")):1; KVState *initial=m->kv; free(initial->kv_start); free(initial); ServeCtx *ctx=calloc(nctx,sizeof(*ctx)); ServeReq *req=calloc(nctx,sizeof(*req)); - for(int i=0;i0)?1:0; if(ready) #endif - if(mux_submit(m,&T,ctx,req,nctx,maxctx,eos)<0) eof=1; + if(mux_submit(m,&T,ctx,req,grd,nctx,maxctx,eos)<0) eof=1; } active=0; for(int i=0;ion && r->temp==0) k=grammar_draft(gd,draft,gd->max); + if(k>0 && sc->len+1+klen+1+k>=(int)sc->kv.max_t) k=(int)sc->kv.max_t-sc->len-2; + if(k<1){ rows[S]=(DecodeRow){&sc->kv,r->pending,sc->len}; slots[S++]=i; continue; } + kv_bind(m,&sc->kv); + int seq[50]; seq[0]=r->pending; + memcpy(seq+1,draft,(size_t)k*sizeof(int)); + float *lo=step_all(m,seq,1+k,sc->len); m->n_fw++; + gd->prop+=(uint64_t)k; + int done=0; + for(int j=0;j<=k && !done;j++){ + g_temp=r->temp; g_nuc=r->top_p; + int next=pick_tok(lo+(int64_t)j*m->c.vocab,m->c.vocab,-1); + sc->len++; /* seq[j] joins the committed history */ + if(next==eos || is_stop(next)){ mux_done(m,sc,r); done=1; break; } + r->pending=next; sc->hist[sc->len]=next; r->emitted++; m->n_emit++; + if(gd->on) gr_feed(gd,next); + mux_data(&T,r->id,next); + if(r->emitted>=r->maximum){ mux_done(m,sc,r); done=1; break; } + if(jacc++; + } + } + free(lo); + continue; /* handled outside the shared batch */ + } + rows[S]=(DecodeRow){&sc->kv,r->pending,sc->len}; slots[S++]=i; } + if(S==0) continue; /* every active slot drafted this round */ double tf0=g_prof?now_s():0; float *lo=step_decode_batch(m,rows,S); if(!lo){fprintf(stderr,"decode batch failed\n");break;} m->n_fw++; @@ -4508,13 +5233,15 @@ static void run_serve_mux(Model *m, const char *snap){ int next=pick_tok(lo+(int64_t)s*m->c.vocab,m->c.vocab,-1); if(next==eos || is_stop(next)){mux_done(m,sc,r);continue;} r->pending=next; sc->hist[sc->len]=next; r->emitted++; m->n_emit++; + if(grd[i].on) gr_feed(&grd[i],next); /* walker stays in sync when not drafting */ mux_data(&T,r->id,next); if(r->emitted>=r->maximum) mux_done(m,sc,r); } free(lo); } usage_save(m); - for(int i=0;ikv=NULL; m->Lc=m->Rc=m->Ic=NULL; m->kv_start=NULL; m->max_t=0; } @@ -4538,7 +5265,7 @@ static void run_serve(Model *m, const char *snap){ Tok T; tok_load(&T,tkp); int eos=tok_id_of(&T,"<|endoftext|>"); stops_arm_tok(&m->c, eos, &T); - grammar_setup(&T); /* metodo F: GRAMMAR=file.gbnf (#48) */ + grammar_setup(&g_grd,&T); /* metodo F: GRAMMAR=file.gbnf (#48) */ if(g_temp<0) g_temp=0.7f; /* auto: 0.7, NON l'1.0 ufficiale — la coda della * distribuzione int4 e' rumore di quantizzazione */ int ngen=getenv("NGEN")?atoi(getenv("NGEN")):256; @@ -4657,7 +5384,7 @@ static void run_serve(Model *m, const char *snap){ else logit=step(m,hist+len-1,1,len-1); /* prompt identico/prefisso: rigenera i logits */ EmitStream es={&T,m,now_s(),0,1}; int prod=0; - grammar_reset(); /* nuova risposta = nuovo documento (MORE invece continua) */ + grammar_reset(&g_grd); /* nuova risposta = nuovo documento (MORE invece continua) */ if(cur>0) prod=spec_decode(m,hist,len,cur,eos,logit,emit_stream,&es,&len); else free(logit); double tdt=now_s()-tt0; if(tdt<1e-6) tdt=1e-6; @@ -4870,6 +5597,56 @@ typedef struct { int l,e; uint32_t c; } PinRec; static int pin_rec_cmp(const void *a,const void *b){ const PinRec *x=a,*y=b; return x->cc?1:x->c>y->c?-1:0; } + +#ifdef __linux__ +/* #419: bind the pinned hot-store as ONE arena per layer instead of one mbind + * per slab. Per-slab policies cost ~2 unmergeable VMAs each; a PIN_GB=all load + * (19,456 experts x slab+fslab) crosses the default vm.max_map_count=65530 and + * posix_memalign dies with terabytes free. Experts of one layer share a tensor + * shape, so a layer's pins pack into two arenas (weights + scales) at a fixed + * stride: 2 binds and a handful of VMAs per layer instead of ~500. Slices are + * pre-attached to the slots (slab_cap covers expert_load's realloc check, so + * its alloc branch never fires); aslab marks arena ownership for the + * release/ensure paths. Arena-OOM just leaves the slots on the individual path. */ +static void pin_arena_bind(Model *m, PinRec *r, int *slot_of, int from, int to){ + if(g_numa_nodes<2 || g_mmap || from>=to) return; + Cfg *c=&m->c; int NR=c->n_layers+1; + int *cnt=calloc((size_t)NR,sizeof(int)); int *first=malloc((size_t)NR*sizeof(int)); + if(!cnt||!first){ free(cnt); free(first); return; } + for(int i=0;iS,nm), *tq=st_find(&m->S,qn); + if(!tw||!tq) ok=0; else { wtot+=tw->nbytes; qtot+=tq->nbytes; } + } + if(!ok) continue; /* unquantized fallback: individual allocs */ + size_t ws=((size_t)wtot+8192+4095)&~(size_t)4095; + size_t fs=((size_t)(qtot/4)*sizeof(float)+4095)&~(size_t)4095; + uint8_t *aw=NULL; float *af=NULL; + if(posix_memalign((void**)&aw,4096,(size_t)cnt[l]*ws)) continue; + if(posix_memalign((void**)&af,4096,(size_t)cnt[l]*fs)){ free(aw); continue; } + numa_slab_bind(aw,(size_t)cnt[l]*ws); + numa_slab_bind(af,(size_t)cnt[l]*fs); + int i=0; + for(int a=from;apin[l][slot_of[a]]; + s->slab=aw+(size_t)i*ws; s->slab_cap=(int64_t)ws; s->aslab=s->slab; + s->fslab=(float*)((uint8_t*)af+(size_t)i*fs); + s->fslab_cap=(int64_t)(fs/sizeof(float)); s->afslab=s->fslab; + i++; + } + } + free(cnt); free(first); +} +#endif static double expert_avail(Model *m, double ram_gb, int ebits, int max_ctx); /* def. sotto */ static void pin_load(Model *m, const char *statspath, double gb){ FILE *f=fopen(statspath,"r"); if(!f){ perror(statspath); return; } @@ -4908,6 +5685,29 @@ static void pin_load(Model *m, const char *statspath, double gb){ double avail=expert_avail(m,ram_env,m->ebits,est_ctx); npin=avail>0?(int)(avail/eb):0; } else npin=(int)(gb*1e9/eb); +#ifdef COLI_CUDA + /* The VRAM budget must be known BEFORE npin is finalized: with + * CUDA_RELEASE_HOST the VRAM-ranked prefix's host slabs are freed right + * after upload, so those slots must NOT consume the RAM pin budget. + * Before this, a 6-GPU host lost its top ~9k ranked experts from the RAM + * count and pinned only the leftovers (measured: 9,280 VRAM + 1,721 RAM + * on a box whose RAM fits ~11k — the cold tail then paid disk forever). */ + double remaining[COLI_CUDA_MAX_DEVICES]={0}, placed_b[COLI_CUDA_MAX_DEVICES]={0}; + int placed_n[COLI_CUDA_MAX_DEVICES]={0}, gpu_prefix=0, prefix_est=0; + double budget=g_cuda_expert_gb*1e9, safe_total=0; + if(g_cuda_enabled&&(g_cuda_expert_gb>0||g_cuda_expert_auto)) for(int i=0;isafe_total) budget=safe_total; + if(g_cuda_enabled&&g_cuda_release_host&&budget>0){ + prefix_est=(int)(budget/eb)+g_cuda_ndev; + npin+=prefix_est; /* additive: prefix RAM is returned after upload */ + } +#endif if(npin>n) npin=n; if(npin<1){ free(r); return; } int *cnt_l=calloc(c->n_layers+1,sizeof(int)); /* +1: riga MTP */ @@ -4918,26 +5718,26 @@ static void pin_load(Model *m, const char *statspath, double gb){ for(int i=0;i<=c->n_layers;i++) m->npin[i]=cnt_l[i]; double t0=now_s(); #ifdef COLI_CUDA - double remaining[COLI_CUDA_MAX_DEVICES]={0}, placed_b[COLI_CUDA_MAX_DEVICES]={0}; - int placed_n[COLI_CUDA_MAX_DEVICES]={0}, gpu_prefix=0; - double budget=g_cuda_expert_gb*1e9, safe_total=0; - if(g_cuda_enabled&&(g_cuda_expert_gb>0||g_cuda_expert_auto)) for(int i=0;isafe_total) budget=safe_total; - if(g_cuda_enabled&&g_cuda_release_host&&budget>0){ gpu_prefix=(int)(budget/eb)+g_cuda_ndev; if(gpu_prefix>npin)gpu_prefix=npin; } + if(prefix_est>0){ gpu_prefix=prefix_est; if(gpu_prefix>npin) gpu_prefix=npin; } #else int gpu_prefix=0; +#endif +#ifdef __linux__ + /* CPU-resident pins only: the GPU prefix stays on individual allocs (its + * host backing is released after upload; arena slices are never freed). */ + pin_arena_bind(m,r,slot_of,gpu_prefix,npin); #endif /* Load the VRAM-ranked prefix first. Once uploaded its host backing is * released before the disjoint RAM-ranked suffix is allocated. */ +#ifdef __linux__ + if(gpu_prefix>0) g_numa_skip_bind=1; /* prefix slabs = transient upload staging: don't bind (#419) */ +#endif #pragma omp parallel for schedule(dynamic,1) for(int a=0;a<(gpu_prefix?gpu_prefix:npin);a++) - expert_load(m,r[a].l,r[a].e,&m->pin[r[a].l][slot_of[a]],1); + expert_load(m,r[a].l,r[a].e,&m->pin[r[a].l][slot_of[a]],1,0); /* startup pin load; demand=0, never classified */ +#ifdef __linux__ + g_numa_skip_bind=0; +#endif m->resident_bytes+=(int64_t)(gpu_prefix?gpu_prefix:npin)*eb; #ifdef COLI_CUDA if(g_cuda_enabled && budget>0){ @@ -4981,7 +5781,7 @@ static void pin_load(Model *m, const char *statspath, double gb){ if(gpu_prefix>0&&gpu_prefixpin[r[a].l][slot_of[a]],1); + expert_load(m,r[a].l,r[a].e,&m->pin[r[a].l][slot_of[a]],1,0); /* startup pin load; demand=0, never classified */ m->resident_bytes+=(int64_t)(npin-gpu_prefix)*eb; } fprintf(stderr,"[PIN] placement: %d VRAM + %d RAM expert (%.1f GB warm) in %.0fs da %s\n", @@ -5317,9 +6117,16 @@ int main(int argc, char **argv){ 0 #endif ; + if(pipe_workers_imply_pipe(getenv("PIPE"),getenv("PIPE_WORKERS"),g_pipe)){ + g_pipe=1; + fprintf(stderr,"[PIPE] PIPE_WORKERS is set — enabling the async pipe (PIPE=1 implied; set PIPE=0 to override)\n"); + } g_pipe_nw = getenv("PIPE_WORKERS")?atoi(getenv("PIPE_WORKERS")):8; /* I/O worker threads */ if(g_pipe_nw<1) g_pipe_nw=1; + g_pipe_block = getenv("COLI_PIPE_BLOCK")?atoi(getenv("COLI_PIPE_BLOCK")):0; /* blocking pipe_wait (default: spin) */ g_direct = getenv("DIRECT")?atoi(getenv("DIRECT")):0; + { const char *dh=getenv("COLI_DISKCLASS_WINDOW"); /* DISK-CLASS recency window, see its declaration */ + if(dh){ g_direct_heat_ticks=(uint32_t)strtoul(dh,NULL,10); g_direct_heat_explicit=1; } } g_uring = getenv("URING")?atoi(getenv("URING")):0; if(g_uring){ #ifdef __linux__ @@ -5382,6 +6189,7 @@ int main(int argc, char **argv){ } g_cuda_dense=getenv("CUDA_DENSE")?atoi(getenv("CUDA_DENSE")):0; g_cuda_pipe=getenv("COLI_CUDA_PIPE")?atoi(getenv("COLI_CUDA_PIPE")):0; + g_cuda_router=getenv("COLI_CUDA_ROUTER")?atoi(getenv("COLI_CUDA_ROUTER")):0; const char *cuda_expert=getenv("CUDA_EXPERT_GB"); g_cuda_expert_auto=cuda_expert&&!strcmp(cuda_expert,"auto"); g_cuda_expert_gb=cuda_expert&&!g_cuda_expert_auto?atof(cuda_expert):0; @@ -5421,6 +6229,19 @@ int main(int argc, char **argv){ printf("== GLM C engine (glm_moe_dsa), cache=%d experts/layer | experts@%d-bit dense@%d-bit | idot: " IDOT_KERNEL " ==\n", cap, ebits, dbits); g_mem_avail_boot = mem_available_gb(); Model m; double t0=now_s(); model_init(&m,snap,cap,ebits,dbits); + if(!g_direct_heat_explicit){ /* COLI_DISKCLASS_WINDOW default, needs m.c (topk/n_layers) */ + /* CURRENT-STATE CALIBRATION: the "8" multiplier (recency window ~= the last 8 + * tokens' worth of routing) is a measured-config constant, not a derived truth. + * Coordinates: measured 2026-07, macOS 26.5, M5 Max 128 GB, base caa49f7, + * GLM-5.2 int4 (topk=8, n_layers=78). On that setup it splits the classes + * cleanly (decode cold share ~71% of classified bytes, 1061.85/1491.49 GB); + * other models, page-cache pressures, or disks may want a different window -- + * COLI_DISKCLASS_WINDOW overrides, and the DISK-CLASS line itself is the + * tuning instrument. */ + uint32_t nl=(uint32_t)(m.c.n_layers>0?m.c.n_layers:1), k=(uint32_t)(m.c.topk>0?m.c.topk:1); + g_direct_heat_ticks = k*nl*8u; /* ~last 8 tokens' worth of routing, see the declaration */ + if(!g_direct_heat_ticks) g_direct_heat_ticks=1; + } if(g_draft<0){ #ifdef COLI_CUDA /* MTP is disabled under CUDA by default: cold (streaming) experts still diff --git a/c/decode_batch.h b/c/decode_batch.h index ad3d2ed..9f4dd34 100644 --- a/c/decode_batch.h +++ b/c/decode_batch.h @@ -13,22 +13,32 @@ static inline float *coli_kv_row(float *base, int position, int width) } typedef struct { - unsigned long long id, bytes; + unsigned long long id, bytes, gbytes; int slot, max_tokens; float temperature, top_p; } ColiSubmit; /* Parse the textual header. The payload is read separately using `bytes`, so - * it may contain newlines. Reject trailing fields to keep framing unambiguous. */ + * it may contain newlines. Reject trailing fields to keep framing unambiguous. + * Optional 7th field `gbytes`: length of a per-request grammar (raw GBNF, or a + * JSON-Schema compiled engine-side) appended to the payload AFTER the prompt + * bytes. 6-field headers remain valid (gbytes = 0). */ static inline int coli_submit_parse(const char *line, ColiSubmit *s) { char tail; - if (!line || !s || - sscanf(line, "SUBMIT %llu %d %llu %d %f %f %c", &s->id, &s->slot, + if (!line || !s) return 0; + s->gbytes = 0; + if (sscanf(line, "SUBMIT %llu %d %llu %d %f %f %llu %c", &s->id, &s->slot, &s->bytes, &s->max_tokens, &s->temperature, &s->top_p, - &tail) != 6) - return 0; - return s->id > 0 && s->bytes <= (16u << 20) && s->slot >= 0 && s->max_tokens >= 1 && + &s->gbytes, &tail) != 7) { + s->gbytes = 0; + if (sscanf(line, "SUBMIT %llu %d %llu %d %f %f %c", &s->id, &s->slot, + &s->bytes, &s->max_tokens, &s->temperature, &s->top_p, + &tail) != 6) + return 0; + } + return s->id > 0 && s->bytes <= (16u << 20) && s->gbytes <= (1u << 20) && + s->slot >= 0 && s->max_tokens >= 1 && isfinite(s->temperature) && isfinite(s->top_p) && s->temperature >= 0 && s->temperature <= 2 && s->top_p > 0 && s->top_p <= 1; diff --git a/c/openai_server.py b/c/openai_server.py index 3900ba6..ae1032c 100644 --- a/c/openai_server.py +++ b/c/openai_server.py @@ -370,10 +370,46 @@ def render_chat(messages, enable_thinking=False, reasoning_effort=None, tools=No return "".join(prompt) +# Generic whitespace-tolerant JSON grammar for response_format {"type": "json_object"}. +# Draft-source semantics: positions with one legal byte draft; jws points just keep +# the walker alive through the model's own spacing (see docs/grammar-draft.md). +GENERIC_JSON_GBNF = ( + 'root ::= jws jval jws\n' + 'jval ::= jobj | jarr | jstr | jnum | "true" | "false" | "null"\n' + 'jobj ::= "{" jws ( jstr jws ":" jws jval jws ( "," jws jstr jws ":" jws jval jws )* )? "}"\n' + 'jarr ::= "[" jws ( jval jws ( "," jws jval jws )* )? "]"\n' + 'jstr ::= "\\"" jchar* "\\""\n' + 'jchar ::= [^"\\\\\\x00-\\x1f] | "\\\\" ( ["\\\\/bfnrt] | "u" jhex jhex jhex jhex )\n' + 'jhex ::= [0-9a-fA-F]\n' + 'jnum ::= "-"? ( "0" | [1-9] [0-9]* ) ( "." [0-9]+ )? ( ( "e" | "E" ) ( "+" | "-" )? [0-9]+ )?\n' + 'jws ::= ( " " | "\\t" | "\\n" | "\\r" )*\n' +) + def generation_options(body, limit): if body.get("n", 1) != 1: raise APIError(400, "Colibri currently supports `n=1` only.", "n", "unsupported_value") # `tools`/`functions` are handled by render_chat (declaration) + parse_tool_calls (output). + # Validate tools/functions structure early so malformed input fails with a clear error. + tools_raw = body.get("tools") or body.get("functions") + if tools_raw is not None: + if not isinstance(tools_raw, list): + raise APIError(400, "`tools` must be a non-empty array.", "tools", "invalid_value") + if not tools_raw: + raise APIError(400, "`tools` must be a non-empty array.", "tools", "invalid_value") + for idx, tool in enumerate(tools_raw): + if not isinstance(tool, dict): + raise APIError(400, f"Each tool must be an object, got {type(tool).__name__} at index {idx}.", + f"tools.{idx}", "invalid_value") + fn = tool.get("function", tool) if isinstance(tool, dict) else {} + if not isinstance(fn, dict): + raise APIError(400, f"Tool function must be an object at index {idx}.", + f"tools.{idx}.function", "invalid_value") + if not fn.get("name"): + raise APIError(400, f"Each tool must have a `name` at index {idx}.", + f"tools.{idx}.function.name", "invalid_value") + if not isinstance(fn["name"], str): + raise APIError(400, f"Tool `name` must be a string at index {idx}.", + f"tools.{idx}.function.name", "invalid_value") choice = body.get("tool_choice") if choice is not None: if isinstance(choice, str): @@ -403,10 +439,38 @@ def generation_options(body, limit): raise APIError(400, "Token penalties are not supported yet.", None, "unsupported_parameter") if body.get("seed") is not None: raise APIError(400, "Per-request seeds are not supported yet.", "seed", "unsupported_parameter") + # response_format -> optional per-request grammar for the engine's grammar-forced + # draft source (#70/#148). NEVER a sampling constraint: drafts are verified, so a + # schema the engine cannot compile degrades to "no speedup", not to an error and + # not to changed output. json_schema payloads are forwarded as-is (the engine + # compiles them via schema_gbnf.h); {"type": "gbnf"} is a raw-GBNF extension. + grammar = None response_format = body.get("response_format") - if response_format not in (None, {"type": "text"}): - raise APIError(400, "Only the default text response format is supported.", - "response_format", "unsupported_parameter") + if response_format is not None and response_format != {"type": "text"}: + if not isinstance(response_format, dict) or "type" not in response_format: + raise APIError(400, "`response_format` must be an object with a `type`.", + "response_format", "invalid_value") + ftype = response_format["type"] + if ftype == "json_object": + grammar = GENERIC_JSON_GBNF + elif ftype == "json_schema": + schema = (response_format.get("json_schema") or {}).get("schema") + if not isinstance(schema, dict): + raise APIError(400, "`response_format.json_schema.schema` must be an object.", + "response_format", "invalid_value") + grammar = json.dumps(schema) + elif ftype == "gbnf": + grammar = response_format.get("grammar") + if not isinstance(grammar, str) or not grammar.strip(): + raise APIError(400, "`response_format.grammar` must be a non-empty GBNF string.", + "response_format", "invalid_value") + else: + raise APIError(400, "`response_format.type` must be \"text\", \"json_object\", " + "\"json_schema\" or \"gbnf\".", + "response_format", "unsupported_value") + if grammar is not None and len(grammar.encode("utf-8")) > (1 << 20): + raise APIError(400, "`response_format` grammar/schema exceeds 1 MiB.", + "response_format", "invalid_value") maximum = body.get("max_completion_tokens") maximum_param = "max_completion_tokens" @@ -433,7 +497,7 @@ def generation_options(body, limit): if (isinstance(top_p, bool) or not isinstance(top_p, (int, float)) or not math.isfinite(top_p) or not 0 < top_p <= 1): raise APIError(400, "`top_p` must be greater than 0 and at most 1.", "top_p") - return maximum, float(temperature), float(top_p) + return maximum, float(temperature), float(top_p), grammar def read_engine_turn(stream, sentinel, on_bytes): @@ -597,12 +661,15 @@ class Engine: self._fail_pending(error) def generate(self, prompt, max_tokens, temperature, top_p, on_text, cache_slot=0, - cancelled=None): + cancelled=None, grammar=None): if isinstance(cache_slot, bool) or not isinstance(cache_slot, int) or not 0 <= cache_slot < self.kv_slots: raise APIError(400, "Invalid cache slot.", "cache_slot") payload = prompt.encode("utf-8") if b"\0" in payload: raise APIError(400, "NUL bytes are not supported in prompts.", "messages") + gpayload = grammar.encode("utf-8") if grammar else b"" + if b"\0" in gpayload: + raise APIError(400, "NUL bytes are not supported in grammars.", "response_format") decoder = codecs.getincrementaldecoder("utf-8")("replace") def decode(data): @@ -622,12 +689,13 @@ class Engine: self.next_request_id += 1 self.pending[request_id] = events header = (f"SUBMIT {request_id} {cache_slot} {len(payload)} {max_tokens} " - f"{temperature:.8g} {top_p:.8g}\n").encode() + f"{temperature:.8g} {top_p:.8g}" + + (f" {len(gpayload)}" if gpayload else "") + "\n").encode() try: with self.write_lock: if self.process.poll() is not None: raise RuntimeError("colibri engine is not running") - self.process.stdin.write(header + payload + b"\n") + self.process.stdin.write(header + payload + gpayload + b"\n") self.process.stdin.flush() except Exception: with self.pending_lock: @@ -863,7 +931,7 @@ class APIHandler(BaseHTTPRequestHandler): except OSError: pass - def generation(self, body, prompt, request_id, chat): + def generation(self, body, prompt, request_id, chat, tools=None, tool_choice=None): # COLI_DEBUG tees the engine transaction to stderr: 1 = decoded output stream only, # 2 = both sides (rendered prompt + output). render_chat already folds prior turns and # tool results into `prompt`, so level 2 is the full conversation the engine saw. @@ -874,9 +942,9 @@ class APIHandler(BaseHTTPRequestHandler): if dbg >= 2: sys.stderr.write(f"\n===== PROMPT [{request_id}] =====\n{prompt}\n===== OUTPUT [{request_id}] =====\n") sys.stderr.flush() - maximum, temperature, top_p = generation_options(body, self.server.max_tokens) - tools = (body.get("tools") or body.get("functions") or None) if chat else None - if body.get("tool_choice") == "none": + maximum, temperature, top_p, grammar = generation_options(body, self.server.max_tokens) + # tools and tool_choice come from chat_completion() already processed/filtered + if chat and tool_choice == "none": tools = None # client forbade tools: never surface tool_calls cache_slot = body.get("cache_slot") if (cache_slot is not None and @@ -903,7 +971,7 @@ class APIHandler(BaseHTTPRequestHandler): output = [] stats = self.server.engine.generate( prompt, maximum, temperature, top_p, output.append, cache_slot, - self.client_disconnected) + self.client_disconnected, grammar=grammar) text = "".join(output) length_finish = "length" if stats["length_limited"] else "stop" if chat and tools: @@ -1010,7 +1078,7 @@ class APIHandler(BaseHTTPRequestHandler): sp["buf"] = sp["buf"][flush:] stats = self.server.engine.generate( prompt, maximum, temperature, top_p, emit_tools, cache_slot, - lambda: not connected) + lambda: not connected, grammar=grammar) if not sp["tool"] and sp["buf"]: emit(sp["buf"]) # no tool call happened: flush held tail _content, calls = parse_tool_calls("".join(raw), tools) @@ -1027,7 +1095,7 @@ class APIHandler(BaseHTTPRequestHandler): emit(chunk) stats = self.server.engine.generate( prompt, maximum, temperature, top_p, emit_plain, cache_slot, - lambda: not connected) + lambda: not connected, grammar=grammar) finish = "length" if stats["length_limited"] else "stop" ka_stop.set() # generation done: stop the keepalive pump ka_thread.join(timeout=2) @@ -1078,9 +1146,10 @@ class APIHandler(BaseHTTPRequestHandler): if not isinstance(enable_thinking, bool): raise APIError(400, "`enable_thinking` must be a boolean.", "enable_thinking") tools = body.get("tools") or body.get("functions") or None + tool_choice = body.get("tool_choice") prompt = render_chat(body.get("messages"), enable_thinking, reasoning_effort, tools, - body.get("tool_choice")) - self.generation(body, prompt, request_id, True) + tool_choice) + self.generation(body, prompt, request_id, True, tools, tool_choice) def completion(self, body, request_id): prompt = body.get("prompt") diff --git a/c/quant.h b/c/quant.h index 928de96..b375538 100644 --- a/c/quant.h +++ b/c/quant.h @@ -268,6 +268,68 @@ static void matmul_i2(float *y, const float *x, const uint8_t *q2, const float * y[(int64_t)s*O+o]=a*sc; } } } +/* ---- int3-g64 (fmt=5): 3-bit weights with ONE f32 scale per 64-input group - + * Per group: 16B low plane (2 bits/val, int2 layout) + 8B high plane (1 bit/val), + * values in [-4,3] stored v+4. 3.5 bits/weight effective — the quality/size point + * the #132 OLMoE ablation measured BEATING per-row int4. */ +#define I3_GROUP 64 +#define I3_GBYTES 24 /* 16B low plane + 8B high plane per group */ +static inline int64_t i3_groups(int I){ return ((int64_t)I + I3_GROUP - 1) / I3_GROUP; } +static inline int64_t i3_rowbytes(int I){ return i3_groups(I) * I3_GBYTES; } + +/* Dequant-on-use with PER-GROUP scale. Exact f32 path only (no IDOT in v1: int8 + * activations don't compose with per-group accumulation without a kernel + * restructure — follow-up). NEON: low plane = matmul_i2's unpack, high plane + * expanded via vtst on bit masks; x86 stays scalar for now (follow-up). */ +static void matmul_i3(float *y, const float *x, const uint8_t *q3, const float *scale, int S, int I, int O){ + int64_t ng=i3_groups(I), rb=i3_rowbytes(I); + #pragma omp parallel for schedule(static) + for(int o=0;o>2), 4); /* 4 bytes = 16 low-plane values */ + uint8x8_t by=vreinterpret_u8_u32(vdup_n_u32(wd)); + uint8x8x2_t z01=vzip_u8(vand_u8(by,m2v), vand_u8(vshr_n_u8(by,2),m2v)); + uint8x8x2_t z23=vzip_u8(vand_u8(vshr_n_u8(by,4),m2v), vshr_n_u8(by,6)); + uint16x4x2_t zz=vzip_u16(vreinterpret_u16_u8(z01.val[0]), vreinterpret_u16_u8(z23.val[0])); + uint8x16_t lov=vcombine_u8(vreinterpret_u8_u16(zz.val[0]), vreinterpret_u8_u16(zz.val[1])); + uint8x16_t hv=vcombine_u8(vdup_n_u8(hi[k>>3]), vdup_n_u8(hi[(k>>3)+1])); + uint8x16_t hb=vandq_u8(vtstq_u8(hv,bitm), fourq); /* 4 where high bit set */ + int8x16_t wq=vsubq_s8(vreinterpretq_s8_u8(vaddq_u8(lov,hb)), b4q); /* [-4,3] in order */ + int16x8_t w0=vmovl_s8(vget_low_s8(wq)), w1=vmovl_s8(vget_high_s8(wq)); + ac0=vfmaq_f32(ac0, vld1q_f32(xs+base+k), vcvtq_f32_s32(vmovl_s16(vget_low_s16(w0)))); + ac1=vfmaq_f32(ac1, vld1q_f32(xs+base+k+4), vcvtq_f32_s32(vmovl_s16(vget_high_s16(w0)))); + ac0=vfmaq_f32(ac0, vld1q_f32(xs+base+k+8), vcvtq_f32_s32(vmovl_s16(vget_low_s16(w1)))); + ac1=vfmaq_f32(ac1, vld1q_f32(xs+base+k+12), vcvtq_f32_s32(vmovl_s16(vget_high_s16(w1)))); + } + a=vaddvq_f32(vaddq_f32(ac0,ac1)); + } +#endif + for(;k>2]>>((k&3)*2))&3) | (((hi[k>>3]>>(k&7))&1)<<2); + a += xs[base+k]*(float)((int)u-4); + } + acc += a*srow[g]; + } + y[(int64_t)s*O+o]=acc; + } + } +} + /* ---- IDOT: integer dot kernels (int8-quantized activations) --------------- */ #if defined(__AVX512VNNI__) && defined(__AVX512BW__) #define IDOT_KERNEL "avx512-vnni" @@ -314,12 +376,27 @@ static inline int32_t dot_i8i8(const int8_t *w, const int8_t *x, int I){ } sum=_mm512_reduce_add_epi32(acc); #elif defined(__AVXVNNI__) && defined(__AVX2__) - __m128i acc=_mm_setzero_si128(); + /* 4 accumulatori indipendenti (64 byte/iter): un solo acc incatena i vpdpbusd + * (latenza-bound ~5c). Somme intere associative -> bit-identico. Stessa struttura + * dei 4 accumulatori del ramo NEON piu' sotto. + * EN: four independent accumulators break the serial vpdpbusd->acc chain; integer + * adds are associative, so the result is bit-identical (mirrors the NEON path). */ + __m128i a0=_mm_setzero_si128(),a1=_mm_setzero_si128(),a2=_mm_setzero_si128(),a3=_mm_setzero_si128(); + for(;i+64<=I;i+=64){ + __m128i w0=_mm_loadu_si128((const __m128i*)(w+i)), x0=_mm_loadu_si128((const __m128i*)(x+i)); + __m128i w1=_mm_loadu_si128((const __m128i*)(w+i+16)), x1=_mm_loadu_si128((const __m128i*)(x+i+16)); + __m128i w2=_mm_loadu_si128((const __m128i*)(w+i+32)), x2=_mm_loadu_si128((const __m128i*)(x+i+32)); + __m128i w3=_mm_loadu_si128((const __m128i*)(w+i+48)), x3=_mm_loadu_si128((const __m128i*)(x+i+48)); + a0=_mm_dpbusd_epi32(a0,_mm_abs_epi8(w0),_mm_sign_epi8(x0,w0)); + a1=_mm_dpbusd_epi32(a1,_mm_abs_epi8(w1),_mm_sign_epi8(x1,w1)); + a2=_mm_dpbusd_epi32(a2,_mm_abs_epi8(w2),_mm_sign_epi8(x2,w2)); + a3=_mm_dpbusd_epi32(a3,_mm_abs_epi8(w3),_mm_sign_epi8(x3,w3)); + } + __m128i acc=_mm_add_epi32(_mm_add_epi32(a0,a1),_mm_add_epi32(a2,a3)); for(;i+16<=I;i+=16){ __m128i wv=_mm_loadu_si128((const __m128i*)(w+i)); __m128i xv=_mm_loadu_si128((const __m128i*)(x+i)); - __m128i xs=_mm_sign_epi8(xv,wv); - acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(wv),xs); + acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(wv),_mm_sign_epi8(xv,wv)); } sum=hsum128_i32(acc); #elif defined(__AVX2__) @@ -390,13 +467,32 @@ static inline int32_t dot_i4i8(const uint8_t *w4, const int8_t *x, int I){ } sum=_mm512_reduce_add_epi32(acc); #elif defined(__AVXVNNI__) && defined(__AVX2__) + /* 4 accumulatori indipendenti (64 elementi = 32 byte packed/iter): un solo acc + * incatena i vpdpbusd (latenza-bound ~5c). Somme intere associative -> bit-identico. + * Stessa struttura dei 4 accumulatori del ramo NEON piu' sotto. + * EN: four independent accumulators break the serial vpdpbusd->acc chain; integer + * adds are associative, so the result is bit-identical (mirrors the NEON path). */ const __m128i m4=_mm_set1_epi8(0x0F); const __m128i b8=_mm_set1_epi8(8); - __m128i acc=_mm_setzero_si128(); - for(;i+32<=I;i+=32){ + __m128i a0=_mm_setzero_si128(),a1=_mm_setzero_si128(),a2=_mm_setzero_si128(),a3=_mm_setzero_si128(); + for(;i+64<=I;i+=64){ + __m128i by0=_mm_loadu_si128((const __m128i*)(w4+(i>>1))); /* elem i..i+31 */ + __m128i by1=_mm_loadu_si128((const __m128i*)(w4+(i>>1)+16)); /* elem i+32..i+63 */ + __m128i lo0=_mm_and_si128(by0,m4), hi0=_mm_and_si128(_mm_srli_epi16(by0,4),m4); + __m128i lo1=_mm_and_si128(by1,m4), hi1=_mm_and_si128(_mm_srli_epi16(by1,4),m4); + __m128i w0=_mm_sub_epi8(_mm_unpacklo_epi8(lo0,hi0),b8), w1=_mm_sub_epi8(_mm_unpackhi_epi8(lo0,hi0),b8); + __m128i w2=_mm_sub_epi8(_mm_unpacklo_epi8(lo1,hi1),b8), w3=_mm_sub_epi8(_mm_unpackhi_epi8(lo1,hi1),b8); + __m128i x0=_mm_loadu_si128((const __m128i*)(x+i)), x1=_mm_loadu_si128((const __m128i*)(x+i+16)); + __m128i x2=_mm_loadu_si128((const __m128i*)(x+i+32)), x3=_mm_loadu_si128((const __m128i*)(x+i+48)); + a0=_mm_dpbusd_epi32(a0,_mm_abs_epi8(w0),_mm_sign_epi8(x0,w0)); + a1=_mm_dpbusd_epi32(a1,_mm_abs_epi8(w1),_mm_sign_epi8(x1,w1)); + a2=_mm_dpbusd_epi32(a2,_mm_abs_epi8(w2),_mm_sign_epi8(x2,w2)); + a3=_mm_dpbusd_epi32(a3,_mm_abs_epi8(w3),_mm_sign_epi8(x3,w3)); + } + __m128i acc=_mm_add_epi32(_mm_add_epi32(a0,a1),_mm_add_epi32(a2,a3)); + for(;i+32<=I;i+=32){ /* 32-nibble remainder: 2 dpbusd, same unpack */ __m128i by=_mm_loadu_si128((const __m128i*)(w4+(i>>1))); __m128i lo=_mm_and_si128(by,m4), hi=_mm_and_si128(_mm_srli_epi16(by,4),m4); - __m128i n0=_mm_unpacklo_epi8(lo,hi), n1=_mm_unpackhi_epi8(lo,hi); - __m128i w0=_mm_sub_epi8(n0,b8), w1=_mm_sub_epi8(n1,b8); + __m128i w0=_mm_sub_epi8(_mm_unpacklo_epi8(lo,hi),b8), w1=_mm_sub_epi8(_mm_unpackhi_epi8(lo,hi),b8); __m128i x0=_mm_loadu_si128((const __m128i*)(x+i)); __m128i x1=_mm_loadu_si128((const __m128i*)(x+i+16)); acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(w0),_mm_sign_epi8(x0,w0)); @@ -655,6 +751,34 @@ static void pack_int4(const float *w, uint8_t *q4, float *scale, int O, int I, i } } } +/* quantize w[O,I] f32 -> int3-g64 (fmt=5): per 64-input group, symmetric absmax + * (qmax=3, clamp [-4,3], stored v+4), 16B low plane + 8B high plane, ONE f32 scale + * per group. Same math as tools/quant_ablation.py `_quant_last_dim(bits=3, group=64)` + * (#132), here with real bit packing. */ +static void pack_int3_g64(const float *w, uint8_t *q3, float *scale, int O, int I){ + int64_t ng=i3_groups(I), rb=i3_rowbytes(I); + #pragma omp parallel for schedule(static) + for(int o=0;oamax)amax=a; } + float s=amax/3.f; if(s<1e-8f)s=1e-8f; sr[g]=s; + uint8_t *lo=qr+g*I3_GBYTES, *hi=lo+16; + memset(lo,0,I3_GBYTES); + for(int k=0;k3)v=3; if(v<-4)v=-4; + unsigned u=(unsigned)(v+4); /* 0..7 */ + lo[k>>2] |= (uint8_t)((u&3)<<((k&3)*2)); + hi[k>>3] |= (uint8_t)(((u>>2)&1)<<(k&7)); + } + } + } +} + static void pack_int2(const float *w, uint8_t *q2, float *scale, int O, int I, int bits){ int qmax=(1<<(bits-1))-1, rb=(I+3)/4; #pragma omp parallel for schedule(static) diff --git a/c/resource_plan.py b/c/resource_plan.py index 5f17a6a..1374a3b 100644 --- a/c/resource_plan.py +++ b/c/resource_plan.py @@ -166,14 +166,33 @@ def discover_gpus(): return devices +def _physical_cores_warn(message): + """Visibility for a mis-detected core count: a silent "1" here becomes + OMP_NUM_THREADS=1 and pins the whole run to a single core (#325). Emit on + stderr so it surfaces in the [PLAN]/[OMP] stream without being swallowed.""" + print(f"[plan] warning: {message}", file=sys.stderr) + + def physical_cpu_count(): + """Number of physical CPU cores (not SMT siblings). + + Per-expert matmul regions are tiny and back-to-back; two SMT siblings share + one AVX-512 unit and contend, so logical (SMT) counts over-subscribe and + hurt throughput. We want true physical cores. A silent 1 here propagates to + OMP_NUM_THREADS=1 and pins the run to one core (#325), so every fallback + must be visible, never just ``or 1``. + """ if sys.platform == "win32": - # os.cpu_count() conta i processori logici (SMT): 2 thread/core saturano - # le unita' AVX-512 e peggiorano il matmul. Contiamo i core fisici veri - # con GetLogicalProcessorInformationEx(RelationProcessorCore). + # Contiamo i core fisici veri con GetLogicalProcessorInformationEx + # (RelationProcessorCore). Le firme vanno dichiarate: su Python a 64 bit + # una WinAPI non dichiarata ritorna c_int (32 bit) e riceve i puntatori + # come c_int di default, quindi il probe puo' fallire silenziosamente. try: import ctypes k32 = ctypes.windll.kernel32 + k32.GetLogicalProcessorInformationEx.argtypes = [ + ctypes.c_uint, ctypes.c_void_p, ctypes.POINTER(ctypes.c_ulong)] + k32.GetLogicalProcessorInformationEx.restype = ctypes.c_int need = ctypes.c_ulong(0) k32.GetLogicalProcessorInformationEx(0, None, ctypes.byref(need)) buf = (ctypes.c_char * need.value)() @@ -189,18 +208,69 @@ def physical_cpu_count(): off += size if cores: return cores - except (OSError, ValueError, AttributeError): - pass + _physical_cores_warn("GetLogicalProcessorInformationEx returned no cores") + except (OSError, ValueError, AttributeError) as error: + _physical_cores_warn(f"Windows core probe failed: {error}") try: + # Ask lscpu for exactly core,socket and dedupe on (core, socket). + # Counting un-deduplicated rows would return logical threads (SMT), + # which was the original over-subscription bug. Empty fields ("-") + # mark an offline core/socket and fail int() -> skipped. + # + # Column layout robustness: `lscpu -p=` emits *exactly* the + # requested columns (no CPU prefix), while bare `lscpu -p` prepends + # CPU. We requested two columns, but take the LAST TWO fields so the + # parser stays correct whether or not a CPU column is present + # (JustVugg review: the previous fields[1]/fields[2] indexing assumed + # a 3-column layout and regressed 2-column output to the logical + # count -- the opposite of the fix). result = subprocess.run(["lscpu", "-p=core,socket"], text=True, capture_output=True, check=True, timeout=5) - cores = {tuple(map(int, line.split(","))) for line in result.stdout.splitlines() - if line and not line.startswith("#")} + cores = set() + for line in result.stdout.splitlines(): + if not line or line.startswith("#"): + continue + fields = line.split(",") + if len(fields) < 2: + continue + try: + core, socket = int(fields[-2]), int(fields[-1]) + except ValueError: + continue # "-" for an offline core/socket + cores.add((core, socket)) if cores: return len(cores) - except (OSError, ValueError, subprocess.SubprocessError): - pass - return os.cpu_count() or 1 + except (OSError, ValueError, subprocess.SubprocessError) as error: + _physical_cores_warn(f"lscpu core probe failed: {error}") + logical = os.cpu_count() + if not logical: + _physical_cores_warn( + "could not detect any CPU cores; falling back to 1. " + "Set OMP_NUM_THREADS manually to fix single-core decode (#325).") + return 1 + _physical_cores_warn( + f"physical-core probes unavailable; using {logical} logical CPUs " + f"(SMT may over-subscribe). Set OMP_NUM_THREADS to physical cores if slow.") + return logical + + +def _resolve_physical_cores(physical_cpus): + """Coerce the build_plan() physical-core argument to a sane positive int. + + A None/0/None-ish value reaching here means physical_cpu_count() already + warned; clamp to 1 (so the engine always gets a positive team size) but keep + that clamp visible rather than silently masking it as the old ``max(1, int())`` + did (#325).""" + try: + count = int(physical_cpus or 0) + except (TypeError, ValueError): + count = 0 + if count < 1: + _physical_cores_warn( + "physical core count resolved to 0; defaulting to 1. " + "Set OMP_NUM_THREADS to fix single-core decode (#325).") + return 1 + return count def cpu_socket_count(): @@ -219,6 +289,54 @@ def cpu_socket_count(): return 1 +def _auto_tune(bottleneck_class, projected_hit, gpus, cpu_sockets, plan_has_metal): + """Derive tuning knobs from the bottleneck classification.""" + tune = {} + has_gpu = bool(gpus) + n_gpu = len(gpus) + + # MTP: costs more than it saves when compute-bound (#389 measured 42% loss) + if bottleneck_class == "compute": + tune["DRAFT"] = {"value": "0", + "reason": "compute-bound: MTP batch overhead exceeds yield"} + elif bottleneck_class == "disk" and projected_hit < 0.90: + tune["DRAFT"] = {"value": "0", + "reason": "low hit rate: MTP widens expert union, adds disk reads"} + # otherwise leave DRAFT unset (engine default: auto) + + # PIPE: resident pipeline mode depends on GPU count + if has_gpu and n_gpu == 1: + tune["COLI_CUDA_PIPE"] = {"value": "1", + "reason": "single GPU: S=1 pipeline gate"} + elif has_gpu and n_gpu > 1: + tune["COLI_CUDA_PIPE"] = {"value": "2", + "reason": "multi-GPU: residual stays on-device across layers"} + elif not has_gpu and bottleneck_class == "disk": + tune["PIPE"] = {"value": "1", + "reason": "overlap disk reads with resident expert compute"} + + # NUMA: selective interleave for GPU hosts, blanket hint for CPU-only + if cpu_sockets > 1 and has_gpu: + tune["COLI_NUMA"] = {"value": "1", + "reason": "multi-socket + GPU: interleave expert slabs, protect DMA buffers"} + elif cpu_sockets > 1 and not has_gpu: + tune["COLI_NUMA"] = {"value": "1", + "reason": "multi-socket CPU-only: interleave expert slabs across nodes"} + tune["_numa_hint"] = "numactl --interleave=all may perform better on CPU-only hosts" + + # OMP: kill hot-thread spin when GPU/Metal owns the power budget + if plan_has_metal: + tune["COLI_NO_OMP_TUNE"] = {"value": "1", + "reason": "Metal: OMP spin-wait steals GPU power budget"} + + # PIN: fully resident if RAM allows and no GPU tier competes + if projected_hit >= 0.99 and not has_gpu: + tune["PIN_GB"] = {"value": "all", + "reason": "enough RAM for full expert residency"} + + return tune + + POLICIES = { "quality": {"preserve_quantization": True, "preserve_router": True}, "balanced": {"preserve_quantization": True, "preserve_router": True}, @@ -290,19 +408,35 @@ def build_plan(model, ram_gb=0, context=4096, gpu_indices=None, vram_gb=0, if cold_bytes: warnings.append("cold expert misses may reach disk; normal decode speed depends on hit rate") + total_expert = info["expert_bytes"] + resident_expert = hot_bytes + warm_bytes + projected_hit = resident_expert / total_expert if total_expert else 1.0 + if cold_bytes: bottleneck = "disk expert misses" - elif warm_bytes: - bottleneck = "CPU expert compute and RAM bandwidth" + bottleneck_class = "disk" + elif warm_bytes and gpus: + bottleneck = "CPU expert tail and GPU compute" + bottleneck_class = "mixed" + elif projected_hit >= 0.99: + if gpus: + bottleneck = "GPU compute and interconnect" + else: + bottleneck = "CPU expert compute (fully resident)" + bottleneck_class = "compute" else: - bottleneck = "GPU compute and interconnect" + bottleneck = "CPU expert compute and RAM bandwidth" + bottleneck_class = "memory" + + tune = _auto_tune(bottleneck_class, projected_hit, gpus, cpu_sockets, + plan_has_metal=False) return { "version": 2, "policy": {"name": policy, **POLICIES[policy], "quality_preserving": policy != "experimental-fast"}, "model": {key: value for key, value in info.items() if key != "config"}, - "cpu": {"physical_cores": max(1, int(physical_cpus)), + "cpu": {"physical_cores": _resolve_physical_cores(physical_cpus), "sockets": max(1, int(cpu_sockets)), "thread_policy": "physical-cores"}, "tiers": { @@ -317,6 +451,9 @@ def build_plan(model, ram_gb=0, context=4096, gpu_indices=None, vram_gb=0, "expert_capacity": vram_experts, "requires_host_backing": False}, }, "expected_bottleneck": bottleneck, + "bottleneck_class": bottleneck_class, + "projected_hit_rate": round(projected_hit, 4), + "tune": tune, "decisions": [ {"target": "VRAM", "reason": "profile-ranked hot experts"}, {"target": "RAM", "reason": "warm experts execute on CPU without quality loss"}, @@ -331,15 +468,23 @@ def environment_for_plan(plan, env=None, cuda_enabled=True): result = dict(env or {}) result.setdefault("COLI_POLICY", plan["policy"]["name"]) result.setdefault("OMP_NUM_THREADS", str(plan["cpu"]["physical_cores"])) - if sys.platform != "win32": - # la libgomp di MinGW non supporta l'affinity su Windows - # ("Affinity not supported on this configuration"): non impostarle li'. - result.setdefault("OMP_PROC_BIND", "spread") - result.setdefault("OMP_PLACES", "cores") - if sys.platform.startswith("linux") and plan["cpu"].get("sockets", 1) > 1: - # Selectively interleave large expert/dense slabs across memory controllers. - # Unlike blanket numactl interleave, this leaves CUDA staging buffers local. - result.setdefault("COLI_NUMA", "1") + # NOTE: we intentionally do NOT set OMP_PROC_BIND / OMP_PLACES here. + # The engine's own hot-thread tuning (glm.c main(), the COLI_OMP_TUNED + # self-exec) sets OMP_PROC_BIND=close with overwrite=0 -- it prefers + # packing the team onto adjacent cores for the tiny back-to-back per-expert + # matmuls. Pre-setting OMP_PROC_BIND=spread here ran first and won (the + # engine's overwrite=0 setenv could not override an already-set var), and + # spread + OMP_PLACES=cores collapsed the team to one CPU on some libgomp / + # multi-socket topologies (#325: --auto-tier pinned decode to 1 core on a + # 64-core box even with OMP_NUM_THREADS=64). Leaving affinity to the engine + # makes --auto-tier match the plain (working) path. A user who wants a + # specific policy can still set OMP_PROC_BIND/OMP_PLACES in the environment + # themselves -- setdefault above only covers OMP_NUM_THREADS. + tune = plan.get("tune", {}) + for key, entry in tune.items(): + if key.startswith("_"): + continue + result.setdefault(key, entry["value"]) if plan["policy"]["name"] == "balanced": result.setdefault("REPIN", "64") ram = plan["tiers"]["ram"] @@ -386,5 +531,18 @@ def format_plan(plan): else: lines.append("VRAM no NVIDIA device detected · CPU path") lines.append(f"limit {plan['expected_bottleneck']}") + hit = plan.get("projected_hit_rate", 0) + lines.append(f"hit {hit:.0%} projected expert residency") + tune = plan.get("tune", {}) + if tune: + lines.append("") + lines.append("auto-tune:") + for key, entry in tune.items(): + if key.startswith("_"): + continue + lines.append(f" {key}={entry['value']:12s} {entry['reason']}") + hint = tune.get("_numa_hint") + if hint: + lines.append(f" hint: {hint}") lines.extend(f"warn {warning}" for warning in plan["warnings"]) return "\n".join(lines) diff --git a/c/sample.h b/c/sample.h index 26db529..414c621 100644 --- a/c/sample.h +++ b/c/sample.h @@ -116,6 +116,16 @@ static void stops_arm_tok(const Cfg *c, int tok_eos, Tok *T){ int nsp = 0; if (T) for (int id = 0; id < T->n_ids && g_nstop < 64; id++) if (T->id_special[id] && !is_stop(id)) { g_stop[g_nstop++] = id; nsp++; } + /* #401: in serve mode keep ONLY <|endoftext|>. Role markers <|user|>/<|observation|> + * (config stops + tokenizer special set) are boundaries the Python server owns; as + * hard stops they cut generation the moment the model opens a block, + * because int4 argmax noise picks a stop-token ID over the correct '<' token. */ + if (getenv("SERVE") && tok_eos >= 0) { + int kept = 0; + for (int i = 0; i < g_nstop; i++) if (g_stop[i] == tok_eos) g_stop[kept++] = g_stop[i]; + if (kept < g_nstop) fprintf(stderr, "[stop] serve mode: filtered %d non-EOS stop tokens (tool-call safety, #401)\n", g_nstop - kept); + g_nstop = kept; nsp = 0; + } fprintf(stderr, "[stop] %d stop tokens:", g_nstop); for (int i = 0; i < g_nstop; i++) fprintf(stderr, " %d", g_stop[i]); if (nsp) fprintf(stderr, " (%d from the tokenizer's special set)", nsp); diff --git a/c/tests/README_efficiency.md b/c/tests/README_efficiency.md new file mode 100644 index 0000000..1f8d339 --- /dev/null +++ b/c/tests/README_efficiency.md @@ -0,0 +1,98 @@ +# Efficiency suite — regression tests + optimization dossier + +Two layers: + +1. **`test_inefficiency.py`** — tiny-model *asserted* regression tests. Fast + (~0.15s/run), gate CI, catch breakage. Run as part of `make test`. +2. **`test_efficiency_report.py`** — an *opt-in optimization dossier* for a real + model. Runs every instrumentation flag, prints a 9-section report answering + *what is doing what, when, with what, is it inefficient, how to improve*. + Never fails CI (it's a report, not a gate). + +## The dossier (what you run when optimizing) + +```bash +# CPU-only (safe, fast to validate): +COLI_EFFICIENCY_MODEL=../glm52_i4_g64 make efficiency-report + +# CUDA (dense + expert tiers — needs a CUDA build, see below): +COLI_EFFICIENCY_MODEL=../glm52_i4_g64 COLI_EFFICIENCY_CUDA=1 make efficiency-report +``` + +It turns ON every observability flag the engine supports — `PROF=1`, +`COLI_CUDA_PROFILE=1`, `CACHE_ROUTE=1` (auto-unlocks `route_agree`/`route_kl`), +`DISK_SPLIT=1`, `LOOKA=1` — so nothing the engine can tell you is left dark. +None of these change the computed output; they only add telemetry. + +The 9 sections, and the question each answers: + +| § | section | answers | +|---|---|---| +| 1 | PROVENANCE | what is running, on what CPU/backend, with what effective config | +| 2 | THROUGHPUT | tok/s + forward-latency p50/p90/p99/max (is the tail healthy?) | +| 3 | WHERE TIME GOES | the 5 PROFILE phases as % of decode + absolute seconds + verdict | +| 3a | ATTENTION BREAKDOWN | attention split into projection/RoPE, score-softmax-value, output | +| 4 | EXPERT CACHE | hit %, experts-loaded/token vs baseline topk | +| 5 | DISK I/O | GB fetched, MB/token, GB/s, read-service vs felt-wait, phase split | +| 5a | DISK-LOAD SPLIT | loads by decode phase (draft/absorb/verify) + MTP-vs-main bytes | +| 6 | ROUTING QUALITY | route_agree %, route_kl, cache swaps | +| 6a | ROUTING PREDICTABILITY | LOOKAHEAD recall per predictor (which prefetch wins) | +| 7 | SPECULATION | tokens/forward, MTP acceptance % | +| 8 | GPU TIERS | resident tensors, expert tier (count/GB/calls), H2D/kernel/D2H ms | + +Every line that crosses an advisory threshold is marked `[FLAG]` with the +concrete lever to pull (raise RAM_GB, add PIN_GB, try DIRECT=1, lower CTX, …), +and all flags repeat in a summary at the end. + +## Tunable thresholds + +The `IS IT INEFFICIENT?` lines are advisory constants at the top of +`test_efficiency_report.py`: + +| constant | default | meaning | +|---|---|---| +| `DISK_WAIT_DOMINANT` | 0.40 | >40% decode waiting on expert reads → I/O-bound | +| `LOW_HIT_RATE` | 0.30 | <30% cache hit → thrashing | +| `LOW_ROUTE_AGREE` | 0.80 | <80% routing overlap → prefetch guessing wrong | +| `HIGH_TAIL_RATIO` | 3.0 | p99 > 3× p50 → decode stalls | +| `LOW_MTP_ACCEPT` | 0.20 | <20% MTP acceptance → draft decoder is dead weight | + +The tiny-model asserted floors live in `tools/efficiency.py` (`TINY_TOK_S_FLOOR`, +`MAX_DISK_WAIT_SHARE`, `MIN_CPU_CUDA_AGREEMENT`). + +## The regression tests (what gates CI) + +`test_inefficiency.py` runs on the bundled `glm_tiny` model and asserts: + +- telemetry parses (no format drift) +- tiny tok/s ≥ floor (throughput regression) +- PROFILE phases present and non-negative (accounting sanity) +- disk-wait not dominant on a resident model (I/O-path regression) +- CPU determinism (two greedy runs agree) +- **CUDA** (skip unless CUDA built): init path, dense uploads VRAM, CPU-vs-CUDA + argmax agreement ≥ 70% (kernel-correctness guard) + +```bash +make efficiency # tiny CPU tests +make efficiency-cuda # tiny CUDA tests (needs CUDA build) +``` + +## CUDA build prerequisite + +The default `make glm.exe` builds **without** CUDA. The CUDA tests and the CUDA +dossier need a host built with `-DCOLI_CUDA` plus the runtime DLL: + +```bash +make clean && make glm.exe CUDA_DLL=1 && make cuda-dll +``` + +`make efficiency-cuda` auto-skips with a clear message if the host is CPU-only +(it scans the binary for the "CPU-only" marker the engine embeds). + +## Files + +- `tools/efficiency.py` — shared harness: `parse_run()` (captures every + telemetry signal), `run_engine()`, thresholds. Reuses `PROFILE_RE`/`SPEED_RE` + from `tools/benchmark_cuda_fixture.py`. +- `tests/test_inefficiency.py` — tiny-model asserted tests (CPU + CUDA). +- `tests/test_efficiency_report.py` — the opt-in optimization dossier. diff --git a/c/tests/bench_idot.c b/c/tests/bench_idot.c new file mode 100644 index 0000000..37b0b30 --- /dev/null +++ b/c/tests/bench_idot.c @@ -0,0 +1,92 @@ +/* Microbenchmark: old (single-accumulator) vs new (independent-accumulator) AVX-VNNI + * int8/int4 dot kernels (quant.h). NOT a unit test -- test_idot.c proves correctness. + * This measures the headline claim: breaking the serial vpdpbusd->acc chain lifts + * per-core kernel throughput, the same win the NEON path already took ("26->63 GB/s, 2.4x"). + * + * It re-implements the OLD single-acc AVX-VNNI kernels inline and calls the REAL (new) + * ones via the include-colibri.c pattern -- one process, identical frozen inputs, warm + * caches. Reports median ns/call + GB/s-of-weights + new/old ratio. Measures the kernel's + * compute ceiling (warm caches), NOT end-to-end tok/s. + * + * Run: make tests/bench_idot ARCH=native && ./tests/bench_idot (not in TEST_BINS -- not a gate) + */ +#define main coli_glm_main_unused +#include "../colibri.c" +#undef main +#include +#include + +static uint32_t rs=0x2545F491u; +static uint32_t xr(void){ rs^=rs<<13; rs^=rs>>17; rs^=rs<<5; return rs; } + +/* ---- OLD kernels: verbatim copies of the pre-change __AVXVNNI__ branches (single acc) ---- */ +#if defined(__AVXVNNI__) && defined(__AVX2__) +static int32_t dot_i8i8_old(const int8_t *w, const int8_t *x, int I){ + int32_t sum=0; int i=0; + __m128i acc=_mm_setzero_si128(); + for(;i+16<=I;i+=16){ + __m128i wv=_mm_loadu_si128((const __m128i*)(w+i)); + __m128i xv=_mm_loadu_si128((const __m128i*)(x+i)); + __m128i xs=_mm_sign_epi8(xv,wv); + acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(wv),xs); + } + sum=hsum128_i32(acc); + for(;i>1))); + __m128i lo=_mm_and_si128(by,m4), hi=_mm_and_si128(_mm_srli_epi16(by,4),m4); + __m128i n0=_mm_unpacklo_epi8(lo,hi), n1=_mm_unpackhi_epi8(lo,hi); + __m128i w0=_mm_sub_epi8(n0,b8), w1=_mm_sub_epi8(n1,b8); + __m128i x0=_mm_loadu_si128((const __m128i*)(x+i)); + __m128i x1=_mm_loadu_si128((const __m128i*)(x+i+16)); + acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(w0),_mm_sign_epi8(x0,w0)); + acc=_mm_dpbusd_epi32(acc,_mm_abs_epi8(w1),_mm_sign_epi8(x1,w1)); + } + sum=hsum128_i32(acc); + for(;i>1]; int v=(i&1)?((int)(b>>4)-8):((int)(b&0xF)-8); sum+=v*x[i]; } + return sum; +} +#else +#error "bench_idot requires an AVX-VNNI build: make tests/bench_idot ARCH=native on an AVX-VNNI CPU" +#endif + +#define I_DIM 6144 +#define N_REPEAT 20000 +static int cmp_d(const void*a,const void*b){ double x=*(const double*)a,y=*(const double*)b; return xy?1:0; } + +int main(void){ + static int8_t w8[I_DIM], x8[I_DIM]; static uint8_t w4[I_DIM/2]; + for(int i=0;i 1 && coli_cuda_tensor_upload(&t8, q8, s8, 1, 4, 2, d1)) return 1; if (!coli_cuda_matmul(&t8, got, x, q8, s8, 1, 2, 4, 2, d0) || !close_enough(got, want8, 4)) return 1; + /* Cached tensor must stay callable without live host pointers + * (CUDA_RELEASE_HOST slots null theirs after upload) — including + * SUSTAINED reuse, not just the first call. */ + for (int rep = 0; rep < 64; rep++) + if (!coli_cuda_matmul(&t8, got, x, nullptr, nullptr, 1, 2, 4, 2, d0) || + !close_enough(got, want8, 4)) return 1; + /* A tensor uploaded from a TEMPORARY host buffer must survive the buffer + * being scribbled and freed (the release-host lifecycle). */ + { + int8_t *tmpw = static_cast(std::malloc(8)); + float *tmps = static_cast(std::malloc(2 * sizeof(float))); + if (!tmpw || !tmps) return 2; + for (int i = 0; i < 8; i++) tmpw[i] = q8[i]; + tmps[0] = s8[0]; tmps[1] = s8[1]; + ColiCudaTensor *tt = nullptr; + if (!coli_cuda_tensor_upload(&tt, tmpw, tmps, 1, 4, 2, d0)) return 1; + for (int i = 0; i < 8; i++) tmpw[i] = 99; + std::free(tmpw); std::free(tmps); + if (!coli_cuda_matmul(&tt, got, x, nullptr, nullptr, 1, 2, 4, 2, d0) || + !close_enough(got, want8, 4)) return 1; + coli_cuda_tensor_free(tt); + } + /* Upload failures must be graceful and must not corrupt accounting — + * and must not poison LATER healthy launches (sticky-error regression). */ + { + size_t c0 = 0, b0 = 0, c1 = 0, b1 = 0; + coli_cuda_stats(-1, &c0, &b0); + ColiCudaTensor *bad = nullptr; + if (coli_cuda_tensor_upload(&bad, q8, s8, 1, 4, 2, 9999)) return 1; + if (coli_cuda_tensor_upload(&bad, q8, s8, 7, 4, 2, d0)) return 1; + if (coli_cuda_tensor_upload(&bad, q8, nullptr, 1, 4, 2, d0)) return 1; + if (coli_cuda_tensor_upload(&bad, nullptr, s8, 1, 4, 2, d0)) return 1; + if (coli_cuda_tensor_upload(&bad, q8, s8, 1, 1 << 20, 1 << 24, d0)) return 1; /* ~16 TB */ + if (bad) return 1; + coli_cuda_stats(-1, &c1, &b1); + if (c0 != c1 || b0 != b1) return 1; + /* healthy launch immediately after the failed allocation */ + if (!coli_cuda_matmul(&t8, got, x, nullptr, nullptr, 1, 2, 4, 2, d0) || + !close_enough(got, want8, 4)) return 1; + } + /* Fault injection hook: on/off, restores cleanly. */ + if (setenv("COLI_GPU_FAIL_AFTER", "0", 1)) return 2; + if (coli_cuda_matmul(&t8, got, x, nullptr, nullptr, 1, 2, 4, 2, d0)) return 1; + if (unsetenv("COLI_GPU_FAIL_AFTER")) return 2; + if (!coli_cuda_matmul(&t8, got, x, nullptr, nullptr, 1, 2, 4, 2, d0) || + !close_enough(got, want8, 4)) return 1; const int8_t q8b[8]={-1,-2,-3,-4, 1,-2,3,-4}; const float s8b[2]={1.f,.5f},want8b[4]={10.f,15.f,-3.f,-2.5f}; if(!coli_cuda_tensor_update(t8,q8b,s8b)|| - !coli_cuda_matmul(&t8,got,x,q8b,s8b,1,2,4,2,d0)|| + !coli_cuda_matmul(&t8,got,x,q8b,s8b,1,2,4,2,d0,0)|| !close_enough(got,want8b,4))return 1; /* Rows [-8,-1,0,7] and [1,2,3,4], packed low nibble first. */ @@ -61,26 +107,26 @@ int main(int argc, char **argv) { const float s4[2] = {1.0f, 0.25f}; const float want4[2] = {-34.0f, -2.5f}; ColiCudaTensor *t4 = nullptr; - if (!coli_cuda_matmul(&t4, got, x, q4, s4, 2, 1, 4, 2, d1) || !close_enough(got, want4, 2)) return 1; + if (!coli_cuda_matmul(&t4, got, x, q4, s4, 2, 1, 4, 2, d1, 0) || !close_enough(got, want4, 2)) return 1; const uint8_t q2[2] = {0xe4, 0x1b}; const float s2[2] = {0.5f, 2.0f}; const float want2[2] = {-2.0f, 12.0f}; ColiCudaTensor *t2 = nullptr; - if (!coli_cuda_matmul(&t2, got, x, q2, s2, 3, 1, 4, 2, d1) || !close_enough(got, want2, 2)) return 1; + if (!coli_cuda_matmul(&t2, got, x, q2, s2, 3, 1, 4, 2, d1, 0) || !close_enough(got, want2, 2)) return 1; const float wf[8] = {1, 0, -1, 2, 0.5f, 0.5f, 0.5f, 0.5f}; const float wantf[2] = {-10.0f, -1.0f}; ColiCudaTensor *tf = nullptr; - if (!coli_cuda_matmul(&tf, got, x, wf, nullptr, 0, 1, 4, 2, d0) || !close_enough(got, wantf, 2)) return 1; + if (!coli_cuda_matmul(&tf, got, x, wf, nullptr, 0, 1, 4, 2, d0, 0) || !close_enough(got, wantf, 2)) return 1; const float eg[8] = {1,0,0,0, 0,1,0,0}; const float eu[8] = {1,0,0,0, 0,1,0,0}; const float ed[8] = {1,0, 0,1, 1,1, 1,-1}; ColiCudaTensor *tg=nullptr,*tu=nullptr,*td=nullptr; - if (!coli_cuda_tensor_upload(&tg,eg,nullptr,0,4,2,d0) || - !coli_cuda_tensor_upload(&tu,eu,nullptr,0,4,2,d0) || - !coli_cuda_tensor_upload(&td,ed,nullptr,0,2,4,d0)) return 1; + if (!coli_cuda_tensor_upload(&tg,eg,nullptr,0,4,2,d0,0) || + !coli_cuda_tensor_upload(&tu,eu,nullptr,0,4,2,d0,0) || + !coli_cuda_tensor_upload(&td,ed,nullptr,0,2,4,d0,0)) return 1; float expert[8], want_expert[8]; for(int s=0;s<2;s++){ float a=x[s*4], b=x[s*4+1]; @@ -98,7 +144,7 @@ int main(int argc, char **argv) { const float aw[16]={1,0,0,0, 0,1,0,0, 0,0,1,0, 0,0,0,1}; const float aq[4]={1,2,.5f,-.5f},al[12]={1,0,0,0, 0,1,0,0, 0,0,1,0}; const float ar[6]={1,0, 0,1, 1,1};float actx[2],aref[2]; - ColiCudaTensor *at=nullptr;if(!coli_cuda_tensor_upload(&at,aw,nullptr,0,4,4,d0))return 1; + ColiCudaTensor *at=nullptr;if(!coli_cuda_tensor_upload(&at,aw,nullptr,0,4,4,d0,0))return 1; float score[3];for(int t=0;t<3;t++)score[t]=aq[0]*al[t*4]+aq[1]*al[t*4+1]+aq[2]*ar[t*2]+aq[3]*ar[t*2+1]; float mx=score[0],z=0;for(int t=1;t<3;t++)mx=score[t]>mx?score[t]:mx; for(int t=0;t<3;t++){score[t]=std::exp(score[t]-mx);z+=score[t];}for(int t=0;t<3;t++)score[t]/=z; @@ -117,9 +163,9 @@ int main(int argc, char **argv) { for(int i=0;i<32;i++)ws4[i]=0.01f+(i%5)*0.002f; for(int i=0;i<64;i++)gx4[i]=std::sin((float)(i+1)*0.17f)*2.f; ColiCudaTensor *g4=nullptr,*u4=nullptr,*d4=nullptr; - if(!coli_cuda_tensor_upload(&g4,w4,ws4,2,32,32,d0)|| - !coli_cuda_tensor_upload(&u4,w4,ws4,2,32,32,d0)|| - !coli_cuda_tensor_upload(&d4,w4,ws4,2,32,32,d0))return 1; + if(!coli_cuda_tensor_upload(&g4,w4,ws4,2,32,32,d0,0)|| + !coli_cuda_tensor_upload(&u4,w4,ws4,2,32,32,d0,0)|| + !coli_cuda_tensor_upload(&d4,w4,ws4,2,32,32,d0,0))return 1; ColiCudaTensor *gg4[2]={g4,g4},*ug4[2]={u4,u4},*dg4[2]={d4,d4}; if(!coli_cuda_expert_group(gg4,ug4,dg4,group_rows,2,scalar4,gx4))return 1; setenv("COLI_CUDA_TC_INT4","1",1); diff --git a/c/tests/test_backend_metal.mm b/c/tests/test_backend_metal.mm index ffdd39b..bff0444 100644 --- a/c/tests/test_backend_metal.mm +++ b/c/tests/test_backend_metal.mm @@ -177,6 +177,68 @@ static int run_attn(int S, int pos_base, const char* name){ return pass?0:1; } +// serial r_top8 vs parallel r_top8_par on the ENGINE build's own compiled shaders — the +// exact-match contract (same indices, same order, same weights bitwise, same keff) +// enforced with memcmp, per adversarial input family. `mode` selects the input +// construction; see the inventory at the call sites in main(). E is a parameter (not +// hardcoded 256) so the same helper drives both the original E=256 fuzz and the +// expert-count-generality cases (E=24 <32-lane-width, E=168 REAP-pruned, E=200 +// lane-straddling boundary, E=257 out-of-contract auto-serial-fallback proof). +static int run_rtop8(int mode, int S, int E, float topp, int normk, float rscale, const char *name) { + const int K=8, Ksel=8; + std::vector sig((size_t)S*E), bias(E); + srand(4242+mode*17+S+E); + for (int e=0;e=E-4)?1.0f:(float)(rand()%10000)/10000.f; break; + default: *v=(float)(rand()%10000)/10000.f; break; + } + } + if (mode==1) for (int e=0;e is((size_t)S*K), ip((size_t)S*K); std::vector ws((size_t)S*K), wp((size_t)S*K); + std::vector ks(S), kp(S); + if (!coli_metal_rtop8(0,sig.data(),bias.data(),S,E,K,Ksel,topp,normk,rscale,is.data(),ws.data(),ks.data()) || + !coli_metal_rtop8(1,sig.data(),bias.data(),S,E,K,Ksel,topp,normk,rscale,ip.data(),wp.data(),kp.data())) { + printf(" %-34s FAIL (rtop8 runner returned 0)\n", name); return 1; } + int ok = memcmp(is.data(),ip.data(),(size_t)S*K*4)==0 && + memcmp(ws.data(),wp.data(),(size_t)S*K*4)==0 && // bitwise: same ops, same order + memcmp(ks.data(),kp.data(),(size_t)S*4)==0; + if (mode==5 && ok) { + // Don't just trust the input design -- confirm the straddling lane's valid segment + // (E-4..E-1) was actually selected, in EVERY row, so this case can't silently + // degrade into an unrelated pass if the input construction above ever changes. + for (int s=0;s=E-4 && ip[(size_t)s*K+k]256, auto-serial-fallback)"); printf(fail? "metal backend tests: FAILED\n" : "metal backend tests: ok\n"); coli_metal_shutdown(); return fail; diff --git a/c/tests/test_decode_batch.c b/c/tests/test_decode_batch.c index d370d2d..f5b216e 100644 --- a/c/tests/test_decode_batch.c +++ b/c/tests/test_decode_batch.c @@ -44,6 +44,12 @@ static void test_submit_header(void) assert(coli_submit_parse("SUBMIT 1 0 16777216 3 1 1", &sub)); assert(!coli_submit_parse("SUBMIT 1 0 16777217 3 1 1", &sub)); assert(!coli_submit_parse("SUBMIT 1 0 2 3 1 1 trailing", &sub)); + /* optional 7th field: per-request grammar length (0 when absent) */ + assert(coli_submit_parse("SUBMIT 42 3 17 64 0.7 0.95", &sub) && sub.gbytes == 0); + assert(coli_submit_parse("SUBMIT 42 3 17 64 0.7 0.95 512", &sub) && sub.gbytes == 512); + assert(coli_submit_parse("SUBMIT 42 3 17 64 0.7 0.95 1048576", &sub)); + assert(!coli_submit_parse("SUBMIT 42 3 17 64 0.7 0.95 1048577", &sub)); + assert(!coli_submit_parse("SUBMIT 42 3 17 64 0.7 0.95 512 extra", &sub)); } int main(void) diff --git a/c/tests/test_efficiency_report.py b/c/tests/test_efficiency_report.py new file mode 100644 index 0000000..ba01b13 --- /dev/null +++ b/c/tests/test_efficiency_report.py @@ -0,0 +1,315 @@ +#!/usr/bin/env python3 +"""Exhaustive optimization dossier for a colibri engine run. + +This is NOT a pass/fail test. It runs the engine with every instrumentation flag +on (PROF, COLI_CUDA_PROFILE, CACHE_ROUTE, DISK_SPLIT, LOOKA) and prints a section- +by-section report answering, for each subsystem: + + WHAT is doing it — which phase/kernel/tier + WHEN it is doing it — how much of decode wall-time it owns + WITH WHAT — the config/weights/tier it used + IS IT INEFFICIENT? — a verdict, with the threshold + HOW TO IMPROVE — the concrete knob, named + +Activation (opt-in only — NOT in `make test`): + COLI_EFFICIENCY_MODEL= python tests/test_efficiency_report.py + +Optional env: + COLI_EFFICIENCY_CUDA=1 also exercise the CUDA dense/expert tiers + COLI_EFFICIENCY_NGEN=N decode tokens (default 24) + COLI_EFFICIENCY_RAM_GB=N RAM budget (default 28) + COLI_EFFICIENCY_VRAM_GB=N CUDA expert-tier budget GB (default 4) + COLI_EFFICIENCY_PROMPT=... prompt (default: a code-gen prompt) + +Exit code is always 0 (it's a dossier, not a gate). Lines marked FLAG point at +the most likely lever to move tok/s for the observed bottleneck. +""" +import os +import sys +import time +from pathlib import Path + +sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) +from tools.efficiency import run_engine, disk_wait_share # noqa: E402 + + +# --- advisory thresholds (the "IS IT INEFFICIENT?" lines) --- +DISK_WAIT_DOMINANT = 0.40 # >40% decode waiting on expert reads -> I/O-bound +LOW_HIT_RATE = 0.30 # <30% cache hit -> thrashing (cap too small) +LOW_ROUTE_AGREE = 0.80 # <80% routing overlap -> prefetch is guessing wrong +HIGH_TAIL_RATIO = 3.0 # p99 > 3x p50 -> decode stalls (I/O hiccups / KV grow) +LOW_MTP_ACCEPT = 0.20 # <20% MTP acceptance -> draft decoder is dead weight +VRAM_WASTE_CALLS = 0 # experts pinned in VRAM but 0 calls served + + +def _flag(ok): return "OK " if ok else "FLAG" + + +def _bar(frac, width=24): + """A simple ASCII bar for share visualization.""" + n = max(0, min(width, round(frac * width))) + return "#" * n + "." * (width - n) + + +def _line(label, value, flag=None, note=""): + tag = f" [{flag}]" if flag else "" + print(f" {label:<22} {value}{tag} {note}" if note else f" {label:<22} {value}{tag}") + + +def main() -> int: + model = os.environ.get("COLI_EFFICIENCY_MODEL") + if not model: + print(__doc__) + print("\nNot activated: set COLI_EFFICIENCY_MODEL= to run.") + return 0 + model = str(Path(model).resolve()) + if not Path(model).is_dir(): + print(f"ERROR: {model} is not a directory", file=sys.stderr) + return 0 + + ngen = int(os.environ.get("COLI_EFFICIENCY_NGEN", "24")) + ram_gb = os.environ.get("COLI_EFFICIENCY_RAM_GB", "28") + vram_gb = os.environ.get("COLI_EFFICIENCY_VRAM_GB", "4") + prompt = os.environ.get( + "COLI_EFFICIENCY_PROMPT", + "Write a Python function that computes the factorial of a number. " + "Include error handling and a docstring.") + use_cuda = os.environ.get("COLI_EFFICIENCY_CUDA") == "1" + + # Turn ON every instrumentation flag so the dossier has maximum detail. + # These are all observability toggles (PROF/COLI_CUDA_PROFILE/CACHE_ROUTE/ + # DISK_SPLIT/LOOKA); none change the computed output. + overlay = dict( + NGEN=str(ngen), TEMP="0", RAM_GB=ram_gb, PROMPT=prompt, + PROF="1", CACHE_ROUTE="1", DISK_SPLIT="1", LOOKA="1", ROUTE_AGREE="1", + ) + if use_cuda: + overlay.update(COLI_CUDA="1", COLI_GPU="0", CUDA_DENSE="1", + COLI_CUDA_PROFILE="1", CUDA_EXPERT_GB=vram_gb) + + print("=" * 78) + print(f"OPTIMIZATION DOSSIER — {Path(model).name}") + print(f" mode : {'CUDA (dense+expert tiers)' if use_cuda else 'CPU-only'} " + f"ngen : {ngen} ram : {ram_gb} GB" + + (f" vram : {vram_gb} GB" if use_cuda else "")) + print("=" * 78) + + t0 = time.time() + t, proc = run_engine(overlay, snap=model, timeout=3600.0) + wall = time.time() - t0 + flags = [] # collected FLAG lines for the summary + + print(f"\n[0] RUN") + _line("wall clock", f"{wall:.0f}s") + _line("exit code", proc.returncode, + None if proc.returncode == 0 else "FLAG", + "" if proc.returncode == 0 else "non-zero exit") + if proc.returncode != 0: + print(" stderr tail:") + for ln in proc.stderr.strip().splitlines()[-8:]: + print(f" {ln}") + return 0 + + # ---------------------------------------------------------------- [1] WHO ---- + print(f"\n[1] PROVENANCE — what is running, on what, with what config") + if t.get("machine"): + m = t["machine"] + _line("CPU", m["cpu"]) + _line("cores / omp", f"{m['cores']} cores") + _line("backend", m["backend"]) + if t.get("load"): + ld = t["load"] + _line("model load time", f"{ld['load_s']:.2f}s") + _line("resident dense", f"{ld['resident_dense_mb']:.1f} MB") + _line("layers / experts", f"{ld['layers']} layers, {ld['experts']} experts") + _line("MTP", f"{ld['mtp_status']} (draft={ld['draft']})") + if t.get("config_str"): + _line("resolved config", t["config_str"]) + print(" (this is the EFFECTIVE config after auto-budgeting — not your env verbatim)") + + # ---------------------------------------------------------------- [2] SPEED -- + print(f"\n[2] THROUGHPUT — is it fast, is the tail healthy") + if t.get("tok_s") is not None: + _line("tok/s", f"{t['tok_s']:.3f}") + else: + flags.append("throughput line missing — engine output format may have changed") + if t.get("latency"): + la = t["latency"] + _line("decode forwards", f"{int(la['forwards'])}") + _line("p50 / p90", f"{la['p50_ms']:.1f} / {la['p90_ms']:.1f} ms") + _line("p99 / max", f"{la['p99_ms']:.1f} / {la['max_ms']:.1f} ms") + tail_ok = la["p99_ms"] <= HIGH_TAIL_RATIO * la["p50_ms"] + _line("tail ratio (p99/p50)", f"{la['p99_ms']/max(la['p50_ms'],1e-9):.2f}x", + _flag(tail_ok), + "high tail = decode stalls (I/O hiccups, KV growth, re-pin)") + if not tail_ok: + flags.append(f"tail latency p99={la['p99_ms']:.1f}ms >> p50={la['p50_ms']:.1f}ms " + "(look for REPIN swaps or disk stalls)") + + # ---------------------------------------------------------------- [3] TIME --- + print(f"\n[3] WHERE TIME GOES — what is doing it, when (share of decode)") + ts = t.get("time_shares") + prof = t.get("profile") + if ts: + order = [("io", "expert-disk I/O", DISK_WAIT_DOMINANT, "the cache is too small / disk is slow"), + ("matmul", "expert matmul", 0.40, "compute-bound; more cores or a GPU expert tier"), + ("attention", "attention", 0.35, "context length is the cost; lower CTX"), + ("head", "lm_head", 0.10, "vocab projection; unusual to dominate"), + ("other", "other", 0.30, "scheduling / KV bookkeeping overhead")] + for key, name, thresh, lever in order: + f = ts[key] + ok = f < thresh + _line(name, f"{f:5.0%} {_bar(f)}", _flag(ok), + "" if ok else f"->{lever}") + if not ok: + flags.append(f"{name} dominates ({f:.0%}) -> {lever}") + if t.get("verdict"): + print(f" engine verdict : {t['verdict']}") + elif prof: + print(" (no [PROF] time shares — set PROF=1 for phase percentages)") + if prof: + print(" absolute seconds :") + for k in ("disk", "expert_matmul", "attention", "lm_head", "other"): + _line(k, f"{prof[k]:.3f}s") + + # attention sub-breakdown: how is attention being read + ab = t.get("attn_breakdown") + if ab: + print(f"\n[3a] ATTENTION BREAKDOWN — how the attention phase is spent") + atot = sum(ab.values()) or 1.0 + for k, label in (("proj_rope", "projection + RoPE"), + ("score_sm_value", "score-softmax-value"), + ("out_proj", "output projection")): + _line(label, f"{ab[k]:.3f}s ({ab[k]/atot:.0%} of attn)") + + # ---------------------------------------------------------------- [4] CACHE -- + print(f"\n[4] EXPERT CACHE — is the cache efficient") + hit = t.get("hit_pct") + if hit is not None: + ok = hit >= LOW_HIT_RATE * 100 + _line("hit rate", f"{hit:.1f}%", _flag(ok), + "" if ok else "<30% = thrashing; raise RAM_GB or cap") + if not ok: + flags.append(f"cache hit {hit:.1f}% is low -> raise RAM_GB (or cap), add PIN_GB") + el = t.get("experts_loaded") + if el: + per_tok = el["per_tok"] + _line("experts loaded/token", f"{per_tok:.1f}") + _line(" per-layer", f"{el['per_layer']:.2f} across {el['n_sparse_layers']} sparse layers") + _line(" baseline", f"topk={el['baseline_topk']} active experts/token") + base_topk = el["baseline_topk"] + if base_topk > 0 and per_tok > 2 * base_topk: + flags.append(f"loading {per_tok:.0f} experts/token vs topk={base_topk} " + "-> redundant I/O; cache is re-fetching evicted experts") + + # ---------------------------------------------------------------- [5] DISK --- + print(f"\n[5] DISK I/O — is I/O the bottleneck, and where") + eio = t.get("expert_io") + if eio: + _line("total fetched", f"{eio['gb_fetched']:.3f} GB") + _line("per token", f"{eio['mb_per_tok']:.1f} MB/token") + _line("disk throughput", f"{eio['gb_per_s']:.2f} GB/s over the run") + _line("read service", f"{eio['read_service_s']:.2f}s (on I/O threads)") + _line("felt wait", f"{eio['felt_wait_s']:.2f}s (stall compute felt)") + if eio["felt_wait_s"] > eio["read_service_s"] * 0.5 and eio["read_service_s"] > 0: + flags.append("felt wait is a large fraction of read service -> PIPE=1 may not be " + "overlapping fully, or DIRECT=1 on NVMe") + ds = t.get("disk_split") + if ds: + print(f"\n[5a] DISK-LOAD SPLIT — which decode phase reads the bytes") + _line("draft phase", f"{ds['draft']} loads") + _line("absorb phase", f"{ds['absorb']} loads") + _line("verify/main", f"{ds['verify_main']} loads") + _line("MTP-layer bytes", f"{ds['mtp_loads']} loads, {ds['mtp_gb']:.2f} GB") + _line("main-layer bytes", f"{ds['main_loads']} loads, {ds['main_gb']:.2f} GB") + if ds.get("mtp_bytes_pct") is not None: + _line("MTP share of bytes", f"{ds['mtp_bytes_pct']:.1f}%") + share = disk_wait_share(t) + if share is not None: + ok = share < DISK_WAIT_DOMINANT + _line("disk-wait share", f"{share:.0%}", _flag(ok), + "" if ok else "I/O-bound (see levers in [3])") + + # ---------------------------------------------------------------- [6] ROUTE -- + print(f"\n[6] ROUTING QUALITY — is the router / prefetch accurate") + ra = t.get("route_agree") + if ra: + ok = ra["agree_pct"] >= LOW_ROUTE_AGREE * 100 + _line("route_agree", f"{ra['agree_pct']:.1f}% overlap with true top-K", + _flag(ok), + "" if ok else "prefetch is guessing wrong; CACHE_ROUTE params may need tuning") + _line("route_kl", f"{ra['kl']:.4f} mean KL (lower = closer to true routing)") + if not ok: + flags.append(f"route_agree {ra['agree_pct']:.1f}% low -> tune ROUTE_J/M/P, " + "or prefetch is hurting more than helping") + sw = t.get("swap") + if sw: + _line("cache swaps", f"{sw['swaps']}/{sw['slots']} ({sw['pct']:.1f}%)", + None, "high swap = churn between turns") + la = t.get("lookahead") + if la: + print(f"\n[6a] ROUTING PREDICTABILITY — recall of true experts in predicted top-8") + print(" (which predictor should drive prefetch? highest recall wins)") + for row in la: + _line(row["predictor"][:34], f"{row['pct']:5.1f}% ({row['hit']}/{row['tot']})") + + # ---------------------------------------------------------------- [7] SPEC --- + print(f"\n[7] SPECULATION — is the draft decoder pulling weight") + sp = t.get("speculation") + if sp: + _line("tokens/forward", f"{sp['tok_per_fw']:.2f} (>1.0 means speculation helps)") + _line("forwards/tokens", f"{sp['forwards']} forwards for {sp['tokens']} tokens") + # Speculation helps only if acceptance is high enough that tok/forward > 1. + # tok_per_fw already == 1.0 when nothing verifies, so judge by acceptance. + ok = sp["mtp_accept_pct"] >= LOW_MTP_ACCEPT * 100 + _line("MTP acceptance", f"{sp['mtp_accept_pct']:.0f}%", _flag(ok), + "" if ok else "<20% -> drafts rarely verify; DRAFT=0 may be faster") + if not ok: + flags.append(f"MTP acceptance {sp['mtp_accept_pct']:.0f}% low -> " + "drafts cost more I/O than they save; try DRAFT=0") + + # ---------------------------------------------------------------- [8] GPU ---- + print(f"\n[8] GPU TIERS — is the GPU actually used") + cuda = t.get("cuda") or {} + if not cuda.get("enabled"): + print(" (CUDA not enabled — CPU-only run)") + else: + if cuda.get("resident_tensors") is not None: + _line("resident dense tensors", f"{cuda['resident_tensors']} tensors, " + f"{cuda['resident_gb']:.2f} GB") + if cuda.get("expert_count") is not None: + waste = cuda["calls_served"] <= VRAM_WASTE_CALLS + _line("expert tier", f"{cuda['expert_count']} experts pinned " + f"({cuda['expert_gb']:.2f} GB)", _flag(not waste)) + _line(" calls served", f"{cuda['calls_served']} from VRAM", + _flag(not waste), + "" if not waste else "WASTE: pinned but never routed -> lower CUDA_EXPERT_GB") + if waste: + flags.append("VRAM expert tier has 0 calls served -> experts pinned but unused; " + "PIN stats may not match this workload") + if cuda.get("groups"): + g = cuda["groups"] + _line("expert groups", f"{g['calls']} calls, {g['experts']} experts, " + f"{g['rows']} rows ({g['experts_per_call']:.1f} experts/call)") + if cuda.get("groups_timing"): + gt = cuda["groups_timing"] + _line("GPU timing", f"H2D {gt['h2d_ms']:.1f} ms | kernel {gt['kernel_ms']:.1f} ms | " + f"D2H {gt['d2h_ms']:.1f} ms") + if gt["h2d_ms"] + gt["d2h_ms"] > gt["kernel_ms"]: + flags.append("CUDA H2D+D2H > kernel time -> transfer-bound; " + "consider larger expert tier to keep weights resident") + + # ---------------------------------------------------------------- summary ---- + print("\n" + "=" * 78) + if flags: + print(f" {len(flags)} FLAG(s) — the most likely levers to move tok/s:") + for f in flags: + print(f" - {f}") + else: + print(" no flags — every measured subsystem is within advisory thresholds.") + print("=" * 78) + return 0 + + +if __name__ == "__main__": + raise SystemExit(main()) diff --git a/c/tests/test_env_defaults.py b/c/tests/test_env_defaults.py index 12a3c99..3bb08fc 100644 --- a/c/tests/test_env_defaults.py +++ b/c/tests/test_env_defaults.py @@ -32,9 +32,16 @@ def args(**over): class EnvDefaultsTest(unittest.TestCase): - def env_for_with(self, environ, platform): + def env_for_with(self, environ, platform, cuda=False): + """Run env_for on a bare-chat args() under a faked env + platform. + + cuda=False by default so the existing default-I/O tests stay + deterministic: the Windows auto-enable branch calls cuda_binary() and + (if True) discover_gpus(), both of which reach the real machine — faking + False keeps these tests independent of the host's GPU.""" with mock.patch.dict(os.environ, environ, clear=True), \ - mock.patch.object(sys, "platform", platform): + mock.patch.object(sys, "platform", platform), \ + mock.patch.object(coli, "cuda_binary", return_value=cuda): return coli.env_for(args()) def test_win32_sets_measured_defaults(self): @@ -64,5 +71,80 @@ class EnvDefaultsTest(unittest.TestCase): self.assertNotIn(k, e) +class CudaAutoEnableTest(unittest.TestCase): + """Windows bare `coli chat` (no --gpu/--vram/--auto-tier) used to ALWAYS run + CPU-only even on a CUDA build with a GPU present. env_for now auto-enables + CUDA on win32 when cuda_binary() is True and a GPU is discoverable; falls + back to CPU with a warning if nvidia-smi (discover_gpus) is missing; stays + silent on a CPU build; and never touches the Linux path.""" + + def _env_for(self, platform, cuda, gpus, plan=None): + # Patch discover_gpus / build_plan / environment_for_plan at the + # resource_plan module (env_for imports them lazily on each call, so the + # patches are live when those imports run). Stubbing the planner keeps + # the test independent of a real model dir (args().model == "X"). + import resource_plan + a = args() + GPB = 1024 ** 3 + if plan is None: + plan = {"tiers": {"ram": {"budget_bytes": 16 * GPB, "cache_slots_per_layer": 4}, + "vram": {"budget_bytes": int(8.0 * GPB), "devices": gpus}}} + + def fake_environment_for_plan(p, env, cuda_enabled=True): + # Mirror the real contract: size CUDA_EXPERT_GB from the plan's VRAM + # budget (this is the value env_for propagates into the engine env). + r = dict(env) + if cuda_enabled and p["tiers"]["vram"]["devices"] and p["tiers"]["vram"]["budget_bytes"] > 0: + r["CUDA_EXPERT_GB"] = f"{p['tiers']['vram']['budget_bytes'] / GPB:.3f}" + return r + + with mock.patch.dict(os.environ, {}, clear=True), \ + mock.patch.object(sys, "platform", platform), \ + mock.patch.object(coli, "cuda_binary", return_value=cuda), \ + mock.patch.object(resource_plan, "discover_gpus", return_value=gpus), \ + mock.patch.object(resource_plan, "build_plan", return_value=plan), \ + mock.patch.object(resource_plan, "environment_for_plan", + side_effect=fake_environment_for_plan): + return coli.env_for(a) + + def _fake_gpu(self, index=0, name="NVIDIA GeForce RTX 5070 Ti", + total_mib=16384, free_mib=15000): + return {"index": index, "name": name, + "total_bytes": total_mib * 1024 * 1024, + "free_bytes": free_mib * 1024 * 1024} + + def test_win32_auto_enables_cuda_when_gpu_present(self): + e = self._env_for("win32", cuda=True, gpus=[self._fake_gpu()]) + self.assertEqual(e["COLI_CUDA"], "1") + self.assertEqual(e["COLI_GPUS"], "0") + # VRAM budget is sized from free VRAM by build_plan (real minus reserve), + # so it must be present and positive — never a guess or zero. + self.assertIn("CUDA_EXPERT_GB", e) + self.assertGreater(float(e["CUDA_EXPERT_GB"]), 0.0) + # Dense offload is an explicit opt-in (matches --auto-tier): not set here. + self.assertNotIn("CUDA_DENSE", e) + + def test_win32_falls_back_to_cpu_when_nvidia_smi_missing(self): + # coli_cuda.dll present (cuda=True) but nvidia-smi absent (no GPUs found) + # -> warn + CPU-only, never crash, never set COLI_CUDA. + e = self._env_for("win32", cuda=True, gpus=[]) + self.assertNotIn("COLI_CUDA", e) + self.assertNotIn("COLI_GPUS", e) + self.assertNotIn("CUDA_EXPERT_GB", e) + + def test_win32_cpu_build_stays_silent(self): + # No coli_cuda.dll (cuda=False) -> CPU build, nothing GPU-related emitted. + e = self._env_for("win32", cuda=False, gpus=[self._fake_gpu()]) + self.assertNotIn("COLI_CUDA", e) + self.assertNotIn("COLI_GPUS", e) + + def test_linux_bare_chat_not_auto_enabled(self): + # The auto-enable is scoped to win32: a Linux bare chat with a GPU + # present must NOT turn CUDA on (Linux keeps the explicit-flag UX). + e = self._env_for("linux", cuda=True, gpus=[self._fake_gpu()]) + self.assertNotIn("COLI_CUDA", e) + self.assertNotIn("CUDA_EXPERT_GB", e) + + if __name__ == "__main__": unittest.main() diff --git a/c/tests/test_grouped_g4_cuda.cu b/c/tests/test_grouped_g4_cuda.cu new file mode 100644 index 0000000..b370cdd --- /dev/null +++ b/c/tests/test_grouped_g4_cuda.cu @@ -0,0 +1,108 @@ +/* Grouped-int4 (fmt=4) CUDA kernel oracle (#334). + * + * Feeds random offset-binary nibble weights + [O, ng] group scales through + * grouped_hidden_g4_dual / grouped_down_g4 and checks against a CPU reference + * that replicates matmul_i4_grouped's semantics (value = nibble - 8, per-group + * partial dot x scale). Covers gs=64, a non-divisible tail group, and a + * per-row (gs=0) member riding in the same launch — the fmt=2-compat case. + * + * The device buffers get the same XOR 0x88 offset->signed conversion the + * upload path applies, so the kernels are exercised exactly as deployed. + * + * Build: nvcc -O2 -std=c++17 -arch=native tests/test_grouped_g4_cuda.cu -o tests/test_grouped_g4 + */ +#include +#include +#include +#include +#include + +#include "../backend_cuda.cu" + +static void cpu_gemv_g4(const uint8_t *q,const float *sc,int K,int O,int gs, + const float *x,float *y){ + int rb=(K+1)/2, ng=gs>0?(K+gs-1)/gs:1, egs=gs>0?gs:K; + for(int o=0;oK) glen=K-base; + double p=0; + for(int i=base;i>1]; int n=(i&1)?(v>>4):(v&15); + p+=(double)x[i]*(n-8); + } + a+=p*scl[g]; + } + y[o]=(float)a; + } +} + +int main(void){ + srand(7); + const int D=200, I=96, gs=64; /* tail group: 200 % 64 = 8 */ + const int COUNT=3; /* expert 0,1: fmt4 gs=64; expert 2: per-row (gs=0) */ + const int rbD=(D+1)/2, rbI=(I+1)/2; + const int ngD=(D+gs-1)/gs, ngI=(I+gs-1)/gs; + int trials=50, bad=0; + for(int t=0;t>>(qg[c],(size_t)I*rbD); + offset_to_signed_s4<<<64,256>>>(qu[c],(size_t)I*rbD); + offset_to_signed_s4<<<64,256>>>(qd[c],(size_t)D*rbI); + cudaMemcpy(sg[c],hgs[c],(size_t)I*cngD*4,cudaMemcpyHostToDevice); + cudaMemcpy(su[c],hus[c],(size_t)I*cngD*4,cudaMemcpyHostToDevice); + cudaMemcpy(sd[c],hds[c],(size_t)D*cngI*4,cudaMemcpyHostToDevice); + host[c]={qg[c],qu[c],qd[c],sg[c],su[c],sd[c],4,4,4,1,c,cgs,cgs,cgs}; + } + for(size_t i=0;i<(size_t)COUNT*D;i++) xs[i]=(rand()/(float)RAND_MAX-.5f)*2.f; + GroupDesc *ddesc; cudaMalloc(&ddesc,sizeof(host)); + cudaMemcpy(ddesc,host,sizeof(host),cudaMemcpyHostToDevice); + dim3 hgd((unsigned)I,1,(unsigned)COUNT),ogd((unsigned)D,1,(unsigned)COUNT); + grouped_hidden_g4_dual<<>>(gate,up,xs,ddesc,I,D); + grouped_down_g4<<>>(y,gate,ddesc,D,I); + if(cudaDeviceSynchronize()!=cudaSuccess){ printf("FAIL cuda\n"); return 1; } + for(int c=0;c1e-3f*(fabsf(rg[o])+1e-3f)|| + fabsf(up[(size_t)c*I+o]-ru[o])>1e-3f*(fabsf(ru[o])+1e-3f)) bad++; + } + cpu_gemv_g4(hd[c],hds[c],I,D,cgs,(float*)gate+(size_t)c*I,ry); + for(int o=0;o1e-3f*(fabsf(ry[o])+1e-3f)) bad++; + } + for(int c=0;c bool: + """True iff BOTH the built engine AND the tiny fixture are available. + + These tests need glm.exe (a build artifact) AND glm_tiny/ (a generated + fixture, gitignored). CI runs `make check` = "dependency-free tests, no + model downloads" (workflow .github/workflows/check.yml, by design #140), so + neither is present there and these tests must SKIP rather than fail. They + run locally after `make glm.exe` (glm_tiny ships alongside the source, or + is regenerated by tools/make_glm_oracle.py). + """ + return ENGINE.exists() and (TINY / "config.json").exists() + + +def _skip_reason() -> str: + """Name exactly which prerequisite is missing, so the skip is actionable.""" + if not ENGINE.exists(): + return "glm.exe not built (run: make glm.exe)" + if not (TINY / "config.json").exists(): + return "glm_tiny fixture absent (gitignored; ship it locally or run tools/make_glm_oracle.py)" + return "" + + +def _cuda_available() -> bool: + """True iff the engine binary has the CUDA loader compiled in AND the DLL is present. + + The host is built with -DCOLI_CUDA only when CUDA_DLL=1 (Makefile). A binary + built without it embeds the string "this binary is CPU-only; rebuild" and + exits 2 on any CUDA env var — so we detect the CPU-only build by scanning + the binary for that marker (avoids a slow ldd/strings on every import; we + only read enough to find it). On Windows the DLL is also required + (backend_loader.c loads it at runtime); on Linux it's direct-linked. + """ + if not _engine_present(): + return False + try: + # Read once; the marker is near the read-only string table. 256 KB is + # plenty for this string and avoids loading a 1 MB binary into memory. + with open(ENGINE, "rb") as f: + blob = f.read(2 * 1024 * 1024) + if b"this binary is CPU-only" in blob: + return False # CPU-only build: COLI_CUDA=1 would exit 2. + except OSError: + pass + # Windows: DLL required at runtime. Linux: direct-linked (no DLL). + if os.name == "nt": + return (C_DIR / "coli_cuda.dll").exists() + import subprocess + try: + out = subprocess.run(["ldd", str(ENGINE)], capture_output=True, text=True) + return "libcudart" in out.stdout + except (FileNotFoundError, OSError): + return False + + +@unittest.skipUnless(_engine_present(), _skip_reason() or "glm.exe + glm_tiny required") +class TinyEfficiencyTest(unittest.TestCase): + """Asserted regression tests on the resident tiny model. Gates CI.""" + + def _run(self, **overlay): + return run_engine(overlay, engine=str(ENGINE), snap=str(TINY))[0] + + # -- telemetry contract --------------------------------------------------- + + def test_telemetry_parses(self): + """A REPLAY run must emit the throughput + PROFILE lines the suite keys on. + + If this fails, either the engine changed its output format (update the + parsers in tools/efficiency.py) or the run crashed early.""" + t = self._run(REPLAY="1", TEMP="0", NGEN="4") + self.assertEqual(t["returncode"], 0, f"engine exited non-zero:\n{t['stderr']}") + self.assertIn("tok_s", t["parsed"], f"missing tok/s line:\n{t['stderr']}") + self.assertIn("profile", t["parsed"], f"missing PROFILE line:\n{t['stderr']}") + self.assertIn("hit_pct", t["parsed"], f"missing expert hit line:\n{t['stderr']}") + self.assertIsNotNone(t["tok_s"]) + self.assertIsNotNone(t["profile"]) + + # -- throughput floor ----------------------------------------------------- + + def test_tiny_tok_s_floor(self): + """Tiny decode must beat TINY_TOK_S_FLOOR. + + The tiny model is fully resident and runs ~200 tok/s; the 20 tok/s + default floor is a 10x margin that catches broken builds or a pathological + config cascade (the cap=1 trap from ISSUE_new_model_resource_regression.md) + without flapping on machine noise.""" + t = self._run(REPLAY="1", TEMP="0", NGEN="8") + self.assertEqual(t["returncode"], 0, f"engine exited non-zero:\n{t['stderr']}") + self.assertGreaterEqual( + t["tok_s"], TINY_TOK_S_FLOOR, + f"tok/s {t['tok_s']:.1f} below floor {TINY_TOK_S_FLOOR} " + f"(regression, or a config cascade starving the cache)", + ) + + # -- accounting sanity ---------------------------------------------------- + + def test_profile_phases_present_and_nonneg(self): + """Every PROFILE phase must be present and non-negative. + + `other` can go slightly negative from timer overhead (the engine allows + it), but a large negative means the timers are double-counting.""" + t = self._run(REPLAY="1", TEMP="0", NGEN="4") + p = t["profile"] + for phase in ("disk", "expert_matmul", "attention", "lm_head"): + self.assertGreaterEqual(p[phase], -0.01, f"{phase} went negative: {p}") + # 'other' is a residual; allow a small negative from timer overlap. + self.assertGreaterEqual(p["other"], -0.05, f"other too negative (double-count): {p}") + + # -- disk-wait not dominant on a resident model --------------------------- + + def test_disk_wait_not_dominant(self): + """A fully-resident tiny model must NOT be I/O-bound. + + Everything fits in RAM; the expert-disk wait share should be ~0. If it + exceeds MAX_DISK_WAIT_SHARE, the cache/I/O path regressed — on a real + model this same regression would make decode I/O-bound (the exact + failure mode the [PROF] verdict flags).""" + t = self._run(REPLAY="1", TEMP="0", NGEN="8", PROF="1") + self.assertIn("time_shares", t["parsed"], f"missing [PROF] time shares:\n{t['stderr']}") + share = disk_wait_share(t) + self.assertIsNotNone(share) + self.assertLess( + share, MAX_DISK_WAIT_SHARE, + f"expert-I/O share {share:.0%} on a resident model — I/O path regressed", + ) + + # -- determinism ---------------------------------------------------------- + + def test_cpu_vs_cpu_determinism(self): + """Two greedy REPLAY runs with the same seed produce identical telemetry. + + TEMP=0 = greedy (no sampling), so tok/s and hit-rate must be reproducible. + A drift here means non-determinism crept into the decode path (stray + threading, uninitialized state) — which on a real model would make A/B + comparisons meaningless.""" + a = self._run(REPLAY="1", TEMP="0", NGEN="8", SEED="1") + b = self._run(REPLAY="1", TEMP="0", NGEN="8", SEED="1") + self.assertEqual(a["returncode"], 0) + self.assertEqual(b["returncode"], 0) + self.assertEqual(a["hit_pct"], b["hit_pct"], "greedy hit-rate drifted between runs") + # tok/s within 25% — exact equality is too strict across scheduler noise. + self.assertLess(abs(a["tok_s"] - b["tok_s"]) / max(a["tok_s"], b["tok_s"]), 0.25) + + +@unittest.skipUnless(_cuda_available(), + _skip_reason() or "CUDA build not present (run: make clean && make glm.exe CUDA_DLL=1 && make cuda-dll)") +class TinyCudaEfficiencyTest(unittest.TestCase): + """CUDA-path regression tests on the tiny model. Skip unless CUDA built. + + glm_tiny is small and fully resident, so CUDA here is fast and exercises the + real GPU code path (init, dense upload, kernel correctness) without the + long load time or memory pressure of the full model. These guard the + silent-failure modes that are otherwise invisible: + - COLI_CUDA=1 silently falling back to CPU (loader/DLL missing) + - CUDA_DENSE=1 uploading nothing + - a CUDA kernel producing different argmax than CPU on identical inputs + """ + + def _run(self, **overlay): + return run_engine(overlay, engine=str(ENGINE), snap=str(TINY))[0] + + def test_cuda_init_path(self): + """COLI_CUDA=1 must initialize the device and NOT exit 2. + + Exit 2 is the engine's "requested backend is unavailable" path + (glm.c: g_cuda_enabled check). A clean init prints the [CUDA] device + banner to stderr. If this fails, the DLL is broken or the loader can't + resolve symbols (ABI drift between backend_cuda.h and the dll).""" + t = self._run(COLI_CUDA="1", COLI_GPU="0", REPLAY="1", TEMP="0", NGEN="4") + self.assertNotEqual(t["returncode"], 2, + f"engine refused CUDA backend:\n{t['stderr']}") + self.assertTrue(t["cuda"]["enabled"], + f"no [CUDA] device banner on stderr:\n{t['stderr']}") + + def test_cuda_dense_uses_vram(self): + """CUDA_DENSE=1 must actually upload dense tensors to VRAM. + + The minimal GPU-exercising config (per backend_loader.c analysis): + COLI_CUDA=1 + CUDA_DENSE=1. Without CUDA_DENSE the dense path stays on + CPU and [CUDA] resident set reports 0 tensors — a silent no-op. This + catches that regression: after the run, resident_tensors > 0.""" + t = self._run(COLI_CUDA="1", COLI_GPU="0", CUDA_DENSE="1", + REPLAY="1", TEMP="0", NGEN="4") + self.assertEqual(t["returncode"], 0, f"engine exited non-zero:\n{t['stderr']}") + rt = t["cuda"]["resident_tensors"] + self.assertIsNotNone(rt, f"no [CUDA] resident set line:\n{t['stderr']}") + self.assertGreater(rt, 0, + f"CUDA_DENSE=1 but {rt} tensors resident — silent CPU fallback") + + def test_cpu_vs_cuda_tf_match(self): + """CPU and CUDA teacher-forcing must agree on most positions (DIRECTLY). + + Both paths prefill the SAME oracle sequence on the SAME weights, so their + argmaxes should match position-for-position — but not exactly: the two + backends accumulate dot-products in different orders (x86 SIMD vs CUDA + kernel), so a few near-tied logits flip. That divergence is expected + numeric behavior, not a kernel bug. A *catastrophic* kernel regression + (wrong GEMM, wrong scale, wrong fmt) would collapse agreement toward + random (~1/vocab = ~4%); the MIN_CPU_CUDA_AGREEMENT floor (default 70%) + catches that while tolerating harmless drift. + + We compare CPU-vs-CUDA *directly* (not via the oracle match-counts), + because both backends differ from the oracle at different positions and + the summary line can't tell "CPU≠CUDA" from "CPU≠oracle".""" + import json + ref = json.loads((C_DIR / "ref_glm.json").read_text()) + oracle = ref["tf_pred"] + + cpu = self._run(TF="1", TEMP="0") + cuda = self._run(COLI_CUDA="1", COLI_GPU="0", CUDA_DENSE="1", TF="1", TEMP="0") + self.assertEqual(cpu["returncode"], 0, f"CPU TF run failed:\n{cpu['stderr']}") + self.assertEqual(cuda["returncode"], 0, f"CUDA TF run failed:\n{cuda['stderr']}") + self.assertIn("tf_mismatches", cpu["parsed"], "CPU run missing per-position mismatches") + self.assertIn("tf_mismatches", cuda["parsed"], "CUDA run missing per-position mismatches") + + agree, diff = tf_agreement(cpu, cuda, oracle) + self.assertGreaterEqual( + agree, MIN_CPU_CUDA_AGREEMENT, + f"CPU-vs-CUDA argmax agreement {agree:.0%} below floor " + f"{MIN_CPU_CUDA_AGREEMENT:.0%} — CUDA kernel likely regressed. " + f"Differing positions: {diff[:10]}{'...' if len(diff)>10 else ''}", + ) + + +if __name__ == "__main__": + unittest.main() diff --git a/c/tests/test_int3.c b/c/tests/test_int3.c new file mode 100644 index 0000000..9fa7c73 --- /dev/null +++ b/c/tests/test_int3.c @@ -0,0 +1,147 @@ +/* int3-g64 (fmt=5) tests: pack layout, dequant round-trip vs plain-C reference, + * matmul_i3 (NEON + scalar tail) vs reference dequant-matmul, per-row helpers, + * the .qs-size format tag, and the quality claim in miniature (per-group int3 + * beats per-row int4 on rows with outliers — the #132 result this format ships). */ +#define main coli_glm_main_unused +#include "../colibri.c" +#undef main + +#include +#include +#include + +static int fails = 0; +#define CHECK(c) do{ if(!(c)){ printf("FAIL %s:%d: %s\n", __FILE__, __LINE__, #c); fails++; } }while(0) + +static uint64_t rng = 0x9E3779B97F4A7C15ull; +static float rndf(void){ rng ^= rng << 13; rng ^= rng >> 7; rng ^= rng << 17; + return ((int64_t)(rng & 0xFFFFF) - 0x80000) / (float)0x80000; } + +/* reference: quantize like pack_int3_g64 but keep dequantized f32 (mirrors + * quant_ablation._quant_last_dim(bits=3, group=64)) */ +static void ref_i3_dequant(const float *w, float *dq, int O, int I){ + int64_t ng=i3_groups(I); + for(int o=0;oamax)amax=a; } + float s=amax/3.f; if(s<1e-8f)s=1e-8f; + for(int k=0;k3)v=3; if(v<-4)v=-4; + dq[(int64_t)o*I+base+k]=(float)v*s; + } + } +} +static void unpack_i3(const uint8_t *q3, const float *s, float *dq, int O, int I){ + int64_t ng=i3_groups(I), rb=i3_rowbytes(I); + for(int o=0;o>2]>>((k&3)*2))&3)|(((hi[k>>3]>>(k&7))&1)<<2); + dq[(int64_t)o*I+base+k]=(float)((int)u-4)*s[(int64_t)o*ng+g]; + } + } +} + +int main(void){ + const int Is[]={64,128,192,100,65,7168}; /* incl. short tail groups and one real GLM dim */ + enum { O=7, MAXI=7168 }; + static float w[(int64_t)O*MAXI], dq_ref[(int64_t)O*MAXI], dq_pk[(int64_t)O*MAXI]; + static float x[4*MAXI], y_ref[4*O], y_ker[4*O]; + static uint8_t q3[(int64_t)O*(MAXI/64+1)*24]; + static float sc[(int64_t)O*(MAXI/64+1)]; + + for(unsigned c=0;c unpack == reference quantize-dequantize, bit for bit */ + pack_int3_g64(w, q3, sc, O, I); + ref_i3_dequant(w, dq_ref, O, I); + unpack_i3(q3, sc, dq_pk, O, I); + int bad=0; + for(int64_t i=0;i<(int64_t)O*I;i++) if(dq_pk[i]!=dq_ref[i]) bad++; + CHECK(bad==0); + + /* 2. matmul_i3 == matmul over the dequantized reference (fp tolerance: + * NEON fma order differs from the scalar reference loop) */ + for(int S=1;S<=4;S+=3){ + for(int64_t i=0;i<(int64_t)S*I;i++) x[i]=rndf(); + matmul_i3(y_ker, x, q3, sc, S, I, O); + for(int s=0;s1?fabsf(y_ref[i]):1; + if(d/m>2e-4f){ CHECK(!"matmul_i3 mismatch"); break; } + } + } + + /* 3. QT plumbing: qt_alloc(bits=3) -> qt_fill -> matmul_qt & qt_bytes & helpers */ + QT t; qt_alloc(&t, O, I, 3); + CHECK(t.fmt==5); + qt_fill(&t, w, 3); + CHECK(qt_bytes(&t)==(int64_t)O*i3_rowbytes(I)+(int64_t)O*i3_groups(I)*4); + matmul_qt(y_ker, x, &t, 1); + for(int o=0;o1?fabsf((float)a):1; + CHECK(d/m<=2e-4f); + } + float acc[MAXI]; memset(acc,0,I*sizeof(float)); + qt_addrow(&t, 2, 0.5f, acc); + for(int i=0;i1?fabsf((float)a):1; + CHECK(d/m<=2e-4f); + } + + /* 4. format resolution through the #413 gate: fmt=5 is tagged by its distinct + * WEIGHT byte count. int3-g64 and grouped-int4-at-gs=64 carry the SAME scale + * cardinality O*ceil(I/64), so the pair (weight bytes, scale bytes) must + * disambiguate: same scales, int4 weights -> fmt=4/gs=64; int3 weights -> fmt=5. + * Only well-posed for I > 256: below that, O row scales legitimately match a + * 1-group grouped layout too (detect_group_size probes gs up to 256), so + * per-row vs grouped is not distinguishable from byte counts alone. */ + if(I>256){ int gs=-1; + int64_t ns_g64=(int64_t)O*i3_groups(I)*4, ns_row=(int64_t)O*4; + CHECK(qt_resolve_fmt("t.i3", O, I, (int64_t)O*i3_rowbytes(I), ns_g64, &gs)==5); + CHECK(gs==0); + CHECK(qt_resolve_fmt("t.i8", O, I, (int64_t)O*I, ns_row, &gs)==1); + CHECK(qt_resolve_fmt("t.i4", O, I, (int64_t)O*((I+1)/2), ns_row, &gs)==2); + CHECK(qt_resolve_fmt("t.i4g", O, I, (int64_t)O*((I+1)/2), ns_g64, &gs)==4); + CHECK(gs==64); + CHECK(qt_resolve_fmt("t.i2", O, I, (int64_t)O*((I+3)/4), ns_row, &gs)==3); } + free(t.q4); free(t.s); + } + + /* 5. quality in miniature: on rows with outliers, per-group int3 must beat + * per-row int4 on reconstruction RMS (the #132 finding this format ships). */ + { + int I=1024; + for(int64_t i=0;i<(int64_t)O*I;i++) w[i]=rndf()*0.02f; + for(int o=0;o>1]; + int v=(i&1)?((int)(b>>4)-8):((int)(b&0xF)-8); w4=(float)v*t4.s[o]; } + double d3=w[(int64_t)o*I+i]-dq_ref[(int64_t)o*I+i], d4=w[(int64_t)o*I+i]-w4; + e3+=d3*d3; e4+=d4*d4; + } + CHECK(e3 < e4); + printf(" outlier-rows RMS: int3-g64 %.3e < int4-row %.3e (ratio %.2f)\n", + sqrt(e3/((double)O*I)), sqrt(e4/((double)O*I)), sqrt(e4/e3)); + free(t4.q4); free(t4.s); + } + + if(fails){ printf("int3-g64 tests: %d FAILED\n", fails); return 1; } + printf("int3-g64 tests: ok\n"); + return 0; +} diff --git a/c/tests/test_int3_convert.py b/c/tests/test_int3_convert.py new file mode 100644 index 0000000..2eaa7f7 --- /dev/null +++ b/c/tests/test_int3_convert.py @@ -0,0 +1,70 @@ +"""quant_int3_g64 (tools/convert_fp8_to_int4.py): pack layout + round-trip. + +Decodes the packed bytes with an independent NumPy decoder implementing the +fmt=5 spec (16B low plane / 8B high plane per 64-group, v+4, per-group f32 +scale) and checks the dequantized result equals the reference +quantize-dequantize (same math as quant_ablation._quant_last_dim(3, 64)). +The C side of the same layout is covered by tests/test_int3.c. +""" +import os, sys, unittest +try: + import numpy as np +except ImportError: + raise unittest.SkipTest("numpy not installed") + +sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "tools")) +from convert_fp8_to_int4 import quant_int3_g64 + + +def decode(packed, scales, O, I, group=64): + ng = (I + group - 1) // group + b = packed.reshape(O, ng, 24) + lo, hi = b[:, :, :16], b[:, :, 16:] + k = np.arange(group) + lov = (lo[:, :, k >> 2] >> ((k & 3) * 2)[None, None, :]) & 3 + hiv = (hi[:, :, k >> 3] >> (k & 7)[None, None, :]) & 1 + v = (lov | (hiv << 2)).astype(np.int64) - 4 + dq = v.astype(np.float64) * scales.reshape(O, ng, 1).astype(np.float64) + return dq.reshape(O, ng * group)[:, :I] + + +def reference(w, group=64): + """same math as quant_int3_g64 (which works in f32), replayed exactly, then + dequantized in f64 so it matches decode() bit for bit""" + O, I = w.shape + ng = (I + group - 1) // group + pad = ng * group - I + wp = np.pad(w, ((0, 0), (0, pad))) if pad else w + g = wp.reshape(O, ng, group) + s = np.maximum(np.abs(g).max(axis=2, keepdims=True) / 3.0, 1e-8).astype(np.float32) + q = np.clip(np.rint(g / s), -4, 3).astype(np.int64) + return (q.astype(np.float64) * s.astype(np.float64)).reshape(O, ng * group)[:, :I] + + +class Int3ConvertTest(unittest.TestCase): + def test_round_trip(self): + rng = np.random.default_rng(7) + for I in (64, 128, 100, 65, 7168): + w = (rng.standard_normal((5, I)) * 0.05).astype(np.float32) + w[0, 3] = 1.7; w[2, min(5, I - 1)] = -2.2 + packed, scales = quant_int3_g64(w) + ng = (I + 63) // 64 + self.assertEqual(packed.size, 5 * ng * 24) + self.assertEqual(scales.size, 5 * ng) + np.testing.assert_allclose(decode(packed, scales, 5, I), + reference(w), rtol=0, atol=0) + + def test_outliers_beat_row_int4(self): + rng = np.random.default_rng(11) + w = (rng.standard_normal((8, 1024)) * 0.02).astype(np.float32) + for o in range(8): w[o, (o * 37) % 1024] = 1.5 + packed, scales = quant_int3_g64(w) + e3 = float(((decode(packed, scales, 8, 1024) - w) ** 2).mean()) + s4 = np.maximum(np.abs(w).max(axis=1, keepdims=True) / 7.0, 1e-8) + w4 = np.clip(np.rint(w / s4), -8, 7) * s4 + e4 = float(((w4 - w) ** 2).mean()) + self.assertLess(e3, e4) + + +if __name__ == "__main__": + unittest.main() diff --git a/c/tests/test_int3_load.c b/c/tests/test_int3_load.c new file mode 100644 index 0000000..57634c0 --- /dev/null +++ b/c/tests/test_int3_load.c @@ -0,0 +1,103 @@ +/* Loader-seam test for fmt=5: writes a real .safetensors file containing an + * int3-g64 tensor (U8 payload + per-GROUP .qs) next to an int4 control tensor + * (per-row .qs), indexes it with st_init, loads both through qt_from_disk, and + * checks the byte-count/.qs-size format inference picks fmt=5 vs fmt=2 correctly + * and the loaded weights dequantize identically to pack_int3_g64's output. */ +#define main coli_glm_main_unused +#include "../colibri.c" +#undef main + +#include +#include +#include +#include + +static int fails = 0; +#define CHECK(c) do{ if(!(c)){ printf("FAIL %s:%d: %s\n", __FILE__, __LINE__, #c); fails++; } }while(0) + +static uint64_t rng = 0xA5A5A5A55A5A5A5Aull; +static float rndf(void){ rng ^= rng << 13; rng ^= rng >> 7; rng ^= rng << 17; + return ((int64_t)(rng & 0xFFFFF) - 0x80000) / (float)0x80000; } + +static void deq4(const QT *t, float *dq){ + for(int o=0;oO;o++) for(int i=0;iI;i++){ + if(t->fmt==5){ + int64_t g=i/I3_GROUP; const uint8_t *lo=t->q4+(int64_t)o*i3_rowbytes(t->I)+g*I3_GBYTES, *hi=lo+16; + int k=i%I3_GROUP; + unsigned u=((lo[k>>2]>>((k&3)*2))&3)|(((hi[k>>3]>>(k&7))&1)<<2); + dq[(int64_t)o*t->I+i]=(float)((int)u-4)*t->s[(int64_t)o*i3_groups(t->I)+g]; + } else { /* fmt2 */ + uint8_t b=t->q4[(int64_t)o*((t->I+1)/2)+(i>>1)]; + int v=(i&1)?((int)(b>>4)-8):((int)(b&0xF)-8); + dq[(int64_t)o*t->I+i]=(float)v*t->s[o]; + } + } +} + +int main(void){ + enum { O=5, I=320 }; /* 5 groups per row; I > 256 so the per-row int4 + * control stays fmt=2 (detect_group_size probes + * gs up to 256: any smaller I would make O row + * scales match a legitimate 1-group layout) */ + int64_t ng=i3_groups(I), rb=i3_rowbytes(I); + static float w[O*I]; + for(int i=0;i> 28) & 0xF)) * 0.5 + for l in range(4): + seven = (word >> (7 * l)) & 0x7F + bits = [(seven >> j) & 1 for j in range(7)] + bits.append(sum(bits) & 1) # odd parity closes the 8th + idx = int(rows[r, base + ib * 8 + l * 2 + 0]) + idx2 = int(rows[r, base + ib * 8 + l * 2 + 1]) + mags = list(g[idx]) + list(g[idx2]) + for j in range(8): + pos = sb * P.QK + ib * P.SUB + l * 8 + j + out[r, pos] = mags[j] * db * (-1.0 if bits[j] else 1.0) + return out + + +@unittest.skipIf(P is None, "numpy not available (offline-tooling test)") +class TestIq3Pack(unittest.TestCase): + def setUp(self): + np.random.seed(4242) + self.x = (np.random.randn(6, 1024) * 0.05).astype(np.float32) + + def test_byte_budget(self): + self.assertEqual(P.BLOCK_BYTES, 98) + self.assertAlmostEqual(P.bpw(), 3.0625, places=6) + packed = P.encode(self.x) + self.assertEqual(packed.shape, (6, 1024 // P.QK * 98)) + self.assertEqual(packed.dtype, np.uint8) + + def test_encode_is_deterministic(self): + self.assertTrue(np.array_equal(P.encode(self.x), P.encode(self.x))) + + def test_decode_matches_spec_reader(self): + packed = P.encode(self.x) + fast = P.decode(packed, 1024) + slow = ref_decode(packed, 1024) + self.assertTrue(np.allclose(fast, slow, rtol=1e-6, atol=1e-8), + f"max |Δ| = {np.abs(fast - slow).max()}") + + def test_sign_parity_closes(self): + """Every 8-weight block must have an even number of negatives — that is + what lets the 8th sign be derived instead of stored.""" + y = P.decode(P.encode(self.x), 1024) + neg = (y < 0).reshape(-1, 8).sum(-1) + self.assertTrue(np.all(neg % 2 == 0), "a block stored odd negatives") + + def test_reconstruction_quality(self): + y = P.decode(P.encode(self.x), 1024) + rel = np.sqrt(((y - self.x) ** 2).mean()) / np.sqrt((self.x ** 2).mean()) + # the torch model that won the #453 A/B measures ~0.195 on this input class + self.assertLess(rel, 0.25, f"rel-RMSE {rel:.4f} — worse than the chosen scheme") + self.assertGreater(rel, 0.05, f"rel-RMSE {rel:.4f} — implausibly good, check the test") + + def test_shape_guard(self): + with self.assertRaises(ValueError): + P.encode(np.zeros((2, 300), dtype=np.float32)) + + +if __name__ == "__main__": + unittest.main() diff --git a/c/tests/test_openai_server.py b/c/tests/test_openai_server.py index 4cda9ef..8000b40 100644 --- a/c/tests/test_openai_server.py +++ b/c/tests/test_openai_server.py @@ -20,8 +20,8 @@ class FakeEngine: self.calls = [] def generate(self, prompt, maximum, temperature, top_p, on_text, cache_slot=0, - cancelled=None): - self.calls.append((prompt, maximum, temperature, top_p, cache_slot)) + cancelled=None, grammar=None): + self.calls.append((prompt, maximum, temperature, top_p, cache_slot, grammar)) on_text("Hé") on_text("llo") return {"prompt_tokens": 7, "completion_tokens": 2, "length_limited": False} @@ -34,7 +34,7 @@ class BlockingEngine(FakeEngine): self.release = threading.Event() def generate(self, prompt, maximum, temperature, top_p, on_text, cache_slot=0, - cancelled=None): + cancelled=None, grammar=None): self.entered.set() self.release.wait(2) return super().generate(prompt, maximum, temperature, top_p, on_text, cache_slot, @@ -71,11 +71,11 @@ class TemplateTest(unittest.TestCase): def test_validates_generation_limits(self): self.assertEqual(generation_options({"max_tokens": 4, "temperature": 0, "top_p": 1}, 8), - (4, 0.0, 1.0)) + (4, 0.0, 1.0, None)) # max_tokens above the server cap is clamped, not rejected (#260): OpenAI # clients default to large values; erroring breaks them. self.assertEqual(generation_options({"max_tokens": 9, "temperature": 0, "top_p": 1}, 8), - (8, 0.0, 1.0)) + (8, 0.0, 1.0, None)) # non-positive / non-int max_tokens is still a hard error with self.assertRaises(APIError): generation_options({"max_tokens": 0}, 8) @@ -84,7 +84,31 @@ class TemplateTest(unittest.TestCase): with self.assertRaises(APIError): generation_options({"top_p": math.inf}, 8) self.assertEqual(generation_options({"temperature": None, "top_p": None}, 8), - (8, 0.7, 0.9)) + (8, 0.7, 0.9, None)) + # response_format -> grammar plumbing (draft source, never a constraint) + opts = generation_options({"max_tokens": 4, "response_format": {"type": "json_object"}}, 8) + self.assertIn("root ::=", opts[3]) + schema = {"type": "object", "properties": {"a": {"type": "string"}}, "required": ["a"]} + opts = generation_options({"max_tokens": 4, "response_format": + {"type": "json_schema", "json_schema": {"schema": schema}}}, 8) + self.assertEqual(json.loads(opts[3]), schema) + opts = generation_options({"max_tokens": 4, "response_format": + {"type": "gbnf", "grammar": 'root ::= "x"'}}, 8) + self.assertEqual(opts[3], 'root ::= "x"') + with self.assertRaises(APIError): + generation_options({"response_format": {"type": "yaml"}}, 8) + with self.assertRaises(APIError): + generation_options({"response_format": {"type": "json_schema", "json_schema": {}}}, 8) + with self.assertRaises(APIError): # non-dict response_format + generation_options({"response_format": "json"}, 8) + with self.assertRaises(APIError): # empty gbnf + generation_options({"response_format": {"type": "gbnf", "grammar": " "}}, 8) + with self.assertRaises(APIError): # oversized grammar (> 1 MiB pre-check) + generation_options({"response_format": {"type": "gbnf", "grammar": "x" * ((1 << 20) + 1)}}, 8) + # malformed GBNF passes the gateway by design: the ENGINE fail-softs it + # (draft source only — bad grammar costs the speedup, never the request) + opts = generation_options({"response_format": {"type": "gbnf", "grammar": "not a grammar ::="}}, 8) + self.assertEqual(opts[3], "not a grammar ::=") class ProtocolTest(unittest.TestCase): diff --git a/c/tests/test_pipe_block.c b/c/tests/test_pipe_block.c new file mode 100644 index 0000000..8571695 --- /dev/null +++ b/c/tests/test_pipe_block.c @@ -0,0 +1,169 @@ +/* COLI_PIPE_BLOCK: the pipe pool's condvar waiter must be observably + * equivalent to the sched_yield spin it replaces — same bytes land in the + * same ws[] slots, and no interleaving loses a wakeup (the worker RELEASE- + * stores ready[] BEFORE taking mx to broadcast; the waiter re-checks under + * the lock, so a flag set between its fast-path check and the wait cannot + * be missed). Both waiters are exercised against the same on-disk fixture, + * alternating parked waits (wait issued before the load finishes) with + * fast-path waits (load already done), across enough generations to cycle + * the pool's gen-tagged cursor. + * + * Also pins the PIPE_WORKERS => PIPE implication table: fires ONLY when + * PIPE is unset in the env AND the platform default left the pipe off AND + * PIPE_WORKERS parses positive (PIPE_WORKERS=0/empty/negative must NOT + * silently enable a clamped 1-worker pipe). */ +#include +#include +#include +#include +#include +#include +#define main coli_glm_main_unused +#include "../colibri.c" +#undef main + +static int fail(const char *s){ fprintf(stderr,"FAIL: %s\n",s); return 1; } + +enum { NE=8, LAYER=1 }; /* experts 0..NE-1 on one MoE layer */ +/* per-expert file image: [gate 12][up 12][down 12][gate.qs 12][up.qs 12][down.qs 16] */ +enum { WB=12, QS_G=12, QS_U=12, QS_D=16, ESZ=3*WB+QS_G+QS_U+QS_D }; + +static unsigned char wbyte(int e,int j){ return (unsigned char)(e*31+j+1); } +static float scale(int e,int i){ return (float)(e*8+i)+0.5f; } + +#define TMPF "test_pipe_block.tmp" + +static int write_fixture(void){ + FILE *w=fopen(TMPF,"wb"); if(!w) return fail("create temp"); + for(int e=0;ec.hidden=4; m->c.moe_inter=3; m->ebits=8; + m->S.n=NE*6; m->S.cap=NE*6; m->S.t=calloc(NE*6,sizeof(st_tensor)); + if(!m->S.t) return fail("tensor metadata allocation"); + const char *proj[3]={"gate_proj","up_proj","down_proj"}; + int sbytes[3]={QS_G,QS_U,QS_D}; + for(int e=0;eS.t[e*6+k]=(st_tensor){strdup(name),fd,wo,WB,3,WB}; wo+=WB; + size_t n=strlen(name); memcpy(name+n,".qs",4); + m->S.t[e*6+3+k]=(st_tensor){strdup(name),fd,so,sbytes[k],2,sbytes[k]/4}; so+=sbytes[k]; + } + } + return 0; +} + +static int check_slot(ESlot *s,int e){ + if(s->eid!=e || s->g.fmt!=1 || s->u.fmt!=1 || s->d.fmt!=1){ + fprintf(stderr," slot: eid=%d (want %d) fmt g/u/d=%d/%d/%d (want 1/1/1)\n", + s->eid,e,s->g.fmt,s->u.fmt,s->d.fmt); + return 1; + } + const unsigned char *g=(const unsigned char*)s->g.q8, + *u=(const unsigned char*)s->u.q8, + *d=(const unsigned char*)s->d.q8; /* q8 is int8_t; compare raw bytes */ + for(int j=0;jg.s[i]!=scale(e,i) || s->u.s[i]!=scale(e,3+i)){ + fprintf(stderr," slot e=%d scale %d: g=%g/%g u=%g/%g (got/want)\n",e,i, + (double)s->g.s[i],(double)scale(e,i),(double)s->u.s[i],(double)scale(e,3+i)); + return 1; + } + for(int i=0;i<4;i++) + if(s->d.s[i]!=scale(e,6+i)){ + fprintf(stderr," slot e=%d scale %d: d=%g/%g (got/want)\n",e,i, + (double)s->d.s[i],(double)scale(e,6+i)); + return 1; + } + return 0; +} + +static int run_generations(Model *m,int block,int gens){ + g_pipe_block=block; + for(int gen=0;genws[q],eids[q])) return fail(block?"slot contents (block)":"slot contents (spin)"); + } + } + return 0; +} + +static int test_implication_table(void){ + struct { const char *pipe_env,*pw_env; int pipe_now,want; } T[]={ + {NULL,"4",0,1}, /* pool sized, pipe off, PIPE unset → imply */ + {NULL,"16",0,1}, + {NULL,"0",0,0}, /* PIPE_WORKERS=0 must NOT enable a clamped pipe */ + {NULL,"",0,0}, + {NULL,"-3",0,0}, + {"0","4",0,0}, /* explicit PIPE=0 always wins */ + {"1","4",1,0}, /* explicit PIPE=1: nothing to imply */ + {NULL,"4",1,0}, /* platform default already ON (win32) */ + {NULL,NULL,0,0}, + }; + for(size_t i=0;i",T[i].pw_env?T[i].pw_env:"",T[i].pipe_now); + return 1; + } + return 0; +} + +int main(void){ + if(test_implication_table()) return 1; + + /* Relative to the CWD, like test_compat_direct's TMPF — NOT "/tmp/...": + * the windows job builds native .exe files and "/tmp" is not a Windows + * path. fwrite then reopen read-only: Windows compat has pread, not pwrite. */ + if(write_fixture()) return 1; + int fd=open(TMPF,COMPAT_O_RDONLY); + if(fd<0) return fail("open temp"); + + static Model m; /* zeroed: buffered pread path, no mmap/cuda */ + if(build_fixture(&m,fd)){ close(fd); remove(TMPF); return 1; } + + g_pipe=1; g_pipe_nw=4; + pipe_init(&m); + + /* spin waiter first (control), then the condvar waiter under the same + * dispatch pattern; 200 generations each cycles the gen-tagged cursor + * and alternates parked/fast-path waits. */ + if(run_generations(&m,0,200) || run_generations(&m,1,200)){ close(fd); remove(TMPF); return 1; } + + for(int q=0;q 24 threads, 12 physical cores. + + # The parser must return 12 physical cores under BOTH lscpu layouts: + # - 2-col: `lscpu -p=core,socket` emits exactly [core,socket] (this is + # what the probe actually requests; the previous fields[1]/[2] + # indexing skipped every line here and fell through to the + # logical count -> the regression JustVugg caught). + # - 3-col: bare `lscpu -p` prepends a CPU column -> [cpu,core,socket]. + # Taking the last two fields is correct in both cases. + layouts = { + "2-col (-p=core,socket)": ( + "# core,socket\n" + + "\n".join(f"{core},0" for core in range(12) for _ in range(2))), + "3-col (bare -p, CPU prefix)": ( + "# CPU,Core,Socket\n" + + "\n".join(f"{cpu},{core},0" for core in range(12) for cpu in range(2))), + } + for label, blob in layouts.items(): + with mock.patch("resource_plan.subprocess.run", + return_value=lscpu(blob)), \ + mock.patch.object(sys, "platform", "linux"): + plan = build_plan(self.model, available_memory=16 * GB, + available_disk=1, gpus=[]) + env = environment_for_plan(plan) + self.assertEqual(plan["cpu"]["physical_cores"], 12, label) + self.assertEqual(env["OMP_NUM_THREADS"], "12", label) + + def test_plan_does_not_set_omp_affinity_vars(self): + # The real #325 regression: --auto-tier set OMP_PROC_BIND=spread + + # OMP_PLACES=cores, which ran before the engine's overwrite=0 setenv and + # so won, collapsing the OpenMP team to one CPU on the reporter's 64-core + # Linux box even though OMP_NUM_THREADS was correct. The plan must leave + # affinity to the engine's own hot-thread tuning (which prefers 'close'). + plan = build_plan(self.model, available_memory=16 * GB, + available_disk=1, gpus=[], physical_cpus=64) + env = environment_for_plan(plan) + self.assertEqual(env["OMP_NUM_THREADS"], "64") + self.assertNotIn("OMP_PROC_BIND", env) + self.assertNotIn("OMP_PLACES", env) + + def test_plan_conserves_budget_and_experts_above_256gb(self): + # Regression for #325's reporter: a 512 GB machine loading the whole + # model into RAM. Verify the budget math stays exact at large RAM sizes + # (no integer truncation, no over-allocation, no experts lost between + # tiers). Checked at 256/512/800 GB to bracket the reporter's box. + for ram_gb in (256, 512, 800): + plan = build_plan(self.model, ram_gb=ram_gb, available_disk=1, + gpus=[], physical_cpus=64) + ram = plan["tiers"]["ram"] + # RAM budget never over-allocated: dense + runtime + cache <= budget. + allocated = (ram["dense_bytes"] + ram["runtime_bytes"] + + ram["expert_cache_bytes"]) + self.assertLessEqual(allocated, ram["budget_bytes"], + f"over-allocated RAM at {ram_gb} GB") + # Every expert byte is accounted for exactly once across the tiers. + tiers = plan["tiers"] + tiered = (tiers["vram"]["hot_expert_bytes"] + + ram["warm_expert_bytes"] + + tiers["disk"]["cold_expert_bytes"]) + self.assertEqual(tiered, plan["model"]["expert_bytes"], + f"expert bytes lost/duplicated at {ram_gb} GB") + # A positive RAM budget yields a non-negative cache and a sensible cap. + self.assertGreaterEqual(ram["expert_cache_bytes"], 0) + self.assertGreaterEqual(ram["cache_slots_per_layer"], 0) + def test_filters_requested_devices(self): gpus = [{"index": 0, "name": "a", "total_bytes": 8 * GB, "free_bytes": 8 * GB}] plan = build_plan(self.model, available_memory=16 * GB, available_disk=1, @@ -117,13 +192,13 @@ class ResourcePlanTest(unittest.TestCase): self.assertEqual(env["COLI_CUDA"], "1") self.assertEqual(env["COLI_GPUS"], "1") self.assertEqual(env["OMP_NUM_THREADS"], str(plan["cpu"]["physical_cores"])) - if sys.platform == "win32": - # MinGW libgomp: niente affinity su Windows, le chiavi non vanno emesse - self.assertNotIn("OMP_PROC_BIND", env) - self.assertNotIn("OMP_PLACES", env) - else: - self.assertEqual(env["OMP_PROC_BIND"], "spread") - self.assertEqual(env["OMP_PLACES"], "cores") + # The plan must NOT set OMP_PROC_BIND / OMP_PLACES on any platform: + # the engine's own hot-thread tuning owns affinity (it prefers + # OMP_PROC_BIND=close for the back-to-back per-expert matmuls). Setting + # spread + cores here ran before the engine's overwrite=0 setenv and so + # won, collapsing the team to one CPU on some libgomp topologies (#325). + self.assertNotIn("OMP_PROC_BIND", env) + self.assertNotIn("OMP_PLACES", env) self.assertEqual(env["PIN_GB"], env["CUDA_EXPERT_GB"]) explicit_threads = environment_for_plan(plan, {"OMP_NUM_THREADS": "7", @@ -140,6 +215,81 @@ class ResourcePlanTest(unittest.TestCase): plan = build_plan(self.model, available_memory=16 * GB, available_disk=1, gpus=[], physical_cpus=8, cpu_sockets=1) self.assertNotIn("COLI_NUMA", environment_for_plan(plan)) + + def test_auto_tune_mtp_off_when_compute_bound(self): + # Tiny model with 64 GB RAM and no GPU: all experts fit in RAM with no + # warm tier, so the plan classifies as compute-bound. + plan = build_plan(self.model, ram_gb=64, available_memory=64 * GB, + available_disk=100 * GB, gpus=[], physical_cpus=24, + cpu_sockets=2) + # With such a small model fully in RAM and no GPU, bottleneck is compute + self.assertEqual(plan["bottleneck_class"], "compute") + self.assertIn("DRAFT", plan["tune"]) + self.assertEqual(plan["tune"]["DRAFT"]["value"], "0") + env = environment_for_plan(plan) + self.assertEqual(env["DRAFT"], "0") + explicit = environment_for_plan(plan, {"DRAFT": "3"}) + self.assertEqual(explicit["DRAFT"], "3") + + def test_auto_tune_mtp_off_when_disk_low_hit(self): + # Use a model large enough that 8 GB RAM can't hold all experts. + big = tempfile.TemporaryDirectory() + bigmodel = Path(big.name) + (bigmodel / "config.json").write_text(json.dumps({ + "num_hidden_layers": 2, "n_routed_experts": 4, + "kv_lora_rank": 4, "qk_rope_head_dim": 2, + "qk_nope_head_dim": 3, "v_head_dim": 5, "num_attention_heads": 2, + })) + expert_size = 3 * GB # each expert 3 GB → 12 GB total, won't fit in 8 GB budget + write_shard(bigmodel / "out-00000.safetensors", [ + ("model.embed_tokens.weight", 100), + ("model.layers.0.self_attn.q_a_proj.weight", 200), + ]) + for i in range(4): + write_shard(bigmodel / f"out-{i+1:05d}.safetensors", [ + (f"model.layers.1.mlp.experts.{i}.gate_proj.weight", expert_size), + ]) + plan = build_plan(bigmodel, ram_gb=0, available_memory=4 * GB, + available_disk=100 * GB, gpus=[], physical_cpus=8, + cpu_sockets=1) + big.cleanup() + self.assertEqual(plan["bottleneck_class"], "disk") + self.assertLess(plan["projected_hit_rate"], 0.90) + self.assertEqual(plan["tune"]["DRAFT"]["value"], "0") + + def test_auto_tune_pipe_multi_gpu(self): + gpus = [ + {"index": 0, "name": "a", "total_bytes": 32 * GB, "free_bytes": 30 * GB}, + {"index": 1, "name": "b", "total_bytes": 32 * GB, "free_bytes": 30 * GB}, + ] + plan = build_plan(self.model, ram_gb=16, available_memory=32 * GB, + available_disk=1, gpus=gpus, cpu_sockets=2) + self.assertEqual(plan["tune"]["COLI_CUDA_PIPE"]["value"], "2") + env = environment_for_plan(plan) + self.assertEqual(env["COLI_CUDA_PIPE"], "2") + + def test_auto_tune_pipe_single_gpu(self): + gpus = [{"index": 0, "name": "a", "total_bytes": 12 * GB, "free_bytes": 10 * GB}] + plan = build_plan(self.model, ram_gb=16, available_memory=32 * GB, + available_disk=1, gpus=gpus, cpu_sockets=1) + self.assertEqual(plan["tune"]["COLI_CUDA_PIPE"]["value"], "1") + + def test_auto_tune_numa_hint_for_cpu_only(self): + plan = build_plan(self.model, ram_gb=64, available_memory=64 * GB, + available_disk=1, gpus=[], physical_cpus=64, cpu_sockets=2) + self.assertIn("_numa_hint", plan["tune"]) + self.assertIn("numactl", plan["tune"]["_numa_hint"]) + self.assertIn("auto-tune", format_plan(plan)) + + def test_format_plan_shows_tune_and_hit_rate(self): + plan = build_plan(self.model, ram_gb=64, available_memory=64 * GB, + available_disk=100 * GB, gpus=[], physical_cpus=24, + cpu_sockets=1) + text = format_plan(plan) + self.assertIn("hit", text) + self.assertIn("auto-tune", text) + self.assertIn("DRAFT", text) + def test_cpu_binary_does_not_apply_gpu_tier(self): plan = build_plan(self.model, available_memory=16 * GB, available_disk=1, gpus=[{"index": 0, "name": "a", "total_bytes": 8 * GB, @@ -178,5 +328,73 @@ class ResourcePlanTest(unittest.TestCase): self.assertIn("expected_bottleneck", plan) +class PhysicalCpuCountTest(unittest.TestCase): + """Regression for #325: --auto-tier pinned decode to one core because + physical_cpu_count() silently returned 1. + + Two root causes this locks down: + 1. lscpu -p prepends a CPU column, so `-p=core,socket` emits + CPU,Core,Socket; counting rows counted logical SMT siblings. + 2. any probe failure fell through to ``os.cpu_count() or 1`` and the + ``or 1`` could pin a constrained/cgroup'd box to a single core. + """ + + def _lscpu(self, stdout): + return subprocess.CompletedProcess(args=[], returncode=0, + stdout=stdout, stderr="") + + def _lscpu_topology(self, sockets, cores_per_socket, threads_per_core): + # Real lscpu shape: socket-local core IDs repeat across sockets; the + # CPU column (always prepended) is a unique logical-CPU index. + rows, cpu = [], 0 + for sock in range(sockets): + for core in range(cores_per_socket): + for _ in range(threads_per_core): + rows.append(f"{cpu},{core},{sock}") + cpu += 1 + return "# CPU,Core,Socket\n" + "\n".join(rows) + + def test_counts_physical_cores_not_smt_threads(self): + blob = self._lscpu_topology(sockets=2, cores_per_socket=16, threads_per_core=2) + with mock.patch("resource_plan.subprocess.run", return_value=self._lscpu(blob)), \ + mock.patch.object(sys, "platform", "linux"): + self.assertEqual(physical_cpu_count(), 32) + + def test_single_socket_no_smt(self): + blob = self._lscpu_topology(sockets=1, cores_per_socket=8, threads_per_core=1) + with mock.patch("resource_plan.subprocess.run", return_value=self._lscpu(blob)), \ + mock.patch.object(sys, "platform", "linux"): + self.assertEqual(physical_cpu_count(), 8) + + def test_skips_offline_core_socket_fields(self): + # VMs / large NUMA boxes emit "-" for offline core or socket IDs; that + # used to raise ValueError, discard the whole parse, and fall through + # to the single-core fallback. + blob = "# CPU,Core,Socket\n0,0,0\n1,-,0\n2,1,0\n3,1,0\n" + with mock.patch("resource_plan.subprocess.run", return_value=self._lscpu(blob)), \ + mock.patch.object(sys, "platform", "linux"): + self.assertEqual(physical_cpu_count(), 2) + + def test_lscpu_missing_falls_back_to_logical_not_silent_one(self): + # The bug: lscpu absent -> os.cpu_count() or 1. On a constrained box + # os.cpu_count() can be 1. We still must never silently pick 1 without + # a warning, and when logical cores exist they must be used. + import os + with mock.patch("resource_plan.subprocess.run", side_effect=FileNotFoundError), \ + mock.patch.object(sys, "platform", "linux"), \ + mock.patch("resource_plan.os.cpu_count", return_value=16), \ + mock.patch("sys.stderr"): + self.assertEqual(physical_cpu_count(), 16) + + def test_zero_logical_cores_warns_and_returns_one(self): + # The genuine degenerate case: no probe works and os.cpu_count() is + # None/1. Must return 1 (engine needs a positive team size) but warn. + with mock.patch("resource_plan.subprocess.run", side_effect=FileNotFoundError), \ + mock.patch.object(sys, "platform", "linux"), \ + mock.patch("resource_plan.os.cpu_count", return_value=None), \ + mock.patch("sys.stderr"): + self.assertEqual(physical_cpu_count(), 1) + + if __name__ == "__main__": unittest.main() diff --git a/c/tests/test_router_cuda.cu b/c/tests/test_router_cuda.cu new file mode 100644 index 0000000..4e714a6 --- /dev/null +++ b/c/tests/test_router_cuda.cu @@ -0,0 +1,79 @@ +/* Device-router kernel oracle (#431 PR-A). + * + * Feeds random activations/router weights through pipe_router_logits + + * pipe_router_select and checks against a CPU reference that replicates + * moe()'s plain routing path verbatim (sigmoid -> bias-augmented top-K by + * `choice`, weights from raw `logit`, route-level TOPP truncation, norm_topk, + * routed_scale). The dot/expf rounding may differ from libm at ~1e-6 rel, so + * a handful of near-tie index flips across trials is tolerated; the weight + * math itself must agree to 1e-4 rel on matching selections. + * + * Build: nvcc -O2 -std=c++17 -arch=native tests/test_router_cuda.cu -o tests/test_router_cuda + */ +#include +#include +#include +#include +#include + +/* pull in the kernel definitions (same idiom as the CPU tests' #include "../colibri.c") */ +#include "../backend_cuda.cu" + +static void cpu_ref(const float *x,const float *W,const float *bias,int D,int E, + int Ksel,float topp,int norm_topk,float rscale, + int *idx,float *w,int *keff){ + float *logit=(float*)malloc(E*sizeof(float)),*choice=(float*)malloc(E*sizeof(float)); + for(int e=0;ebv){bv=choice[e];best=e;} } + idx[kk]=best; w[kk]=logit[best]; } + int Ke=Ksel; + if(topp>0.f && topp<1.f){ + for(int a=1;a=0 && w[b]=topp*tot){ Ke=kk+1; break; } } } + if(norm_topk){ float sm=0; for(int kk=0;kk>>(x,W,b,D,lg,ch); + pipe_router_select<<<1,1>>>(lg,ch,E,K,topp,nt,rs,out); + char pack[K*8+4]; + if(cudaMemcpy(pack,out,sizeof(pack),cudaMemcpyDeviceToHost)!=cudaSuccess){ + printf("FAIL cuda\n"); return 1; } + int gidx[K],gkeff; float gw[K]; + memcpy(gidx,pack,K*4); memcpy(gw,pack+K*4,K*4); memcpy(&gkeff,pack+K*8,4); + int ridx[K],rkeff; float rw[K]; + cpu_ref(x,W,b,D,E,K,topp,nt,rs,ridx,rw,&rkeff); + int mism=0; for(int k2=0;k21e-4f*(fabsf(ref)+1e-6f)+1e-6f){ bad++; break; } + } + } + printf("router oracle: %d trials, %d near-tie flips, %d weight mismatches\n",TRIALS,flips,bad); + if(flips>4||bad){ printf("FAIL\n"); return 1; } + printf("OK\n"); return 0; +} diff --git a/c/tests/test_st_pread b/c/tests/test_st_pread new file mode 100644 index 0000000..8cf2710 Binary files /dev/null and b/c/tests/test_st_pread differ diff --git a/c/tests/test_tok_o200k.c b/c/tests/test_tok_o200k.c new file mode 100644 index 0000000..fe62890 --- /dev/null +++ b/c/tests/test_tok_o200k.c @@ -0,0 +1,64 @@ +/* o200k pre-tokenizer validation against HF-tokenizers-generated expectations. + * Self-contained for the test-c harness: loads tests/tok_o200k_tiny.json (a + * synthetic byte-level BPE whose Split regex is the o200k pattern — a few KB, + * no model download) and scores tests/tok_o200k_cases.txt, whose expected ids + * were produced by HF `tokenizers` on the same file. Guards the case-aware + * letter matcher, contractions, digit groups, the [\r\n/]* punctuation tail, + * whitespace branches, and added-token atomicity; round-trips every case. + * The cl100k path is untouched by construction (dispatch requires \p{Lu} in + * the tokenizer's own Split pattern) and stays covered by the GLM oracle. */ +#define _GNU_SOURCE +#include "../tok.h" + +int main(void) { + Tok T; + tok_load(&T, "tests/tok_o200k_tiny.json"); + if (!T.o200k) { fprintf(stderr, "test_tok_o200k: o200k pattern not detected\n"); return 1; } + FILE *f = fopen("tests/tok_o200k_cases.txt", "rb"); + if (!f) { perror("tests/tok_o200k_cases.txt"); return 1; } + /* fgets, not getline: MinGW's UCRT lacks getline and this must run on + * the windows job. Case lines are short; 8 KB is generous. */ + char line[8192]; + int pass = 0, tot = 0, dpass = 0; + while (fgets(line, sizeof(line), f)) { + size_t nr = strlen(line); + while (nr > 0 && (line[nr-1] == '\n' || line[nr-1] == '\r')) line[--nr] = 0; + if (nr == 0) continue; + char *tab = strchr(line, '\t'); if (!tab) continue; + *tab = 0; + const char *text = line, *idstr = tab + 1; + char tbuf[4096]; int tn = 0; + for (const char *q = text; *q && tn < 4095; q++) { + if (q[0]=='\\' && q[1]=='n') { tbuf[tn++]='\n'; q++; } + else if (q[0]=='\\' && q[1]=='t') { tbuf[tn++]='\t'; q++; } + else if (q[0]=='\\' && q[1]=='r') { tbuf[tn++]='\r'; q++; } + else if (q[0]=='\\' && q[1]=='\\') { tbuf[tn++]='\\'; q++; } + else tbuf[tn++] = *q; + } + tbuf[tn] = 0; + int exp[512], ne = 0; + for (const char *q = idstr; *q; ) { + while (*q == ',' || *q == ' ') q++; + if (!*q) break; + exp[ne++] = atoi(q); + while (*q && *q != ',') q++; + } + int got[512]; int ng = tok_encode(&T, tbuf, tn, got, 512); + int ok = (ng == ne); + for (int i = 0; i < ng && ok; i++) ok = (got[i] == exp[i]); + tot++; if (ok) pass++; + char dec[8192]; int dn = tok_decode(&T, got, ng, dec, 8191); + int drt = (dn == tn) && !memcmp(dec, tbuf, tn); + if (drt) dpass++; + if (!ok || !drt) { + fprintf(stderr, "MISMATCH text=%s\n exp(%d):", text, ne); + for (int i = 0; i < ne; i++) fprintf(stderr, " %d", exp[i]); + fprintf(stderr, "\n got(%d):", ng); + for (int i = 0; i < ng; i++) fprintf(stderr, " %d", got[i]); + fprintf(stderr, "\n decode_ok=%d\n", drt); + } + } + fclose(f); + printf("test_tok_o200k: ENCODE %d/%d DECODE %d/%d\n", pass, tot, dpass, tot); + return (pass == tot && dpass == tot) ? 0 : 2; +} diff --git a/c/tests/tok_o200k_cases.txt b/c/tests/tok_o200k_cases.txt new file mode 100644 index 0000000..38e37fa --- /dev/null +++ b/c/tests/tok_o200k_cases.txt @@ -0,0 +1,40 @@ +hello world 259,32,119,111,114,263 +HelloWorld 72,101,257,111,262 +XMLHttpRequest 88,77,76,72,116,116,112,82,101,113,117,101,115,116 +helloWORLDhello 259,87,79,82,76,68,259 +dog's 100,111,103,270 +DON'T 68,79,78,39,84 +don't 100,111,110,39,116 +I'll've 73,39,257,39,118,101 +O'Brien 79,39,66,114,105,101,110 +the theatre 265,32,116,256,97,116,114,101 +12345 268,52,53 +a1b22c333d4444 97,49,98,50,50,99,51,51,51,100,52,52,52,52 +3.14 51,46,49,52 +http://x.com/a/b 104,116,116,112,58,47,47,120,46,99,111,109,47,97,47,98 +path/to/file 112,97,264,47,116,111,47,102,105,108,101 +a//b///c 97,47,47,98,47,47,47,99 +!!\r\n//x 33,33,13,10,47,47,120 +one\ntwo\r\nthree 111,110,101,10,116,119,111,13,10,264,114,101,101 + \n x 32,32,10,32,32,120 + 32,32,32 +a b c 97,32,32,98,32,32,32,99 +tab\there 116,269,9,256,114,101 +Café 67,97,102,101,204,129 +naiveBayes 110,97,105,118,101,66,97,121,101,115 +Éclair 195,137,99,108,97,105,114 +北京大学 229,140,151,228,186,172,229,164,167,229,173,166 +ΑΒαβ 206,145,206,146,206,177,206,178 +Иван 208,152,208,178,208,176,208,189 +ẞßscharf 225,186,158,195,159,115,99,104,97,114,102 +i̇stanbul 105,204,135,115,116,97,110,98,117,108 +hello<|endoftext|>world 259,274,119,111,114,263 +<|message_user|>hi 275,104,105 +mixedCASEand123 109,105,120,101,100,67,65,83,69,97,110,100,268 +'s 270 + 's 32,39,115 +A 65 +aB 97,66 +Ab 65,98 +AB 65,66 +ab 269 diff --git a/c/tests/tok_o200k_tiny.json b/c/tests/tok_o200k_tiny.json new file mode 100644 index 0000000..829301e --- /dev/null +++ b/c/tests/tok_o200k_tiny.json @@ -0,0 +1 @@ +{"version": "1.0", "truncation": null, "padding": null, "added_tokens": [{"id": 274, "content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": false, "special": true}, {"id": 275, "content": "<|message_user|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": false, "special": true}], "normalizer": null, "pre_tokenizer": {"type": "Sequence", "pretokenizers": [{"type": "Split", "pattern": {"Regex": "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"}, "behavior": "Isolated", "invert": false}, {"type": "ByteLevel", "add_prefix_space": false, "trim_offsets": true, "use_regex": false}]}, "post_processor": null, "decoder": {"type": "ByteLevel", "add_prefix_space": true, "trim_offsets": true, "use_regex": true}, "model": {"type": "BPE", "dropout": null, "unk_token": null, "continuing_subword_prefix": null, "end_of_word_suffix": null, "fuse_unk": false, "byte_fallback": false, "ignore_merges": true, "vocab": {"Ā": 0, "ā": 1, "Ă": 2, "ă": 3, "Ą": 4, "ą": 5, "Ć": 6, "ć": 7, "Ĉ": 8, "ĉ": 9, "Ċ": 10, "ċ": 11, "Č": 12, "č": 13, "Ď": 14, "ď": 15, "Đ": 16, "đ": 17, "Ē": 18, "ē": 19, "Ĕ": 20, "ĕ": 21, "Ė": 22, "ė": 23, "Ę": 24, "ę": 25, "Ě": 26, "ě": 27, "Ĝ": 28, "ĝ": 29, "Ğ": 30, "ğ": 31, "Ġ": 32, "!": 33, "\"": 34, "#": 35, "$": 36, "%": 37, "&": 38, "'": 39, "(": 40, ")": 41, "*": 42, "+": 43, ",": 44, "-": 45, ".": 46, "/": 47, "0": 48, "1": 49, "2": 50, "3": 51, "4": 52, "5": 53, "6": 54, "7": 55, "8": 56, "9": 57, ":": 58, ";": 59, "<": 60, "=": 61, ">": 62, "?": 63, "@": 64, "A": 65, "B": 66, "C": 67, "D": 68, "E": 69, "F": 70, "G": 71, "H": 72, "I": 73, "J": 74, "K": 75, "L": 76, "M": 77, "N": 78, "O": 79, "P": 80, "Q": 81, "R": 82, "S": 83, "T": 84, "U": 85, "V": 86, "W": 87, "X": 88, "Y": 89, "Z": 90, "[": 91, "\\": 92, "]": 93, "^": 94, "_": 95, "`": 96, "a": 97, "b": 98, "c": 99, "d": 100, "e": 101, "f": 102, "g": 103, "h": 104, "i": 105, "j": 106, "k": 107, "l": 108, "m": 109, "n": 110, "o": 111, "p": 112, "q": 113, "r": 114, "s": 115, "t": 116, "u": 117, "v": 118, "w": 119, "x": 120, "y": 121, "z": 122, "{": 123, "|": 124, "}": 125, "~": 126, "ġ": 127, "Ģ": 128, "ģ": 129, "Ĥ": 130, "ĥ": 131, "Ħ": 132, "ħ": 133, "Ĩ": 134, "ĩ": 135, "Ī": 136, "ī": 137, "Ĭ": 138, "ĭ": 139, "Į": 140, "į": 141, "İ": 142, "ı": 143, "IJ": 144, "ij": 145, "Ĵ": 146, "ĵ": 147, "Ķ": 148, "ķ": 149, "ĸ": 150, "Ĺ": 151, "ĺ": 152, "Ļ": 153, "ļ": 154, "Ľ": 155, "ľ": 156, "Ŀ": 157, "ŀ": 158, "Ł": 159, "ł": 160, "¡": 161, "¢": 162, "£": 163, "¤": 164, "¥": 165, "¦": 166, "§": 167, "¨": 168, "©": 169, "ª": 170, "«": 171, "¬": 172, "Ń": 173, "®": 174, "¯": 175, "°": 176, "±": 177, "²": 178, "³": 179, "´": 180, "µ": 181, "¶": 182, "·": 183, "¸": 184, "¹": 185, "º": 186, "»": 187, "¼": 188, "½": 189, "¾": 190, "¿": 191, "À": 192, "Á": 193, "Â": 194, "Ã": 195, "Ä": 196, "Å": 197, "Æ": 198, "Ç": 199, "È": 200, "É": 201, "Ê": 202, "Ë": 203, "Ì": 204, "Í": 205, "Î": 206, "Ï": 207, "Ð": 208, "Ñ": 209, "Ò": 210, "Ó": 211, "Ô": 212, "Õ": 213, "Ö": 214, "×": 215, "Ø": 216, "Ù": 217, "Ú": 218, "Û": 219, "Ü": 220, "Ý": 221, "Þ": 222, "ß": 223, "à": 224, "á": 225, "â": 226, "ã": 227, "ä": 228, "å": 229, "æ": 230, "ç": 231, "è": 232, "é": 233, "ê": 234, "ë": 235, "ì": 236, "í": 237, "î": 238, "ï": 239, "ð": 240, "ñ": 241, "ò": 242, "ó": 243, "ô": 244, "õ": 245, "ö": 246, "÷": 247, "ø": 248, "ù": 249, "ú": 250, "û": 251, "ü": 252, "ý": 253, "þ": 254, "ÿ": 255, "he": 256, "ll": 257, "hell": 258, "hello": 259, "Wo": 260, "Wor": 261, "World": 262, "ld": 263, "th": 264, "the": 265, "Ġthe": 266, "12": 267, "123": 268, "ab": 269, "'s": 270, "Ġa": 271, "./": 272, "ĊĊ": 273}, "merges": [["h", "e"], ["l", "l"], ["he", "ll"], ["hell", "o"], ["W", "o"], ["Wo", "r"], ["Wor", "ld"], ["l", "d"], ["t", "h"], ["th", "e"], ["Ġ", "the"], ["1", "2"], ["12", "3"], ["a", "b"], ["'", "s"], ["Ġ", "a"], [".", "/"], ["Ċ", "Ċ"]]}} \ No newline at end of file diff --git a/c/tok.h b/c/tok.h index 7d1140b..6b957df 100644 --- a/c/tok.h +++ b/c/tok.h @@ -19,6 +19,7 @@ #include #include "json.h" #include "tok_unicode.h" +#include "tok_unicode_o200k.h" /* ---------- hash map (chiavi binarie con lunghezza) ---------- */ typedef struct { const char *k; int klen; int v; int used; } ment; @@ -50,6 +51,7 @@ typedef struct { Special *sp; int nsp; /* added tokens, ordinati per lunghezza decrescente */ uint32_t byte2cp[256]; int byte2cp_len[256]; char byte2str[256][3]; int16_t cp2byte[1024]; + int o200k; /* pre_tokenizer regex family: 0 = cl100k (GLM), 1 = o200k (Inkling) */ } Tok; /* ---------- UTF-8 ---------- */ @@ -144,6 +146,17 @@ static void tok_load(Tok *T, const char *path){ } qsort(T->sp,T->nsp,sizeof(Special),cmp_sp_len); /* match piu' lungo per primo */ } + /* pre_tokenizer family: the o200k Split regex is recognizable by its + * case-category classes (\p{Lu}...) which cl100k does not use */ + jval *pt=json_get(root,"pre_tokenizer"); + if(pt){ + jval *ps=json_get(pt,"pretokenizers"); + if(ps&&ps->t==J_ARR) for(int i=0;ilen;i++){ + jval *pat=json_get(ps->kids[i],"pattern"); + jval *rx=pat?json_get(pat,"Regex"):NULL; + if(rx&&rx->t==J_STR&&strstr(rx->str,"\\p{Lu}")) T->o200k=1; + } + } /* arena/buf restano allocati: le stringhe (j_dup) sono malloc indipendenti e ci servono vive */ (void)arena; } @@ -241,6 +254,104 @@ static void pretok_chunk(Tok *T, const unsigned char *p, int a, int b, int *out, free(cp); free(off); } +/* ---------- pre-tokenizer o200k (Inkling / GPT-4o family) ---------- + * Split regex: + * A: [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)? + * B: [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)? + * C: \p{N}{1,3} D: ' ?[^\s\p{L}\p{N}]+[\r\n/]*' E: \s*[\r\n]+ F: \s+(?!\S) G: \s+ + * S1 = Lu|Lt|Lm|Lo|M, S2 = Ll|Lm|Lo|M. The letter matcher below replays the + * regex engine's backtracking order exactly: A with greedy optional prefix and + * maximally-greedy S1* given back until S2+ can take >=1 char, then B. */ +#define O2_S1(c) (is_U(c)||is_X(c)) +#define O2_S2(c) (is_X(c)||(is_L(c)&&!is_U(c))) +static uint32_t o2_low(uint32_t c){ return (c>='A'&&c<='Z')?c+32:c; } +static int o2_contraction(const uint32_t *cp, int n, int k){ + if(k=0; pfx--){ + int j0=i; + if(pfx){ + uint32_t c=cp[i]; + if(c=='\r'||c=='\n'||is_L(c)||is_N(c)||i+1>=n) continue; + j0=i+1; + } + int m1=j0; while(m1=j0; s--){ + if(s=0; pfx--){ + int j0=i; + if(pfx){ + uint32_t c=cp[i]; + if(c=='\r'||c=='\n'||is_L(c)||is_N(c)||i+1>=n) continue; + j0=i+1; + } + int m1=j0; while(m1j0){ + int k=m1; while(ki){ i=e; bpe_piece(T,p,off[start],off[i],out,no,max); continue; } + } + /* C: \p{N}{1,3} */ + if(is_N(c)){ int j=i,k=0; while(ji){ int last=-1; for(int j=i;j=0){ i=last+1; bpe_piece(T,p,off[start],off[i],out,no,max); continue; } + int end = (r id (split sugli added token, poi pretok+BPE) ---------- */ static int tok_encode(Tok *T, const char *text, int len, int *out, int max){ const unsigned char *p=(const unsigned char*)text; int no=0; int i=0; @@ -254,7 +365,10 @@ static int tok_encode(Tok *T, const char *text, int len, int *out, int max){ } } int chunk_end = (hitpos<0) ? len : hitpos; - if(chunk_end>i) pretok_chunk(T,p,i,chunk_end,out,&no,max); + if(chunk_end>i){ + if(T->o200k) pretok_chunk_o200k(T,p,i,chunk_end,out,&no,max); + else pretok_chunk(T,p,i,chunk_end,out,&no,max); + } if(hitpos<0) break; if(no + +static const uint32_t uni_U[][2] = { + {0x41,0x5A},{0xC0,0xD6},{0xD8,0xDE},{0x100,0x100},{0x102,0x102},{0x104,0x104}, + {0x106,0x106},{0x108,0x108},{0x10A,0x10A},{0x10C,0x10C},{0x10E,0x10E},{0x110,0x110}, + {0x112,0x112},{0x114,0x114},{0x116,0x116},{0x118,0x118},{0x11A,0x11A},{0x11C,0x11C}, + {0x11E,0x11E},{0x120,0x120},{0x122,0x122},{0x124,0x124},{0x126,0x126},{0x128,0x128}, + {0x12A,0x12A},{0x12C,0x12C},{0x12E,0x12E},{0x130,0x130},{0x132,0x132},{0x134,0x134}, + {0x136,0x136},{0x139,0x139},{0x13B,0x13B},{0x13D,0x13D},{0x13F,0x13F},{0x141,0x141}, + {0x143,0x143},{0x145,0x145},{0x147,0x147},{0x14A,0x14A},{0x14C,0x14C},{0x14E,0x14E}, + {0x150,0x150},{0x152,0x152},{0x154,0x154},{0x156,0x156},{0x158,0x158},{0x15A,0x15A}, + {0x15C,0x15C},{0x15E,0x15E},{0x160,0x160},{0x162,0x162},{0x164,0x164},{0x166,0x166}, + {0x168,0x168},{0x16A,0x16A},{0x16C,0x16C},{0x16E,0x16E},{0x170,0x170},{0x172,0x172}, + {0x174,0x174},{0x176,0x176},{0x178,0x179},{0x17B,0x17B},{0x17D,0x17D},{0x181,0x182}, + {0x184,0x184},{0x186,0x187},{0x189,0x18B},{0x18E,0x191},{0x193,0x194},{0x196,0x198}, + {0x19C,0x19D},{0x19F,0x1A0},{0x1A2,0x1A2},{0x1A4,0x1A4},{0x1A6,0x1A7},{0x1A9,0x1A9}, + {0x1AC,0x1AC},{0x1AE,0x1AF},{0x1B1,0x1B3},{0x1B5,0x1B5},{0x1B7,0x1B8},{0x1BC,0x1BC}, + {0x1C4,0x1C5},{0x1C7,0x1C8},{0x1CA,0x1CB},{0x1CD,0x1CD},{0x1CF,0x1CF},{0x1D1,0x1D1}, + {0x1D3,0x1D3},{0x1D5,0x1D5},{0x1D7,0x1D7},{0x1D9,0x1D9},{0x1DB,0x1DB},{0x1DE,0x1DE}, + {0x1E0,0x1E0},{0x1E2,0x1E2},{0x1E4,0x1E4},{0x1E6,0x1E6},{0x1E8,0x1E8},{0x1EA,0x1EA}, + {0x1EC,0x1EC},{0x1EE,0x1EE},{0x1F1,0x1F2},{0x1F4,0x1F4},{0x1F6,0x1F8},{0x1FA,0x1FA}, + {0x1FC,0x1FC},{0x1FE,0x1FE},{0x200,0x200},{0x202,0x202},{0x204,0x204},{0x206,0x206}, + {0x208,0x208},{0x20A,0x20A},{0x20C,0x20C},{0x20E,0x20E},{0x210,0x210},{0x212,0x212}, + {0x214,0x214},{0x216,0x216},{0x218,0x218},{0x21A,0x21A},{0x21C,0x21C},{0x21E,0x21E}, + {0x220,0x220},{0x222,0x222},{0x224,0x224},{0x226,0x226},{0x228,0x228},{0x22A,0x22A}, + {0x22C,0x22C},{0x22E,0x22E},{0x230,0x230},{0x232,0x232},{0x23A,0x23B},{0x23D,0x23E}, + {0x241,0x241},{0x243,0x246},{0x248,0x248},{0x24A,0x24A},{0x24C,0x24C},{0x24E,0x24E}, + {0x370,0x370},{0x372,0x372},{0x376,0x376},{0x37F,0x37F},{0x386,0x386},{0x388,0x38A}, + {0x38C,0x38C},{0x38E,0x38F},{0x391,0x3A1},{0x3A3,0x3AB},{0x3CF,0x3CF},{0x3D2,0x3D4}, + {0x3D8,0x3D8},{0x3DA,0x3DA},{0x3DC,0x3DC},{0x3DE,0x3DE},{0x3E0,0x3E0},{0x3E2,0x3E2}, + {0x3E4,0x3E4},{0x3E6,0x3E6},{0x3E8,0x3E8},{0x3EA,0x3EA},{0x3EC,0x3EC},{0x3EE,0x3EE}, + {0x3F4,0x3F4},{0x3F7,0x3F7},{0x3F9,0x3FA},{0x3FD,0x42F},{0x460,0x460},{0x462,0x462}, + {0x464,0x464},{0x466,0x466},{0x468,0x468},{0x46A,0x46A},{0x46C,0x46C},{0x46E,0x46E}, + {0x470,0x470},{0x472,0x472},{0x474,0x474},{0x476,0x476},{0x478,0x478},{0x47A,0x47A}, + {0x47C,0x47C},{0x47E,0x47E},{0x480,0x480},{0x48A,0x48A},{0x48C,0x48C},{0x48E,0x48E}, + {0x490,0x490},{0x492,0x492},{0x494,0x494},{0x496,0x496},{0x498,0x498},{0x49A,0x49A}, + {0x49C,0x49C},{0x49E,0x49E},{0x4A0,0x4A0},{0x4A2,0x4A2},{0x4A4,0x4A4},{0x4A6,0x4A6}, + {0x4A8,0x4A8},{0x4AA,0x4AA},{0x4AC,0x4AC},{0x4AE,0x4AE},{0x4B0,0x4B0},{0x4B2,0x4B2}, + {0x4B4,0x4B4},{0x4B6,0x4B6},{0x4B8,0x4B8},{0x4BA,0x4BA},{0x4BC,0x4BC},{0x4BE,0x4BE}, + {0x4C0,0x4C1},{0x4C3,0x4C3},{0x4C5,0x4C5},{0x4C7,0x4C7},{0x4C9,0x4C9},{0x4CB,0x4CB}, + {0x4CD,0x4CD},{0x4D0,0x4D0},{0x4D2,0x4D2},{0x4D4,0x4D4},{0x4D6,0x4D6},{0x4D8,0x4D8}, + {0x4DA,0x4DA},{0x4DC,0x4DC},{0x4DE,0x4DE},{0x4E0,0x4E0},{0x4E2,0x4E2},{0x4E4,0x4E4}, + {0x4E6,0x4E6},{0x4E8,0x4E8},{0x4EA,0x4EA},{0x4EC,0x4EC},{0x4EE,0x4EE},{0x4F0,0x4F0}, + {0x4F2,0x4F2},{0x4F4,0x4F4},{0x4F6,0x4F6},{0x4F8,0x4F8},{0x4FA,0x4FA},{0x4FC,0x4FC}, + {0x4FE,0x4FE},{0x500,0x500},{0x502,0x502},{0x504,0x504},{0x506,0x506},{0x508,0x508}, + {0x50A,0x50A},{0x50C,0x50C},{0x50E,0x50E},{0x510,0x510},{0x512,0x512},{0x514,0x514}, + {0x516,0x516},{0x518,0x518},{0x51A,0x51A},{0x51C,0x51C},{0x51E,0x51E},{0x520,0x520}, + {0x522,0x522},{0x524,0x524},{0x526,0x526},{0x528,0x528},{0x52A,0x52A},{0x52C,0x52C}, + {0x52E,0x52E},{0x531,0x556},{0x10A0,0x10C5},{0x10C7,0x10C7},{0x10CD,0x10CD},{0x13A0,0x13F5}, + {0x1C90,0x1CBA},{0x1CBD,0x1CBF},{0x1E00,0x1E00},{0x1E02,0x1E02},{0x1E04,0x1E04},{0x1E06,0x1E06}, + {0x1E08,0x1E08},{0x1E0A,0x1E0A},{0x1E0C,0x1E0C},{0x1E0E,0x1E0E},{0x1E10,0x1E10},{0x1E12,0x1E12}, + {0x1E14,0x1E14},{0x1E16,0x1E16},{0x1E18,0x1E18},{0x1E1A,0x1E1A},{0x1E1C,0x1E1C},{0x1E1E,0x1E1E}, + {0x1E20,0x1E20},{0x1E22,0x1E22},{0x1E24,0x1E24},{0x1E26,0x1E26},{0x1E28,0x1E28},{0x1E2A,0x1E2A}, + {0x1E2C,0x1E2C},{0x1E2E,0x1E2E},{0x1E30,0x1E30},{0x1E32,0x1E32},{0x1E34,0x1E34},{0x1E36,0x1E36}, + {0x1E38,0x1E38},{0x1E3A,0x1E3A},{0x1E3C,0x1E3C},{0x1E3E,0x1E3E},{0x1E40,0x1E40},{0x1E42,0x1E42}, + {0x1E44,0x1E44},{0x1E46,0x1E46},{0x1E48,0x1E48},{0x1E4A,0x1E4A},{0x1E4C,0x1E4C},{0x1E4E,0x1E4E}, + {0x1E50,0x1E50},{0x1E52,0x1E52},{0x1E54,0x1E54},{0x1E56,0x1E56},{0x1E58,0x1E58},{0x1E5A,0x1E5A}, + {0x1E5C,0x1E5C},{0x1E5E,0x1E5E},{0x1E60,0x1E60},{0x1E62,0x1E62},{0x1E64,0x1E64},{0x1E66,0x1E66}, + {0x1E68,0x1E68},{0x1E6A,0x1E6A},{0x1E6C,0x1E6C},{0x1E6E,0x1E6E},{0x1E70,0x1E70},{0x1E72,0x1E72}, + {0x1E74,0x1E74},{0x1E76,0x1E76},{0x1E78,0x1E78},{0x1E7A,0x1E7A},{0x1E7C,0x1E7C},{0x1E7E,0x1E7E}, + {0x1E80,0x1E80},{0x1E82,0x1E82},{0x1E84,0x1E84},{0x1E86,0x1E86},{0x1E88,0x1E88},{0x1E8A,0x1E8A}, + {0x1E8C,0x1E8C},{0x1E8E,0x1E8E},{0x1E90,0x1E90},{0x1E92,0x1E92},{0x1E94,0x1E94},{0x1E9E,0x1E9E}, + {0x1EA0,0x1EA0},{0x1EA2,0x1EA2},{0x1EA4,0x1EA4},{0x1EA6,0x1EA6},{0x1EA8,0x1EA8},{0x1EAA,0x1EAA}, + {0x1EAC,0x1EAC},{0x1EAE,0x1EAE},{0x1EB0,0x1EB0},{0x1EB2,0x1EB2},{0x1EB4,0x1EB4},{0x1EB6,0x1EB6}, + {0x1EB8,0x1EB8},{0x1EBA,0x1EBA},{0x1EBC,0x1EBC},{0x1EBE,0x1EBE},{0x1EC0,0x1EC0},{0x1EC2,0x1EC2}, + {0x1EC4,0x1EC4},{0x1EC6,0x1EC6},{0x1EC8,0x1EC8},{0x1ECA,0x1ECA},{0x1ECC,0x1ECC},{0x1ECE,0x1ECE}, + {0x1ED0,0x1ED0},{0x1ED2,0x1ED2},{0x1ED4,0x1ED4},{0x1ED6,0x1ED6},{0x1ED8,0x1ED8},{0x1EDA,0x1EDA}, + {0x1EDC,0x1EDC},{0x1EDE,0x1EDE},{0x1EE0,0x1EE0},{0x1EE2,0x1EE2},{0x1EE4,0x1EE4},{0x1EE6,0x1EE6}, + {0x1EE8,0x1EE8},{0x1EEA,0x1EEA},{0x1EEC,0x1EEC},{0x1EEE,0x1EEE},{0x1EF0,0x1EF0},{0x1EF2,0x1EF2}, + {0x1EF4,0x1EF4},{0x1EF6,0x1EF6},{0x1EF8,0x1EF8},{0x1EFA,0x1EFA},{0x1EFC,0x1EFC},{0x1EFE,0x1EFE}, + {0x1F08,0x1F0F},{0x1F18,0x1F1D},{0x1F28,0x1F2F},{0x1F38,0x1F3F},{0x1F48,0x1F4D},{0x1F59,0x1F59}, + {0x1F5B,0x1F5B},{0x1F5D,0x1F5D},{0x1F5F,0x1F5F},{0x1F68,0x1F6F},{0x1F88,0x1F8F},{0x1F98,0x1F9F}, + {0x1FA8,0x1FAF},{0x1FB8,0x1FBC},{0x1FC8,0x1FCC},{0x1FD8,0x1FDB},{0x1FE8,0x1FEC},{0x1FF8,0x1FFC}, + {0x2102,0x2102},{0x2107,0x2107},{0x210B,0x210D},{0x2110,0x2112},{0x2115,0x2115},{0x2119,0x211D}, + {0x2124,0x2124},{0x2126,0x2126},{0x2128,0x2128},{0x212A,0x212D},{0x2130,0x2133},{0x213E,0x213F}, + {0x2145,0x2145},{0x2183,0x2183},{0x2C00,0x2C2E},{0x2C60,0x2C60},{0x2C62,0x2C64},{0x2C67,0x2C67}, + {0x2C69,0x2C69},{0x2C6B,0x2C6B},{0x2C6D,0x2C70},{0x2C72,0x2C72},{0x2C75,0x2C75},{0x2C7E,0x2C80}, + {0x2C82,0x2C82},{0x2C84,0x2C84},{0x2C86,0x2C86},{0x2C88,0x2C88},{0x2C8A,0x2C8A},{0x2C8C,0x2C8C}, + {0x2C8E,0x2C8E},{0x2C90,0x2C90},{0x2C92,0x2C92},{0x2C94,0x2C94},{0x2C96,0x2C96},{0x2C98,0x2C98}, + {0x2C9A,0x2C9A},{0x2C9C,0x2C9C},{0x2C9E,0x2C9E},{0x2CA0,0x2CA0},{0x2CA2,0x2CA2},{0x2CA4,0x2CA4}, + {0x2CA6,0x2CA6},{0x2CA8,0x2CA8},{0x2CAA,0x2CAA},{0x2CAC,0x2CAC},{0x2CAE,0x2CAE},{0x2CB0,0x2CB0}, + {0x2CB2,0x2CB2},{0x2CB4,0x2CB4},{0x2CB6,0x2CB6},{0x2CB8,0x2CB8},{0x2CBA,0x2CBA},{0x2CBC,0x2CBC}, + {0x2CBE,0x2CBE},{0x2CC0,0x2CC0},{0x2CC2,0x2CC2},{0x2CC4,0x2CC4},{0x2CC6,0x2CC6},{0x2CC8,0x2CC8}, + {0x2CCA,0x2CCA},{0x2CCC,0x2CCC},{0x2CCE,0x2CCE},{0x2CD0,0x2CD0},{0x2CD2,0x2CD2},{0x2CD4,0x2CD4}, + {0x2CD6,0x2CD6},{0x2CD8,0x2CD8},{0x2CDA,0x2CDA},{0x2CDC,0x2CDC},{0x2CDE,0x2CDE},{0x2CE0,0x2CE0}, + {0x2CE2,0x2CE2},{0x2CEB,0x2CEB},{0x2CED,0x2CED},{0x2CF2,0x2CF2},{0xA640,0xA640},{0xA642,0xA642}, + {0xA644,0xA644},{0xA646,0xA646},{0xA648,0xA648},{0xA64A,0xA64A},{0xA64C,0xA64C},{0xA64E,0xA64E}, + {0xA650,0xA650},{0xA652,0xA652},{0xA654,0xA654},{0xA656,0xA656},{0xA658,0xA658},{0xA65A,0xA65A}, + {0xA65C,0xA65C},{0xA65E,0xA65E},{0xA660,0xA660},{0xA662,0xA662},{0xA664,0xA664},{0xA666,0xA666}, + {0xA668,0xA668},{0xA66A,0xA66A},{0xA66C,0xA66C},{0xA680,0xA680},{0xA682,0xA682},{0xA684,0xA684}, + {0xA686,0xA686},{0xA688,0xA688},{0xA68A,0xA68A},{0xA68C,0xA68C},{0xA68E,0xA68E},{0xA690,0xA690}, + {0xA692,0xA692},{0xA694,0xA694},{0xA696,0xA696},{0xA698,0xA698},{0xA69A,0xA69A},{0xA722,0xA722}, + {0xA724,0xA724},{0xA726,0xA726},{0xA728,0xA728},{0xA72A,0xA72A},{0xA72C,0xA72C},{0xA72E,0xA72E}, + {0xA732,0xA732},{0xA734,0xA734},{0xA736,0xA736},{0xA738,0xA738},{0xA73A,0xA73A},{0xA73C,0xA73C}, + {0xA73E,0xA73E},{0xA740,0xA740},{0xA742,0xA742},{0xA744,0xA744},{0xA746,0xA746},{0xA748,0xA748}, + {0xA74A,0xA74A},{0xA74C,0xA74C},{0xA74E,0xA74E},{0xA750,0xA750},{0xA752,0xA752},{0xA754,0xA754}, + {0xA756,0xA756},{0xA758,0xA758},{0xA75A,0xA75A},{0xA75C,0xA75C},{0xA75E,0xA75E},{0xA760,0xA760}, + {0xA762,0xA762},{0xA764,0xA764},{0xA766,0xA766},{0xA768,0xA768},{0xA76A,0xA76A},{0xA76C,0xA76C}, + {0xA76E,0xA76E},{0xA779,0xA779},{0xA77B,0xA77B},{0xA77D,0xA77E},{0xA780,0xA780},{0xA782,0xA782}, + {0xA784,0xA784},{0xA786,0xA786},{0xA78B,0xA78B},{0xA78D,0xA78D},{0xA790,0xA790},{0xA792,0xA792}, + {0xA796,0xA796},{0xA798,0xA798},{0xA79A,0xA79A},{0xA79C,0xA79C},{0xA79E,0xA79E},{0xA7A0,0xA7A0}, + {0xA7A2,0xA7A2},{0xA7A4,0xA7A4},{0xA7A6,0xA7A6},{0xA7A8,0xA7A8},{0xA7AA,0xA7AE},{0xA7B0,0xA7B4}, + {0xA7B6,0xA7B6},{0xA7B8,0xA7B8},{0xA7BA,0xA7BA},{0xA7BC,0xA7BC},{0xA7BE,0xA7BE},{0xA7C2,0xA7C2}, + {0xA7C4,0xA7C7},{0xA7C9,0xA7C9},{0xA7F5,0xA7F5},{0xFF21,0xFF3A},{0x10400,0x10427},{0x104B0,0x104D3}, + {0x10C80,0x10CB2},{0x118A0,0x118BF},{0x16E40,0x16E5F},{0x1D400,0x1D419},{0x1D434,0x1D44D},{0x1D468,0x1D481}, + {0x1D49C,0x1D49C},{0x1D49E,0x1D49F},{0x1D4A2,0x1D4A2},{0x1D4A5,0x1D4A6},{0x1D4A9,0x1D4AC},{0x1D4AE,0x1D4B5}, + {0x1D4D0,0x1D4E9},{0x1D504,0x1D505},{0x1D507,0x1D50A},{0x1D50D,0x1D514},{0x1D516,0x1D51C},{0x1D538,0x1D539}, + {0x1D53B,0x1D53E},{0x1D540,0x1D544},{0x1D546,0x1D546},{0x1D54A,0x1D550},{0x1D56C,0x1D585},{0x1D5A0,0x1D5B9}, + {0x1D5D4,0x1D5ED},{0x1D608,0x1D621},{0x1D63C,0x1D655},{0x1D670,0x1D689},{0x1D6A8,0x1D6C0},{0x1D6E2,0x1D6FA}, + {0x1D71C,0x1D734},{0x1D756,0x1D76E},{0x1D790,0x1D7A8},{0x1D7CA,0x1D7CA},{0x1E900,0x1E921}, +}; +static const int uni_U_n = 641; + +static const uint32_t uni_X[][2] = { + {0xAA,0xAA},{0xBA,0xBA},{0x1BB,0x1BB},{0x1C0,0x1C3},{0x294,0x294},{0x2B0,0x2C1}, + {0x2C6,0x2D1},{0x2E0,0x2E4},{0x2EC,0x2EC},{0x2EE,0x2EE},{0x300,0x36F},{0x374,0x374}, + {0x37A,0x37A},{0x483,0x489},{0x559,0x559},{0x591,0x5BD},{0x5BF,0x5BF},{0x5C1,0x5C2}, + {0x5C4,0x5C5},{0x5C7,0x5C7},{0x5D0,0x5EA},{0x5EF,0x5F2},{0x610,0x61A},{0x620,0x65F}, + {0x66E,0x6D3},{0x6D5,0x6DC},{0x6DF,0x6E8},{0x6EA,0x6EF},{0x6FA,0x6FC},{0x6FF,0x6FF}, + {0x710,0x74A},{0x74D,0x7B1},{0x7CA,0x7F5},{0x7FA,0x7FA},{0x7FD,0x7FD},{0x800,0x82D}, + {0x840,0x85B},{0x860,0x86A},{0x8A0,0x8B4},{0x8B6,0x8C7},{0x8D3,0x8E1},{0x8E3,0x963}, + {0x971,0x983},{0x985,0x98C},{0x98F,0x990},{0x993,0x9A8},{0x9AA,0x9B0},{0x9B2,0x9B2}, + {0x9B6,0x9B9},{0x9BC,0x9C4},{0x9C7,0x9C8},{0x9CB,0x9CE},{0x9D7,0x9D7},{0x9DC,0x9DD}, + {0x9DF,0x9E3},{0x9F0,0x9F1},{0x9FC,0x9FC},{0x9FE,0x9FE},{0xA01,0xA03},{0xA05,0xA0A}, + {0xA0F,0xA10},{0xA13,0xA28},{0xA2A,0xA30},{0xA32,0xA33},{0xA35,0xA36},{0xA38,0xA39}, + {0xA3C,0xA3C},{0xA3E,0xA42},{0xA47,0xA48},{0xA4B,0xA4D},{0xA51,0xA51},{0xA59,0xA5C}, + {0xA5E,0xA5E},{0xA70,0xA75},{0xA81,0xA83},{0xA85,0xA8D},{0xA8F,0xA91},{0xA93,0xAA8}, + {0xAAA,0xAB0},{0xAB2,0xAB3},{0xAB5,0xAB9},{0xABC,0xAC5},{0xAC7,0xAC9},{0xACB,0xACD}, + {0xAD0,0xAD0},{0xAE0,0xAE3},{0xAF9,0xAFF},{0xB01,0xB03},{0xB05,0xB0C},{0xB0F,0xB10}, + {0xB13,0xB28},{0xB2A,0xB30},{0xB32,0xB33},{0xB35,0xB39},{0xB3C,0xB44},{0xB47,0xB48}, + {0xB4B,0xB4D},{0xB55,0xB57},{0xB5C,0xB5D},{0xB5F,0xB63},{0xB71,0xB71},{0xB82,0xB83}, + {0xB85,0xB8A},{0xB8E,0xB90},{0xB92,0xB95},{0xB99,0xB9A},{0xB9C,0xB9C},{0xB9E,0xB9F}, + {0xBA3,0xBA4},{0xBA8,0xBAA},{0xBAE,0xBB9},{0xBBE,0xBC2},{0xBC6,0xBC8},{0xBCA,0xBCD}, + {0xBD0,0xBD0},{0xBD7,0xBD7},{0xC00,0xC0C},{0xC0E,0xC10},{0xC12,0xC28},{0xC2A,0xC39}, + {0xC3D,0xC44},{0xC46,0xC48},{0xC4A,0xC4D},{0xC55,0xC56},{0xC58,0xC5A},{0xC60,0xC63}, + {0xC80,0xC83},{0xC85,0xC8C},{0xC8E,0xC90},{0xC92,0xCA8},{0xCAA,0xCB3},{0xCB5,0xCB9}, + {0xCBC,0xCC4},{0xCC6,0xCC8},{0xCCA,0xCCD},{0xCD5,0xCD6},{0xCDE,0xCDE},{0xCE0,0xCE3}, + {0xCF1,0xCF2},{0xD00,0xD0C},{0xD0E,0xD10},{0xD12,0xD44},{0xD46,0xD48},{0xD4A,0xD4E}, + {0xD54,0xD57},{0xD5F,0xD63},{0xD7A,0xD7F},{0xD81,0xD83},{0xD85,0xD96},{0xD9A,0xDB1}, + {0xDB3,0xDBB},{0xDBD,0xDBD},{0xDC0,0xDC6},{0xDCA,0xDCA},{0xDCF,0xDD4},{0xDD6,0xDD6}, + {0xDD8,0xDDF},{0xDF2,0xDF3},{0xE01,0xE3A},{0xE40,0xE4E},{0xE81,0xE82},{0xE84,0xE84}, + {0xE86,0xE8A},{0xE8C,0xEA3},{0xEA5,0xEA5},{0xEA7,0xEBD},{0xEC0,0xEC4},{0xEC6,0xEC6}, + {0xEC8,0xECD},{0xEDC,0xEDF},{0xF00,0xF00},{0xF18,0xF19},{0xF35,0xF35},{0xF37,0xF37}, + {0xF39,0xF39},{0xF3E,0xF47},{0xF49,0xF6C},{0xF71,0xF84},{0xF86,0xF97},{0xF99,0xFBC}, + {0xFC6,0xFC6},{0x1000,0x103F},{0x1050,0x108F},{0x109A,0x109D},{0x10FC,0x10FC},{0x1100,0x1248}, + {0x124A,0x124D},{0x1250,0x1256},{0x1258,0x1258},{0x125A,0x125D},{0x1260,0x1288},{0x128A,0x128D}, + {0x1290,0x12B0},{0x12B2,0x12B5},{0x12B8,0x12BE},{0x12C0,0x12C0},{0x12C2,0x12C5},{0x12C8,0x12D6}, + {0x12D8,0x1310},{0x1312,0x1315},{0x1318,0x135A},{0x135D,0x135F},{0x1380,0x138F},{0x1401,0x166C}, + {0x166F,0x167F},{0x1681,0x169A},{0x16A0,0x16EA},{0x16F1,0x16F8},{0x1700,0x170C},{0x170E,0x1714}, + {0x1720,0x1734},{0x1740,0x1753},{0x1760,0x176C},{0x176E,0x1770},{0x1772,0x1773},{0x1780,0x17D3}, + {0x17D7,0x17D7},{0x17DC,0x17DD},{0x180B,0x180D},{0x1820,0x1878},{0x1880,0x18AA},{0x18B0,0x18F5}, + {0x1900,0x191E},{0x1920,0x192B},{0x1930,0x193B},{0x1950,0x196D},{0x1970,0x1974},{0x1980,0x19AB}, + {0x19B0,0x19C9},{0x1A00,0x1A1B},{0x1A20,0x1A5E},{0x1A60,0x1A7C},{0x1A7F,0x1A7F},{0x1AA7,0x1AA7}, + {0x1AB0,0x1AC0},{0x1B00,0x1B4B},{0x1B6B,0x1B73},{0x1B80,0x1BAF},{0x1BBA,0x1BF3},{0x1C00,0x1C37}, + {0x1C4D,0x1C4F},{0x1C5A,0x1C7D},{0x1CD0,0x1CD2},{0x1CD4,0x1CFA},{0x1D2C,0x1D6A},{0x1D78,0x1D78}, + {0x1D9B,0x1DF9},{0x1DFB,0x1DFF},{0x2071,0x2071},{0x207F,0x207F},{0x2090,0x209C},{0x20D0,0x20F0}, + {0x2135,0x2138},{0x2C7C,0x2C7D},{0x2CEF,0x2CF1},{0x2D30,0x2D67},{0x2D6F,0x2D6F},{0x2D7F,0x2D96}, + {0x2DA0,0x2DA6},{0x2DA8,0x2DAE},{0x2DB0,0x2DB6},{0x2DB8,0x2DBE},{0x2DC0,0x2DC6},{0x2DC8,0x2DCE}, + {0x2DD0,0x2DD6},{0x2DD8,0x2DDE},{0x2DE0,0x2DFF},{0x2E2F,0x2E2F},{0x3005,0x3006},{0x302A,0x302F}, + {0x3031,0x3035},{0x303B,0x303C},{0x3041,0x3096},{0x3099,0x309A},{0x309D,0x309F},{0x30A1,0x30FA}, + {0x30FC,0x30FF},{0x3105,0x312F},{0x3131,0x318E},{0x31A0,0x31BF},{0x31F0,0x31FF},{0x3400,0x4DBF}, + {0x4E00,0x9FFC},{0xA000,0xA48C},{0xA4D0,0xA4FD},{0xA500,0xA60C},{0xA610,0xA61F},{0xA62A,0xA62B}, + {0xA66E,0xA672},{0xA674,0xA67D},{0xA67F,0xA67F},{0xA69C,0xA6E5},{0xA6F0,0xA6F1},{0xA717,0xA71F}, + {0xA770,0xA770},{0xA788,0xA788},{0xA78F,0xA78F},{0xA7F7,0xA7F9},{0xA7FB,0xA827},{0xA82C,0xA82C}, + {0xA840,0xA873},{0xA880,0xA8C5},{0xA8E0,0xA8F7},{0xA8FB,0xA8FB},{0xA8FD,0xA8FF},{0xA90A,0xA92D}, + {0xA930,0xA953},{0xA960,0xA97C},{0xA980,0xA9C0},{0xA9CF,0xA9CF},{0xA9E0,0xA9EF},{0xA9FA,0xA9FE}, + {0xAA00,0xAA36},{0xAA40,0xAA4D},{0xAA60,0xAA76},{0xAA7A,0xAAC2},{0xAADB,0xAADD},{0xAAE0,0xAAEF}, + {0xAAF2,0xAAF6},{0xAB01,0xAB06},{0xAB09,0xAB0E},{0xAB11,0xAB16},{0xAB20,0xAB26},{0xAB28,0xAB2E}, + {0xAB5C,0xAB5F},{0xAB69,0xAB69},{0xABC0,0xABEA},{0xABEC,0xABED},{0xAC00,0xD7A3},{0xD7B0,0xD7C6}, + {0xD7CB,0xD7FB},{0xF900,0xFA6D},{0xFA70,0xFAD9},{0xFB1D,0xFB28},{0xFB2A,0xFB36},{0xFB38,0xFB3C}, + {0xFB3E,0xFB3E},{0xFB40,0xFB41},{0xFB43,0xFB44},{0xFB46,0xFBB1},{0xFBD3,0xFD3D},{0xFD50,0xFD8F}, + {0xFD92,0xFDC7},{0xFDF0,0xFDFB},{0xFE00,0xFE0F},{0xFE20,0xFE2F},{0xFE70,0xFE74},{0xFE76,0xFEFC}, + {0xFF66,0xFFBE},{0xFFC2,0xFFC7},{0xFFCA,0xFFCF},{0xFFD2,0xFFD7},{0xFFDA,0xFFDC},{0x10000,0x1000B}, + {0x1000D,0x10026},{0x10028,0x1003A},{0x1003C,0x1003D},{0x1003F,0x1004D},{0x10050,0x1005D},{0x10080,0x100FA}, + {0x101FD,0x101FD},{0x10280,0x1029C},{0x102A0,0x102D0},{0x102E0,0x102E0},{0x10300,0x1031F},{0x1032D,0x10340}, + {0x10342,0x10349},{0x10350,0x1037A},{0x10380,0x1039D},{0x103A0,0x103C3},{0x103C8,0x103CF},{0x10450,0x1049D}, + {0x10500,0x10527},{0x10530,0x10563},{0x10600,0x10736},{0x10740,0x10755},{0x10760,0x10767},{0x10800,0x10805}, + {0x10808,0x10808},{0x1080A,0x10835},{0x10837,0x10838},{0x1083C,0x1083C},{0x1083F,0x10855},{0x10860,0x10876}, + {0x10880,0x1089E},{0x108E0,0x108F2},{0x108F4,0x108F5},{0x10900,0x10915},{0x10920,0x10939},{0x10980,0x109B7}, + {0x109BE,0x109BF},{0x10A00,0x10A03},{0x10A05,0x10A06},{0x10A0C,0x10A13},{0x10A15,0x10A17},{0x10A19,0x10A35}, + {0x10A38,0x10A3A},{0x10A3F,0x10A3F},{0x10A60,0x10A7C},{0x10A80,0x10A9C},{0x10AC0,0x10AC7},{0x10AC9,0x10AE6}, + {0x10B00,0x10B35},{0x10B40,0x10B55},{0x10B60,0x10B72},{0x10B80,0x10B91},{0x10C00,0x10C48},{0x10D00,0x10D27}, + {0x10E80,0x10EA9},{0x10EAB,0x10EAC},{0x10EB0,0x10EB1},{0x10F00,0x10F1C},{0x10F27,0x10F27},{0x10F30,0x10F50}, + {0x10FB0,0x10FC4},{0x10FE0,0x10FF6},{0x11000,0x11046},{0x1107F,0x110BA},{0x110D0,0x110E8},{0x11100,0x11134}, + {0x11144,0x11147},{0x11150,0x11173},{0x11176,0x11176},{0x11180,0x111C4},{0x111C9,0x111CC},{0x111CE,0x111CF}, + {0x111DA,0x111DA},{0x111DC,0x111DC},{0x11200,0x11211},{0x11213,0x11237},{0x1123E,0x1123E},{0x11280,0x11286}, + {0x11288,0x11288},{0x1128A,0x1128D},{0x1128F,0x1129D},{0x1129F,0x112A8},{0x112B0,0x112EA},{0x11300,0x11303}, + {0x11305,0x1130C},{0x1130F,0x11310},{0x11313,0x11328},{0x1132A,0x11330},{0x11332,0x11333},{0x11335,0x11339}, + {0x1133B,0x11344},{0x11347,0x11348},{0x1134B,0x1134D},{0x11350,0x11350},{0x11357,0x11357},{0x1135D,0x11363}, + {0x11366,0x1136C},{0x11370,0x11374},{0x11400,0x1144A},{0x1145E,0x11461},{0x11480,0x114C5},{0x114C7,0x114C7}, + {0x11580,0x115B5},{0x115B8,0x115C0},{0x115D8,0x115DD},{0x11600,0x11640},{0x11644,0x11644},{0x11680,0x116B8}, + {0x11700,0x1171A},{0x1171D,0x1172B},{0x11800,0x1183A},{0x118FF,0x11906},{0x11909,0x11909},{0x1190C,0x11913}, + {0x11915,0x11916},{0x11918,0x11935},{0x11937,0x11938},{0x1193B,0x11943},{0x119A0,0x119A7},{0x119AA,0x119D7}, + {0x119DA,0x119E1},{0x119E3,0x119E4},{0x11A00,0x11A3E},{0x11A47,0x11A47},{0x11A50,0x11A99},{0x11A9D,0x11A9D}, + {0x11AC0,0x11AF8},{0x11C00,0x11C08},{0x11C0A,0x11C36},{0x11C38,0x11C40},{0x11C72,0x11C8F},{0x11C92,0x11CA7}, + {0x11CA9,0x11CB6},{0x11D00,0x11D06},{0x11D08,0x11D09},{0x11D0B,0x11D36},{0x11D3A,0x11D3A},{0x11D3C,0x11D3D}, + {0x11D3F,0x11D47},{0x11D60,0x11D65},{0x11D67,0x11D68},{0x11D6A,0x11D8E},{0x11D90,0x11D91},{0x11D93,0x11D98}, + {0x11EE0,0x11EF6},{0x11FB0,0x11FB0},{0x12000,0x12399},{0x12480,0x12543},{0x13000,0x1342E},{0x14400,0x14646}, + {0x16800,0x16A38},{0x16A40,0x16A5E},{0x16AD0,0x16AED},{0x16AF0,0x16AF4},{0x16B00,0x16B36},{0x16B40,0x16B43}, + {0x16B63,0x16B77},{0x16B7D,0x16B8F},{0x16F00,0x16F4A},{0x16F4F,0x16F87},{0x16F8F,0x16F9F},{0x16FE0,0x16FE1}, + {0x16FE3,0x16FE4},{0x16FF0,0x16FF1},{0x17000,0x187F7},{0x18800,0x18CD5},{0x18D00,0x18D08},{0x1B000,0x1B11E}, + {0x1B150,0x1B152},{0x1B164,0x1B167},{0x1B170,0x1B2FB},{0x1BC00,0x1BC6A},{0x1BC70,0x1BC7C},{0x1BC80,0x1BC88}, + {0x1BC90,0x1BC99},{0x1BC9D,0x1BC9E},{0x1D165,0x1D169},{0x1D16D,0x1D172},{0x1D17B,0x1D182},{0x1D185,0x1D18B}, + {0x1D1AA,0x1D1AD},{0x1D242,0x1D244},{0x1DA00,0x1DA36},{0x1DA3B,0x1DA6C},{0x1DA75,0x1DA75},{0x1DA84,0x1DA84}, + {0x1DA9B,0x1DA9F},{0x1DAA1,0x1DAAF},{0x1E000,0x1E006},{0x1E008,0x1E018},{0x1E01B,0x1E021},{0x1E023,0x1E024}, + {0x1E026,0x1E02A},{0x1E100,0x1E12C},{0x1E130,0x1E13D},{0x1E14E,0x1E14E},{0x1E2C0,0x1E2EF},{0x1E800,0x1E8C4}, + {0x1E8D0,0x1E8D6},{0x1E944,0x1E94B},{0x1EE00,0x1EE03},{0x1EE05,0x1EE1F},{0x1EE21,0x1EE22},{0x1EE24,0x1EE24}, + {0x1EE27,0x1EE27},{0x1EE29,0x1EE32},{0x1EE34,0x1EE37},{0x1EE39,0x1EE39},{0x1EE3B,0x1EE3B},{0x1EE42,0x1EE42}, + {0x1EE47,0x1EE47},{0x1EE49,0x1EE49},{0x1EE4B,0x1EE4B},{0x1EE4D,0x1EE4F},{0x1EE51,0x1EE52},{0x1EE54,0x1EE54}, + {0x1EE57,0x1EE57},{0x1EE59,0x1EE59},{0x1EE5B,0x1EE5B},{0x1EE5D,0x1EE5D},{0x1EE5F,0x1EE5F},{0x1EE61,0x1EE62}, + {0x1EE64,0x1EE64},{0x1EE67,0x1EE6A},{0x1EE6C,0x1EE72},{0x1EE74,0x1EE77},{0x1EE79,0x1EE7C},{0x1EE7E,0x1EE7E}, + {0x1EE80,0x1EE89},{0x1EE8B,0x1EE9B},{0x1EEA1,0x1EEA3},{0x1EEA5,0x1EEA9},{0x1EEAB,0x1EEBB},{0x20000,0x2A6DD}, + {0x2A700,0x2B734},{0x2B740,0x2B81D},{0x2B820,0x2CEA1},{0x2CEB0,0x2EBE0},{0x2F800,0x2FA1D},{0x30000,0x3134A}, + {0xE0100,0xE01EF}, +}; +static const int uni_X_n = 595; + +static inline int is_U(uint32_t c){ return uni_in(uni_U,uni_U_n,c); } +static inline int is_X(uint32_t c){ return uni_in(uni_X,uni_X_n,c); } +#endif diff --git a/c/tools/convert_fp8_to_int4.py b/c/tools/convert_fp8_to_int4.py index eac51f9..447654d 100644 --- a/c/tools/convert_fp8_to_int4.py +++ b/c/tools/convert_fp8_to_int4.py @@ -83,6 +83,29 @@ def quant_int4_grouped(w, bits, gs=128): s_flat = s[:, :, 0].astype(np.float32).reshape(-1) return out.reshape(-1), s_flat +def quant_int3_g64(w, bits=3, group=64): # -> (qbytes U8 [O*ceil(I/64)*24], scales f32 [O*ceil(I/64)]) + """int3 with PER-GROUP scales (fmt=5 in colibri.c): per 64-input group, symmetric absmax + (qmax=3, clamp [-4,3], stored v+4), packed as 16B low plane (2 bits/val, int2 layout) + + 8B high plane (1 bit/val). Same math as quant_ablation._quant_last_dim(bits=3, + group=64) (#132), here with real packing. 3.5 bits/weight effective.""" + O, I = w.shape + ng = (I + group - 1) // group + pad = ng * group - I + wp = np.pad(w, ((0, 0), (0, pad))) if pad else w + g = wp.reshape(O, ng, group) + amax = np.abs(g).max(axis=2, keepdims=True) + s = np.maximum(amax / 3.0, 1e-8) + q = (np.clip(np.rint(g / s), -4, 3).astype(np.int32) + 4).astype(np.uint8) # 0..7 + if pad: q[:, -1, group - pad:] = 4 # pad packs as 0 after -4 + lo = np.zeros((O, ng, 16), np.uint8) + for k in range(4): + lo |= ((q[:, :, k::4] & 3) << (k * 2)).astype(np.uint8) + hi = np.zeros((O, ng, 8), np.uint8) + for b in range(8): + hi |= (((q[:, :, b::8] >> 2) & 1) << b).astype(np.uint8) + out = np.concatenate([lo, hi], axis=2) # [O, ng, 24] + return out.reshape(-1), s[:, :, 0].astype(np.float32).reshape(-1) + def quant_int2(w, bits): # -> (qbytes U8 [O*ceil(I/4)], scale f32 [O]); 4/byte O, I = w.shape qmax = (1 << (bits - 1)) - 1 # bits=2 -> qmax=1, valori [-2,1] @@ -214,6 +237,14 @@ def dequant(f, name, keys): return (w * sc).numpy() return f.get_tensor(name).to(torch.float32).numpy() +# Per-projection bit overrides for ROUTED experts (gate_proj/up_proj/down_proj), set from +# --up-bits/--gate-bits/--down-bits in main(). Empty = uniform xbits. Motivated by the +# measured result that up_proj tolerates int3-g64 at ~zero quality cost while int2 craters +# (OLMoE ablation, PR #168 comment): up-only int3 drops ~8% of expert bytes for free. +# NB: the resume manifests (check_or_record_params and the --indir progress file) already +# record dict(PROJ_BITS) — this global is the definition those sites depend on. +PROJ_BITS = {} + def convert_shard(path, out_dict, n_layers, ebits, io_bits, xbits, keep_mtp=False, keep_idx=False, group_size=0, bits_map=None): from safetensors import safe_open @@ -235,9 +266,16 @@ def convert_shard(path, out_dict, n_layers, ebits, io_bits, xbits, # Any unknown kind that fell through classify as "q" if bits_map and kind not in bits_map and kind not in ("io", "x", "sh", "o", "kvb", "attn", "dmlp"): bits = ebits + # Per-projection override for routed experts, applied on top of the type-level bits. + if kind == "x" and PROJ_BITS: # e.g. up_proj -> 3 (int3-g64) while gate/down stay 4 + for proj, pb in PROJ_BITS.items(): + if f".{proj}.weight" in name: bits = pb; break if w.ndim != 2: # es. bias 1D non previsto come 'q' -> tienilo f32 out_dict[name] = w.astype(np.float32); continue - if group_size > 0 and bits <= 4: + if bits == 3: + # int3-g64 (fmt=5): inherently group-64, distinct from grouped-int4. + q, s = quant_int3_g64(w) + elif group_size > 0 and bits <= 4: q, s = quant_int4_grouped(w, bits, group_size) else: q, s = (quant_int2(w, bits) if bits <= 2 else @@ -247,6 +285,32 @@ def convert_shard(path, out_dict, n_layers, ebits, io_bits, xbits, def free_gb(p): return shutil.disk_usage(p).free / 1e9 +def check_or_record_params(outdir, prefix, params): + """#383-class guard, mirrored onto the --repo download loops from the --indir + path's resume manifest (below): a resumed run with DIFFERENT conversion + parameters (bits, group size, PROJ_BITS, ...) must not silently mix bit-widths + across shards in the same outdir -- the #355 failure mode (a second pass with + changed flags overwriting/interleaving with a finished container in silence). + Unlike the --indir manifest this doesn't need to track per-shard completion: + the --repo loops already do that via out-NNNNN.safetensors existence, since + shard index maps directly to output filename there. Only whether the params + used SO FAR match this run's needs checking. Returns False (caller should + abort) on a mismatch, True otherwise; records params on first use.""" + path = os.path.join(outdir, f".{prefix}params.json") + if os.path.exists(path): + try: prev = json.loads(open(path).read()) + except (OSError, ValueError): prev = None + if prev is not None and prev != params: + print(f"ERROR: {path} records a conversion with {prev};\n" + f" this run uses {params}. Refusing to mix conversions in the " + f"same outdir — use a fresh --outdir (or delete {path} and the " + f"{prefix}*.safetensors shards to redo).") + return False + tmp = path + ".tmp" + with open(tmp, "w") as f: json.dump(params, f, indent=1) # atomic write, same reasoning as the --indir manifest + os.replace(tmp, path) + return True + def main(): ap = argparse.ArgumentParser() ap.add_argument("--repo", default=None) @@ -269,6 +333,13 @@ def main(): help="bits for dense MLP (first 3 layers). Default=ebits") ap.add_argument("--group-size", type=int, default=0, # 0 = per-row (backward compat); 128 = group-scaled help="group size for int4 scales: 0=per-row (default), 128=one scale per 128 elements (much better quality)") + # Per-projection bit overrides for routed experts (orthogonal to the type-level flags above). + ap.add_argument("--up-bits", type=int, default=None, + help="bits for up_proj in routed experts (e.g. 3 = int3-g64). Default=xbits") + ap.add_argument("--gate-bits", type=int, default=None, + help="bits for gate_proj in routed experts. Default=xbits") + ap.add_argument("--down-bits", type=int, default=None, + help="bits for down_proj in routed experts. Default=xbits") ap.add_argument("--n-layers", type=int, default=78) ap.add_argument("--min-free-gb", type=float, default=20.0) ap.add_argument("--selftest", action="store_true") @@ -298,6 +369,10 @@ def main(): "embedding half -> MTP acceptance ~0% (issue #8). Use the default --ebits 8, " "or add --group-size 128 for group-scaled int4.") if a.xbits is None: a.xbits = a.ebits + for proj, val in (("gate_proj", a.gate_bits), ("up_proj", a.up_bits), ("down_proj", a.down_bits)): + if val is not None: PROJ_BITS[proj] = val + if PROJ_BITS: + print(f"[per-projection expert bits] {PROJ_BITS} (others -> xbits={a.xbits})") # Build per-type bits map. If a type-specific arg is set, use it; otherwise the # converter falls back to ebits for that type. @@ -440,7 +515,8 @@ def main(): # EN: resume skips only what matches, and different parameters on the same # EN: outdir are refused instead of mixing containers (the #355 failure mode). params = {"ebits": a.ebits, "io_bits": a.io_bits, "xbits": a.xbits, - "group_size": a.group_size, "n_layers": a.n_layers, "bits_map": bits_map} + "group_size": a.group_size, "n_layers": a.n_layers, "bits_map": bits_map, + "proj_bits": dict(PROJ_BITS)} prog_path = os.path.join(a.outdir, f".{prefix}progress.json") prog = {} if os.path.exists(prog_path): @@ -718,6 +794,10 @@ def main(): except Exception: pass tmp = os.path.join(a.outdir, "_inflight"); os.makedirs(tmp, exist_ok=True) if a.mtp: + params = {"ebits": a.ebits, "io_bits": a.io_bits, "xbits": a.xbits, + "group_size": a.group_size, "n_layers": a.n_layers, "bits_map": bits_map, + "proj_bits": dict(PROJ_BITS)} + if not check_or_record_params(a.outdir, "out-mtp-", params): return import urllib.request idx = json.loads(urllib.request.urlopen( f"https://huggingface.co/{a.repo}/resolve/main/model.safetensors.index.json", timeout=30).read())["weight_map"] @@ -737,6 +817,10 @@ def main(): print(f" -> {os.path.basename(outp)} ({os.path.getsize(outp)/1e9:.2f} GB, {len(out)} tensors)", flush=True) shutil.rmtree(tmp, ignore_errors=True); print("[MTP] DONE."); return if a.indexer: + params = {"ebits": a.ebits, "io_bits": a.io_bits, "xbits": a.xbits, + "group_size": a.group_size, "n_layers": a.n_layers, "bits_map": bits_map, + "proj_bits": dict(PROJ_BITS)} + if not check_or_record_params(a.outdir, "out-idx-", params): return import urllib.request idx = json.loads(urllib.request.urlopen( f"https://huggingface.co/{a.repo}/resolve/main/model.safetensors.index.json", timeout=30).read())["weight_map"] @@ -756,6 +840,10 @@ def main(): if os.path.isfile(blob): os.remove(blob) print(f" -> {os.path.basename(outp)} ({len(out)} tensors)", flush=True) shutil.rmtree(tmp, ignore_errors=True); print("[IDX] DONE."); return + params = {"ebits": a.ebits, "io_bits": a.io_bits, "xbits": a.xbits, + "group_size": a.group_size, "n_layers": a.n_layers, "bits_map": bits_map, + "proj_bits": dict(PROJ_BITS)} + if not check_or_record_params(a.outdir, "out-", params): return for i, sh in enumerate(shards): if free_gb(a.outdir) < a.min_free_gb: print(f"STOP: free space is below {a.min_free_gb} GB. Free space and rerun to resume."); break diff --git a/c/tools/diag_harness.py b/c/tools/diag_harness.py new file mode 100644 index 0000000..56a1b22 --- /dev/null +++ b/c/tools/diag_harness.py @@ -0,0 +1,704 @@ +#!/usr/bin/env python3 +""" +diag_harness.py — Comprehensive model diagnostic harness for the colibri GLM-5.2 engine. + +Runs a full campaign of tests against any model snapshot: + Phase 0 (system): startup telemetry — GPU, RAM, cache cap, load time, MTP, idot kernel + Phase 1 (smoke): correctness — curated prompts, coherence checks, corruption detection + Phase 2 (diagnostic): deep x-ray — full PROFILE breakdown, routing, MTP, disk-split, CUDA tier + Phase 3 (quality): benchmark accuracy — hellaswag/arc_challenge/mmlu via eval_glm.py SCORE + Phase 4 (throughput): tok/s with and without MTP speculation + Phase 5 (report): structured JSON + human-readable Markdown summary + +Usage: + python tools/diag_harness.py --snap /path/to/model --phase all + python tools/diag_harness.py --snap /path/to/model --phase smoke --ngen 64 + python tools/diag_harness.py --snap /path/to/model --phase quality --quality-limit 200 + +Output goes to --out (default ./diag_results//). Each phase writes a raw log +(_.log) and all metrics are collected into report.json + report.md. + +Design notes: + - stdout and stderr are captured separately via subprocess.PIPE. The engine streams + generated text to stdout (interleaved with prompt + PROFILE stats), but TOKENS=1 dumps + clean token-id lists to stderr — that is the primary text-capture path. + - Every regex is anchored to the exact printf format strings in glm.c (verified against + profile_print line 3853, run_text line 3948, the banner line 5299, etc.). + - Subprocess calls have a hard timeout (default 600s) and are killed cleanly on expiry. + - A single phase can be run standalone; results accumulate in the output dir. +""" +import os, sys, re, json, time, argparse, subprocess, signal, traceback +from datetime import datetime +from pathlib import Path + +# --------------------------------------------------------------------------- +# PROMPT SUITE — curated across categories. "expect" is a case-insensitive +# substring checked against the generated text (None = coherence-only check). +# --------------------------------------------------------------------------- +PROMPTS = [ + {"id":"fact_capital", "cat":"factual", "prompt":"The capital of France is", + "expect":"Paris", "note":"basic world knowledge"}, + {"id":"fact_boiling", "cat":"factual", "prompt":"What is the boiling point of water in Celsius?", + "expect":"100", "note":"basic science"}, + {"id":"fact_planet", "cat":"factual", "prompt":"What planet is closest to the Sun?", + "expect":"Mercury","note":"astronomy fact"}, + {"id":"math_mult", "cat":"math", "prompt":"What is 15 times 12?", + "expect":"180", "note":"2-digit multiplication"}, + {"id":"math_add", "cat":"math", "prompt":"What is 847 plus 153?", + "expect":"1000", "note":"3-digit addition"}, + {"id":"reason_train", "cat":"reasoning", "prompt":"If a train travels 60 mph for 2.5 hours, how far does it go?", + "expect":"150", "note":"rate-time-distance"}, + {"id":"code_factorial","cat":"code", "prompt":"Write a Python function that computes the factorial of a number.", + "expect":"def", "note":"code generation"}, + {"id":"explain_nn", "cat":"explanation","prompt":"Explain what a neural network is in one sentence.", + "expect":None, "note":"coherence check"}, + {"id":"creative_story","cat":"creative", "prompt":"Write a one-sentence story about a lighthouse.", + "expect":None, "note":"creative coherence"}, + {"id":"edge_hello", "cat":"edge", "prompt":"Hello, how are you today?", + "expect":None, "note":"conversational opener"}, + {"id":"edge_the", "cat":"edge", "prompt":"The weather today is", + "expect":None, "note":"simple continuation"}, + {"id":"edge_repeat", "cat":"edge", "prompt":"The quick brown fox jumps over the lazy dog. The quick brown fox", + "expect":None, "note":"repetition-bait"}, +] + +# --------------------------------------------------------------------------- +# METRIC EXTRACTION — each parser is (name, compiled_regex, group_index, cast). +# Regexes match the EXACT printf format strings in glm.c. +# --------------------------------------------------------------------------- +def _f1(g): return float(g) +def _i1(g): return int(g) + +# Build parsers as (regex, lambda(match)->value) for clarity +RX = { + # stdout: run_text summary line (line 3948) + "decode_toks": (re.compile(r"decode (\d+) tokens in"), lambda m: int(m.group(1))), + "decode_secs": (re.compile(r"decode \d+ tokens in ([\d.]+)s"), lambda m: float(m.group(1))), + "decode_tps": (re.compile(r"decode \d+ tokens in [\d.]+s \(([\d.]+) tok/s\)"), lambda m: float(m.group(1))), + "prefill_toks": (re.compile(r"prefill (\d+) tokens in"), lambda m: int(m.group(1))), + "prefill_secs": (re.compile(r"prefill \d+ tokens in ([\d.]+)s"), lambda m: float(m.group(1))), + "hit_rate": (re.compile(r"expert hit rate ([\d.]+)%"), lambda m: float(m.group(1))), + "rss_gb": (re.compile(r"RSS ([\d.]+) GB"), lambda m: float(m.group(1))), + "experts_per_tok":(re.compile(r"experts loaded/token: ([\d.]+)"), lambda m: float(m.group(1))), + "mtp_accept": (re.compile(r"MTP acceptance ([\d.]+)%"), lambda m: float(m.group(1))), + "mtp_acc_cnt": (re.compile(r"MTP acceptance \d+% \((\d+)/(\d+)\)"), lambda m: (int(m.group(1)), int(m.group(2)))), + "spec_tok_per_fw":(re.compile(r"speculation: ([\d.]+) tokens/forward"),lambda m: float(m.group(1))), + # stdout: PROFILE lines (profile_print, line 3853-3864) + "prof_expert_disk": (re.compile(r"expert-disk ([\d.]+)s service"), lambda m: float(m.group(1))), + "prof_expert_wait": (re.compile(r"service / ([\d.]+)s wait"), lambda m: float(m.group(1))), + "prof_expert_mm": (re.compile(r"expert-matmul ([\d.]+)s"), lambda m: float(m.group(1))), + "prof_attention": (re.compile(r"\| attention ([\d.]+)s"), lambda m: float(m.group(1))), + "prof_kvb": (re.compile(r"including kvb ([\d.]+)s"), lambda m: float(m.group(1))), + "prof_lm_head": (re.compile(r"lm_head ([\d.]+)s"), lambda m: float(m.group(1))), + "prof_other": (re.compile(r"\| other ([\d.]+)s"), lambda m: float(m.group(1))), + # stdout: banner (line 5299) + "load_secs": (re.compile(r"loaded in ([\d.]+)s"), lambda m: float(m.group(1))), + "resident_mb": (re.compile(r"resident dense: ([\d.]+) MB"), lambda m: float(m.group(1))), + "mtp_status": (re.compile(r"MTP (ACTIVE|absent)"), lambda m: m.group(1)), + "n_layers": (re.compile(r"layers=(\d+) experts=(\d+)"), lambda m: int(m.group(1))), + "n_experts": (re.compile(r"layers=(\d+) experts=(\d+)"), lambda m: int(m.group(2))), + # stdout: banner idot kernel + "idot_kernel": (re.compile(r"idot: (\S+) =="), lambda m: m.group(1)), + "cache_cap": (re.compile(r"cache=(\d+) experts/layer"), lambda m: int(m.group(1))), + # stderr: RAM_GB (line 5086-5112) + "ram_budget": (re.compile(r"\[RAM_GB=([\d.]+)"), lambda m: float(m.group(1))), + "cap_lowered": (re.compile(r"cap lowered (\d+)->(\d+)"), lambda m: (int(m.group(1)), int(m.group(2)))), + "cap_raised": (re.compile(r"cap raised (\d+)->(\d+)"), lambda m: (int(m.group(1)), int(m.group(2)))), + "cap_ok": (re.compile(r"cap=(\d+) ok"), lambda m: int(m.group(1))), + # stderr: CUDA (backend_cuda.cu:389) + "cuda_device": (re.compile(r"\[CUDA\] device (\d+): (.*?), ([\d.]+) GB VRAM, sm_(\d)(\d)"), + lambda m: {"id":int(m.group(1)),"name":m.group(2).strip(),"vram_gb":float(m.group(3)), + "sm":f"{m.group(4)}.{m.group(5)}"}), + "cuda_mode": (re.compile(r"\[CUDA\] mode: (.+)"), lambda m: m.group(1)), + "cuda_tier": (re.compile(r"CUDA expert tier: (\d+) resident experts \(([\d.]+) GB\)"), + lambda m: {"resident":int(m.group(1)),"vram_gb":float(m.group(2))}), + # stderr: TOKENS dump (line 4010-4012) + "tokens_dump": (re.compile(r"^\[TOKENS\] (\d+) generated:(.*)$", re.MULTILINE), + lambda m: [int(x) for x in m.group(2).split()]), + # stderr: per-16-token progress (emit_stream line 3742) + "progress_tps": (re.compile(r"t=(\d+)\s+RSS ([\d.]+) GB\s+hit ([\d.]+)%\s+([\d.]+) tok/s\s+([\d.]+) tok/fw"), + lambda m: {"tok":int(m.group(1)),"rss":float(m.group(2)),"hit":float(m.group(3)), + "tps":float(m.group(4)),"tpf":float(m.group(5))}), + # stderr: DSA, USAGE, KV startup lines + "usage_loaded": (re.compile(r"\[USAGE\].*?(\d+) selections"), lambda m: int(m.group(1))), + "kv_slots": (re.compile(r"\[KV\].*?(\d+) context slots"), lambda m: int(m.group(1))), +} + +def extract_metrics(stdout: str, stderr: str) -> dict: + """Parse all metrics from the engine's stdout+stderr output.""" + text_out = stdout or "" + text_err = stderr or "" + metrics = {} + # For most parsers we search BOTH streams (engine is inconsistent about which + # channel a given line lands on). Some are stream-specific (noted below). + combined = text_out + "\n" + text_err + for name, (rx, fn) in RX.items(): + m = rx.search(combined) + if m: + try: metrics[name] = fn(m) + except (ValueError, IndexError): pass + # TOKENS dump is stderr-only and may appear once; grab it explicitly + m = RX["tokens_dump"][0].search(text_err) + if m: + try: metrics["tokens_dump"] = RX["tokens_dump"][1](m) + except: pass + # Multiple CUDA devices — collect all + cuda_devs = [] + for m in re.finditer(r"\[CUDA\] device (\d+): (.*?), ([\d.]+) GB VRAM, sm_(\d)(\d)", text_err): + cuda_devs.append({"id":int(m.group(1)),"name":m.group(2).strip(), + "vram_gb":float(m.group(3)),"sm":f"{m.group(4)}.{m.group(5)}"}) + if cuda_devs: metrics["cuda_devices"] = cuda_devs + # Multiple progress checkpoints — collect the full curve + progress = [] + for m in RX["progress_tps"][0].finditer(text_err): + try: + progress.append(RX["progress_tps"][1](m)) + except: pass + if progress: metrics["progress_curve"] = progress + # Multiple PROFILE lines — there are two (prefill + decode). Keep both. + prof_lines = [(i, line) for i, line in enumerate(text_out.splitlines()) if line.startswith("PROFILE:")] + for idx, (line_no, line) in enumerate(prof_lines): + label = "prefill_profile" if idx == 0 else "decode_profile" + p = {} + for pname, (rx, fn) in RX.items(): + if not pname.startswith("prof_"): continue + mm = rx.search(line) + if mm: + try: p[pname] = fn(mm) + except: pass + if p: metrics[label] = p + return metrics + + +# --------------------------------------------------------------------------- +# ENGINE RUNNER — subprocess wrapper with timeout, signal handling, logging. +# --------------------------------------------------------------------------- +class EngineRunner: + def __init__(self, glm_path, snap, out_dir, default_env=None, timeout=600): + self.glm = str(glm_path) + self.snap = str(snap) + self.out_dir = Path(out_dir) + self.out_dir.mkdir(parents=True, exist_ok=True) + self.default_env = default_env or {} + self.timeout = timeout + self.default_cap = 75 # production default (matches bench_full.sh) + + def run_prompt(self, prompt, ngen=64, env_extra=None, log_name=None, cap=None): + """Run the engine in PROMPT mode and return (stdout, stderr, returncode, elapsed).""" + env = dict(os.environ, SNAP=self.snap, PROMPT=prompt, NGEN=str(ngen)) + env.update(self.default_env) + if env_extra: env.update(env_extra) + env["TOKENS"] = "1" # always capture token ids for reliable text extraction + cmd = [self.glm, str(cap if cap is not None else self.default_cap)] + return self._exec(cmd, env, log_name) + + def run_score(self, score_file, cap=None, env_extra=None, log_name=None): + """Run the engine in SCORE (log-likelihood) mode.""" + env = dict(os.environ, SNAP=self.snap, SCORE=score_file) + env.update(self.default_env) + if env_extra: env.update(env_extra) + cmd = [self.glm, str(cap if cap is not None else self.default_cap)] + return self._exec(cmd, env, log_name) + + def _exec(self, cmd, env, log_name): + t0 = time.time() + log_path = self.out_dir / (log_name or f"run_{int(t0)}.log") + try: + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.PIPE, + text=True, encoding="utf-8", errors="replace") + try: + stdout, stderr = proc.communicate(timeout=self.timeout) + rc = proc.returncode + except subprocess.TimeoutExpired: + proc.kill() + stdout, stderr = proc.communicate() + rc = -1 + stderr = (stderr or "") + f"\n[TIMEOUT after {self.timeout}s]\n" + except Exception as e: + stdout, stderr, rc = "", f"[EXCEPTION] {e}\n{traceback.format_exc()}", -2 + elapsed = time.time() - t0 + # Write raw log (both streams, clearly delimited) + with open(log_path, "w", encoding="utf-8") as f: + f.write(f"=== CMD: {' '.join(cmd)}\n=== ELAPSED: {elapsed:.1f}s\n=== RC: {rc}\n\n") + f.write("--- STDOUT ---\n"); f.write(stdout or ""); f.write("\n") + f.write("--- STDERR ---\n"); f.write(stderr or ""); f.write("\n") + return stdout or "", stderr or "", rc, elapsed + + +# --------------------------------------------------------------------------- +# TEXT RECOVERY — decode the [TOKENS] ids back to text using the tokenizer. +# Falls back to stdout extraction if tokenizer/tokens unavailable. +# --------------------------------------------------------------------------- +def recover_text(stdout, stderr, prompt, tokenizer=None): + """Extract generated text. Primary: TOKENS dump decoded. Fallback: stdout parse.""" + # Method 1: decode TOKENS ids via tokenizer + if tokenizer: + m = re.search(r"^\[TOKENS\] \d+ generated:(.*)$", stderr, re.MULTILINE) + if m: + ids = [int(x) for x in m.group(1).split()] + if ids: + try: + return tokenizer.decode(ids), "tokens" + except Exception: pass + # Method 2: stdout — text is between the prompt string and "PROFILO" or "\n---" + text = stdout or "" + # Find the prompt in stdout, take everything after it up to PROFILO + idx = text.find(prompt) + if idx >= 0: + after = text[idx + len(prompt):] + cut = after.find("PROFILO") + if cut >= 0: + return after[:cut].strip(), "stdout" + cut = after.find("\n---") + if cut >= 0: + return after[:cut].strip(), "stdout" + return "", "none" + + +# --------------------------------------------------------------------------- +# CORRUPTION / COHERENCE CHECKS +# --------------------------------------------------------------------------- +def check_repetition(token_ids): + """Detect degenerate repetition: same token 3+ consecutive times.""" + if not token_ids or len(token_ids) < 6: + return False, 0 + max_run = 1; cur = 1 + for i in range(1, len(token_ids)): + if token_ids[i] == token_ids[i-1]: cur += 1; max_run = max(max_run, cur) + else: cur = 1 + return max_run >= 3, max_run + +def check_expected(text, expect): + """Check if expected substring appears case-insensitively in the first 200 chars.""" + if not expect or not text: return None + return expect.lower() in text[:200].lower() + + +# --------------------------------------------------------------------------- +# PHASES +# --------------------------------------------------------------------------- +class DiagnosticHarness: + def __init__(self, args): + self.args = args + self.snap = args.snap + self.glm = args.glm + ts = datetime.now().strftime("%Y%m%d_%H%M%S") + self.out_dir = Path(args.out or f"./diag_results/{ts}") + self.out_dir.mkdir(parents=True, exist_ok=True) + # Base env for all runs + base_env = {"TEMP": "0", "PIPE": "1", "PIPE_WORKERS": "8", "DIRECT": "1"} + if args.ram: base_env["RAM_GB"] = str(args.ram) + if args.cuda: + base_env.update({ + "COLI_CUDA": "1", + "COLI_GPU": str(args.gpu) if args.gpu is not None else "0", + "CUDA_DENSE": "1", + "COLI_CUDA_ATTN": "1", + "COLI_CUDA_PIPE": "2", + "COLI_CUDA_PIPE_S_MIN": "1", + }) + self.runner = EngineRunner(self.glm, self.snap, self.out_dir, base_env, timeout=args.timeout) + self.runner.default_cap = args.cap + # Load tokenizer for text recovery + self.tokenizer = None + tok_path = os.path.join(self.snap, "tokenizer.json") + if os.path.exists(tok_path): + try: + from tokenizers import Tokenizer + self.tokenizer = Tokenizer.from_file(tok_path) + except Exception as e: + print(f"[warn] could not load tokenizer: {e}", file=sys.stderr) + self.results = {"meta": {"snap": self.snap, "glm": self.glm, + "timestamp": ts, "args": vars(args)}, + "phases": {}} + + def phase_system(self): + """Phase 0: minimal run to capture startup telemetry.""" + print("\n" + "="*60 + "\nPHASE 0: SYSTEM PROBE\n" + "="*60) + stdout, stderr, rc, elapsed = self.runner.run_prompt( + "Hello", ngen=1, log_name="system_probe.log") + metrics = extract_metrics(stdout, stderr) + result = { + "rc": rc, "elapsed": elapsed, + "load_secs": metrics.get("load_secs"), + "resident_mb": metrics.get("resident_mb"), + "n_layers": metrics.get("n_layers"), + "n_experts": metrics.get("n_experts"), + "mtp_status": metrics.get("mtp_status"), + "idot_kernel": metrics.get("idot_kernel"), + "cache_cap_requested": metrics.get("cache_cap"), + "ram_budget_gb": metrics.get("ram_budget"), + "cap_lowered": metrics.get("cap_lowered"), + "cap_raised": metrics.get("cap_raised"), + "cap_final": metrics.get("cap_ok"), + "cuda_devices": metrics.get("cuda_devices", []), + "cuda_mode": metrics.get("cuda_mode"), + } + # Print summary + print(f" load time: {result['load_secs']:.2f}s" if result['load_secs'] else " load time: FAILED") + print(f" resident: {result['resident_mb']:.0f} MB" if result['resident_mb'] else "") + print(f" layers/exp: {result['n_layers']}/{result['n_experts']}" if result['n_layers'] else "") + print(f" MTP: {result['mtp_status']}") + print(f" idot kernel: {result['idot_kernel']}") + print(f" RAM budget: {result['ram_budget_gb']} GB" if result['ram_budget_gb'] else " RAM budget: auto") + if result['cap_lowered']: + print(f" cache cap: {result['cap_lowered'][0]} -> {result['cap_lowered'][1]} (RAM-lowered)") + elif result['cap_final']: + print(f" cache cap: {result['cap_final']} (ok)") + else: + print(f" cache cap: {result['cache_cap_requested']}") + for d in result['cuda_devices']: + print(f" GPU {d['id']}: {d['name']}, {d['vram_gb']:.1f} GB, sm_{d['sm']}") + if rc != 0 and rc != -1: + print(f" WARNING: engine returned rc={rc}") + self.results["phases"]["system"] = result + return result + + def phase_smoke(self): + """Phase 1: correctness smoke test across curated prompts.""" + print("\n" + "="*60 + "\nPHASE 1: CORRECTNESS SMOKE TEST\n" + "="*60) + ngen = self.args.ngen + prompt_results = [] + pass_count = 0; total = len(PROMPTS) + for p in PROMPTS: + stdout, stderr, rc, elapsed = self.runner.run_prompt( + p["prompt"], ngen=ngen, log_name=f"smoke_{p['id']}.log") + metrics = extract_metrics(stdout, stderr) + token_ids = metrics.get("tokens_dump", []) + text, method = recover_text(stdout, stderr, p["prompt"], self.tokenizer) + is_rep, max_run = check_repetition(token_ids) + expect_ok = check_expected(text, p.get("expect")) + # Verdict: pass if text is non-empty, no severe repetition, and (if factual) expected found + has_text = bool(text.strip()) + ok = has_text and not is_rep + if p.get("expect") and expect_ok is False: ok = False + if ok: pass_count += 1 + # Diagnose failure mode for reporting + if not has_text: + fail_reason = "no tokens generated (immediate EOS)" + elif is_rep: + fail_reason = f"repetition loop (max run={max_run})" + elif p.get("expect") and expect_ok is False: + fail_reason = f"expected '{p['expect']}' not found" + else: + fail_reason = "" + prompt_results.append({ + "id": p["id"], "cat": p["cat"], "prompt": p["prompt"], + "expect": p.get("expect"), "generated": text[:300], + "expect_match": expect_ok, "repetition": is_rep, "max_run": max_run, + "toks_generated": len(token_ids), "decode_tps": metrics.get("decode_tps"), + "hit_rate": metrics.get("hit_rate"), "rc": rc, "elapsed": elapsed, + "text_method": method, "pass": ok, "fail_reason": fail_reason, + }) + status = "PASS" if ok else "FAIL" + extra = f" expect={'Y' if expect_ok else ('N' if expect_ok is False else '-')}" if p.get("expect") else "" + tps = f" {metrics.get('decode_tps',0):.2f}t/s" if metrics.get("decode_tps") else "" + print(f" [{status}] {p['id']:<22} ({p['cat']:<11}) rep={'Y' if is_rep else 'N'}{extra}{tps}") + if text: + preview = text.replace("\n", " ")[:80] + print(f" -> {preview}") + elif rc != 0: + print(f" -> [engine rc={rc}]") + else: + print(f" -> [{fail_reason}]") + result = {"prompts": prompt_results, "pass": pass_count, "total": total, + "pass_rate": 100.0 * pass_count / total if total else 0} + print(f"\n SMOKE SUMMARY: {pass_count}/{total} passed ({result['pass_rate']:.0f}%)") + self.results["phases"]["smoke"] = result + return result + + def phase_diagnostic(self): + """Phase 2: single deep-instrumented run — full PROFILE + routing + MTP.""" + print("\n" + "="*60 + "\nPHASE 2: FULL SYSTEM DIAGNOSTIC\n" + "="*60) + diag_env = { + "LOOKA": "1", "DISK_SPLIT": "1", "ROUTE_AGREE": "1", + } + if self.args.cuda: + diag_env["COLI_CUDA_PROFILE"] = "1" + prompt = ("Write a short paragraph explaining how photosynthesis works. " + "Include the roles of sunlight, water, and carbon dioxide.") + stdout, stderr, rc, elapsed = self.runner.run_prompt( + prompt, ngen=self.args.ngen, env_extra=diag_env, log_name="diagnostic.log") + metrics = extract_metrics(stdout, stderr) + text, _ = recover_text(stdout, stderr, prompt, self.tokenizer) + result = { + "rc": rc, "elapsed": elapsed, + "generated_text": text[:500], + "prefill_profile": metrics.get("prefill_profile", {}), + "decode_profile": metrics.get("decode_profile", {}), + "decode_tps": metrics.get("decode_tps"), + "prefill_secs": metrics.get("prefill_secs"), + "hit_rate": metrics.get("hit_rate"), + "rss_gb": metrics.get("rss_gb"), + "experts_per_tok": metrics.get("experts_per_tok"), + "mtp_accept": metrics.get("mtp_accept"), + "mtp_counts": metrics.get("mtp_acc_cnt"), + "spec_tok_per_fw": metrics.get("spec_tok_per_fw"), + "cuda_tier": metrics.get("cuda_tier"), + "progress_curve": metrics.get("progress_curve", []), + } + # Print the PROFILE breakdown + dp = result["decode_profile"] + print(f"\n DECODE TIMING BREAKDOWN (per-bucket, seconds):") + print(f" expert-disk: {dp.get('prof_expert_disk', '?'):>8}") + print(f" expert-matmul: {dp.get('prof_expert_mm', '?'):>8}") + print(f" attention: {dp.get('prof_attention', '?'):>8}") + print(f" lm_head: {dp.get('prof_lm_head', '?'):>8}") + print(f" other: {dp.get('prof_other', '?'):>8}") + print(f"\n PERFORMANCE:") + print(f" prefill: {result['prefill_secs']:.2f}s" if result['prefill_secs'] else " prefill: ?") + print(f" decode: {result['decode_tps']:.3f} tok/s" if result['decode_tps'] else " decode: ?") + print(f" hit rate: {result['hit_rate']:.1f}%" if result.get('hit_rate') is not None else " hit rate: ?") + print(f" RSS: {result['rss_gb']:.1f} GB" if result.get('rss_gb') else " RSS: ?") + print(f" exp/tok: {result['experts_per_tok']:.1f}" if result.get('experts_per_tok') else " exp/tok: ?") + print(f" MTP: {result['mtp_accept']:.0f}% accept" if result.get('mtp_accept') is not None else " MTP: ?") + if result.get('cuda_tier'): + ct = result['cuda_tier'] + print(f" CUDA tier: {ct['resident']} experts, {ct['vram_gb']:.1f} GB VRAM") + # Show generated text preview + if text: + print(f"\n GENERATED TEXT (first 200 chars):") + print(f" {text[:200].replace(chr(10), ' ')}") + else: + print(f"\n GENERATED TEXT: [none recovered]") + self.results["phases"]["diagnostic"] = result + return result + + def phase_quality(self): + """Phase 3: benchmark accuracy via eval_glm.py SCORE mode.""" + print("\n" + "="*60 + "\nPHASE 3: QUALITY BENCHMARKS\n" + "="*60) + eval_script = os.path.join(os.path.dirname(__file__), "eval_glm.py") + bench_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), "bench") + if not os.path.exists(eval_script): + print(f" [SKIP] eval_glm.py not found at {eval_script}") + self.results["phases"]["quality"] = {"error": "eval_glm.py not found"} + return None + tasks = ["hellaswag", "arc_challenge", "mmlu"] + missing = [t for t in tasks if not os.path.exists(os.path.join(bench_dir, f"{t}.jsonl"))] + if missing: + print(f" [SKIP] benchmark data missing: {missing}") + print(f" run: python tools/fetch_benchmarks.py --out {bench_dir}") + self.results["phases"]["quality"] = {"error": f"missing benchmark data: {missing}"} + return None + limit = self.args.quality_limit + py = sys.executable + cmd = [py, eval_script, "--snap", self.snap, "--glm", self.glm, + "--data", bench_dir, "--tasks", ",".join(tasks), "--limit", str(limit)] + env = dict(os.environ) + if self.args.ram: env["RAM_GB"] = str(self.args.ram) + print(f" Running eval_glm.py (tasks={tasks}, n={limit})...") + print(f" This takes ~{limit*3*4/0.05:.0f}s at 0.05 tok/s (worst case)...") + t0 = time.time() + log_path = self.out_dir / "quality_eval.log" + try: + with open(log_path, "w", encoding="utf-8") as logf: + proc = subprocess.run(cmd, env=env, capture_output=True, text=True, timeout=self.args.timeout*3, + encoding="utf-8", errors="replace") + logf.write(proc.stdout); logf.write("\n--- STDERR ---\n"); logf.write(proc.stderr) + elapsed = time.time() - t0 + # Parse the acc/acc_norm table from eval_glm.py output + scores = {} + for line in proc.stdout.splitlines(): + # lines like: "hellaswag 40 45.0% 50.0%" + m = re.match(r"(\w+)\s+(\d+)\s+([\d.]+)%\s+([\d.]+)%", line.strip()) + if m: + scores[m.group(1)] = {"n": int(m.group(2)), "acc": float(m.group(3)), + "acc_norm": float(m.group(4))} + mean_m = re.search(r"MEAN acc_norm:\s*([\d.]+)%", proc.stdout) + result = {"scores": scores, "mean_acc_norm": float(mean_m.group(1)) if mean_m else None, + "elapsed": elapsed, "rc": proc.returncode} + print(f" Completed in {elapsed:.0f}s\n") + print(f" {'task':<18} {'n':>4} {'acc':>7} {'acc_norm':>9}") + for t, s in scores.items(): + print(f" {t:<18} {s['n']:>4} {s['acc']:>6.1f}% {s['acc_norm']:>8.1f}%") + if result["mean_acc_norm"] is not None: + print(f"\n MEAN acc_norm: {result['mean_acc_norm']:.1f}%") + except subprocess.TimeoutExpired: + result = {"error": f"eval timed out after {self.args.timeout*3}s"} + print(f" [TIMEOUT] eval_glm.py exceeded {self.args.timeout*3}s") + except Exception as e: + result = {"error": str(e)} + print(f" [ERROR] {e}") + self.results["phases"]["quality"] = result + return result + + def phase_throughput(self): + """Phase 4: tok/s comparison — MTP on vs off.""" + print("\n" + "="*60 + "\nPHASE 4: THROUGHPUT BENCHMARK\n" + "="*60) + prompt = "Summarize the plot of Romeo and Juliet in three sentences." + ngen = self.args.ngen + results = {} + # Run 1: with MTP (default) + print(f" [1/2] MTP ON (draft=3)...") + stdout, stderr, rc, t = self.runner.run_prompt( + prompt, ngen=ngen, log_name="throughput_mtp_on.log") + m_on = extract_metrics(stdout, stderr) + results["mtp_on"] = {"tps": m_on.get("decode_tps"), "hit": m_on.get("hit_rate"), + "mtp_accept": m_on.get("mtp_accept"), "elapsed": t} + print(f" {m_on.get('decode_tps',0):.3f} tok/s | hit {m_on.get('hit_rate',0):.1f}% | " + f"MTP {m_on.get('mtp_accept',0):.0f}%") + # Run 2: MTP off + print(f" [2/2] MTP OFF (MTP=0)...") + stdout, stderr, rc, t = self.runner.run_prompt( + prompt, ngen=ngen, env_extra={"MTP": "0"}, log_name="throughput_mtp_off.log") + m_off = extract_metrics(stdout, stderr) + results["mtp_off"] = {"tps": m_off.get("decode_tps"), "hit": m_off.get("hit_rate"), + "elapsed": t} + print(f" {m_off.get('decode_tps',0):.3f} tok/s | hit {m_off.get('hit_rate',0):.1f}%") + # Compute speedup + if results["mtp_on"]["tps"] and results["mtp_off"]["tps"] and results["mtp_off"]["tps"] > 0: + sp = results["mtp_on"]["tps"] / results["mtp_off"]["tps"] + results["mtp_speedup"] = sp + print(f"\n MTP speedup: {sp:.2f}x ({'MTP helps' if sp > 1.05 else 'MTP hurts' if sp < 0.95 else 'no effect'})") + else: + print(f"\n MTP speedup: (insufficient data)") + self.results["phases"]["throughput"] = results + return results + + def write_report(self): + """Write report.json and report.md.""" + ts = self.results["meta"]["timestamp"] + json_path = self.out_dir / "report.json" + md_path = self.out_dir / "report.md" + # JSON + with open(json_path, "w", encoding="utf-8") as f: + json.dump(self.results, f, indent=2, default=str) + # Markdown + lines = [ + f"# Diagnostic Report — {ts}", + f"", + f"**Model:** `{self.snap}`", + f"**Engine:** `{self.glm}`", + f"**Date:** {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}", + f"", + ] + phases = self.results["phases"] + # System + if "system" in phases: + s = phases["system"] + lines += ["## Phase 0: System Probe", ""] + lines += [f"- Load time: **{s.get('load_secs','?')}s**"] + lines += [f"- Layers/experts: {s.get('n_layers','?')}/{s.get('n_experts','?')}"] + lines += [f"- MTP: {s.get('mtp_status','?')}"] + lines += [f"- idot kernel: `{s.get('idot_kernel','?')}`"] + lines += [f"- RAM budget: {s.get('ram_budget_gb','auto')} GB"] + if s.get("cap_lowered"): + lines += [f"- Cache cap: {s['cap_lowered'][0]}→{s['cap_lowered'][1]} (RAM-lowered)"] + elif s.get("cap_final"): + lines += [f"- Cache cap: {s['cap_final']}"] + for d in s.get("cuda_devices", []): + lines += [f"- GPU {d['id']}: {d['name']}, {d['vram_gb']:.1f} GB, sm_{d['sm']}"] + lines.append("") + # Smoke + if "smoke" in phases: + sm = phases["smoke"] + lines += [f"## Phase 1: Correctness Smoke", "", + f"**{sm['pass']}/{sm['total']} prompts passed ({sm['pass_rate']:.0f}%)**", "", + "| ID | Category | Pass | Expect | Repetition | tok/s | Generated (first 80 chars) |", + "|---|---|---|---|---|---|---|"] + for p in sm["prompts"]: + gen = p["generated"][:80].replace("|", "\\|").replace("\n", " ") if p["generated"] else "" + exp = "Y" if p.get("expect_match") else ("N" if p.get("expect_match") is False else "-") + tps = f"{p.get('decode_tps',0):.2f}" if p.get("decode_tps") else "-" + lines.append(f"| {p['id']} | {p['cat']} | {'✅' if p['pass'] else '❌'} | {exp} | " + f"{'⚠️' if p['repetition'] else '-'} (run={p.get('max_run',0)}) | {tps} | {gen} |") + lines.append("") + # Diagnostic + if "diagnostic" in phases: + d = phases["diagnostic"] + lines += ["## Phase 2: Full System Diagnostic", ""] + dp = d.get("decode_profile", {}) + lines += ["### Decode Timing Breakdown", "", + "| Bucket | Seconds |", "|---|---|"] + for k, label in [("prof_expert_disk","expert-disk"),("prof_expert_mm","expert-matmul"), + ("prof_attention","attention"),("prof_lm_head","lm_head"), + ("prof_other","other")]: + v = dp.get(k, "?") + lines.append(f"| {label} | {v} |") + lines += [f"", f"### Performance", f"- Decode: **{d.get('decode_tps','?')} tok/s**", + f"- Prefill: {d.get('prefill_secs','?')}s", + f"- Expert hit rate: {d.get('hit_rate','?')}%", + f"- RSS: {d.get('rss_gb','?')} GB", + f"- Experts/token: {d.get('experts_per_tok','?')}", + f"- MTP acceptance: {d.get('mtp_accept','?')}%", ""] + if d.get("generated_text"): + lines += [f"### Generated Text", f"```\n{d['generated_text'][:300]}\n```", ""] + # Quality + if "quality" in phases: + q = phases["quality"] + lines += ["## Phase 3: Quality Benchmarks", ""] + if "error" in q: + lines += [f"⚠️ {q['error']}", ""] + else: + lines += [f"**Mean acc_norm: {q.get('mean_acc_norm','?')}%**", "", + "| Task | n | acc | acc_norm |", "|---|---|---|---|"] + for t, s in q.get("scores", {}).items(): + lines.append(f"| {t} | {s['n']} | {s['acc']:.1f}% | {s['acc_norm']:.1f}% |") + lines.append("") + # Throughput + if "throughput" in phases: + th = phases["throughput"] + lines += ["## Phase 4: Throughput", "", + "| Mode | tok/s | hit% | MTP accept |", "|---|---|---|---|"] + on = th.get("mtp_on", {}); off = th.get("mtp_off", {}) + lines.append(f"| MTP ON | {on.get('tps','?')} | {on.get('hit','?')}% | {on.get('mtp_accept','?')}% |") + lines.append(f"| MTP OFF | {off.get('tps','?')} | {off.get('hit','?')}% | — |") + if th.get("mtp_speedup"): + lines.append(f"\n**MTP speedup: {th['mtp_speedup']:.2f}x**") + lines.append("") + with open(md_path, "w", encoding="utf-8") as f: + f.write("\n".join(lines)) + print(f"\n{'='*60}") + print(f"Report written:") + print(f" JSON: {json_path}") + print(f" MD: {md_path}") + print(f" Logs: {self.out_dir}/*.log") + print(f"{'='*60}") + + def run(self): + phase = self.args.phase + if phase == "all": + self.phase_system() + self.phase_smoke() + self.phase_diagnostic() + self.phase_quality() + self.phase_throughput() + elif phase == "system": self.phase_system() + elif phase == "smoke": self.phase_smoke() + elif phase == "diagnostic": self.phase_diagnostic() + elif phase == "quality": self.phase_quality() + elif phase == "throughput": self.phase_throughput() + else: + print(f"Unknown phase: {phase}", file=sys.stderr); sys.exit(1) + self.write_report() + + +def main(): + ap = argparse.ArgumentParser(description="Comprehensive model diagnostic harness for colibri GLM-5.2") + ap.add_argument("--snap", required=True, help="model snapshot directory") + ap.add_argument("--glm", default=None, help="engine binary path (default: ./glm.exe or ./glm)") + ap.add_argument("--phase", default="all", + choices=["all","system","smoke","diagnostic","quality","throughput"], + help="which test phase to run") + ap.add_argument("--out", default=None, help="output directory (default: ./diag_results/)") + ap.add_argument("--ngen", type=int, default=64, help="generation length for smoke/throughput (default 64)") + ap.add_argument("--quality-limit", type=int, default=40, help="questions per benchmark task (default 40)") + ap.add_argument("--ram", type=float, default=0, help="RAM_GB override (0=auto)") + ap.add_argument("--cuda", action="store_true", help="enable COLI_CUDA GPU tier") + ap.add_argument("--gpu", type=int, default=None, help="GPU device ordinal (with --cuda)") + ap.add_argument("--cap", type=int, default=75, help="experts-per-layer cache cap (default 75)") + ap.add_argument("--timeout", type=int, default=600, help="per-run timeout in seconds (default 600)") + a = ap.parse_args() + # Auto-detect engine binary + if a.glm is None: + for cand in ["./glm.exe", "./glm", "../glm.exe", os.path.join(os.path.dirname(__file__), "..", "glm.exe")]: + if os.path.exists(cand): a.glm = os.path.abspath(cand); break + if a.glm is None: + print("ERROR: could not find glm/glm.exe. Specify with --glm", file=sys.stderr); sys.exit(1) + if not os.path.isdir(a.snap): + print(f"ERROR: snapshot dir does not exist: {a.snap}", file=sys.stderr); sys.exit(1) + harness = DiagnosticHarness(a) + harness.run() + +if __name__ == "__main__": + main() diff --git a/c/tools/efficiency.py b/c/tools/efficiency.py new file mode 100644 index 0000000..775e8cf --- /dev/null +++ b/c/tools/efficiency.py @@ -0,0 +1,458 @@ +"""Efficiency / regression harness for the colibri engine. + +The engine already emits rich telemetry (REPLAY tok/s, PROFILE phase timings, +[PROF] time shares + verdict, CUDA expert-tier utilization). Until now every +consumer of that telemetry — `benchmark_cuda_fixture.py`, `bench_full.sh`, +`bench_ux.sh` — has only *printed* it for a human to eyeball. This module turns +each signal into a parseable field so tests can assert on it. + +Design: + - Reuses SPEED_RE / PROFILE_RE from tools.benchmark_cuda_fixture (no drift). + - parse_run() is pure: stdout+stderr in, dict out. Easy to unit-test against + captured strings (like the existing test_benchmark_cuda_fixture does). + - run_engine() is the subprocess wrapper. Captures stdout and stderr + separately, because the engine splits them: PROFILE/REPLAY/CUDA-tier go to + stdout, the [CUDA]/[PROF]/[prefill] banners go to stderr. + - Floor defaults are module constants (tunable in one place, not scattered). + +No model file is required to import this module; only run_engine() invokes the +binary. parse_run() works on any captured text, so most test surface is covered +by string fixtures without spinning the engine at all. +""" +from __future__ import annotations + +import os +import re +import subprocess +from pathlib import Path +from typing import Optional + +# Reuse the validated regexes from the existing A/B benchmark harness so the +# PROFILE field order (disk, expert_matmul, attention, lm_head, other) and the +# tok/s capture stay identical. Drift here would silently break every consumer. +from tools.benchmark_cuda_fixture import SPEED_RE as _SPEED_RE_REPLAY, PROFILE_RE, PROFILE_KEYS + +# SPEED_RE (from benchmark_cuda_fixture) matches the REPLAY-mode line only: +# "REPLAY decode: ... | 12.34 tok/s | ..." +# run_text / PROMPT mode uses a DIFFERENT format (glm.c:4682): +# "decode N tokens in X.XXs (12.34 tok/s) | expert hit rate ..." +# This alt regex catches the parenthesized form so the full-model report (which +# uses PROMPT mode) gets a real tok/s instead of reporting it missing. +SPEED_RE_TEXT = re.compile(r"decode \d+ tokens in [0-9.]+s \(([0-9.]+) tok/s\)") + + +def _first_speed(stdout: str): + """Find tok/s in whichever run-mode format the engine used.""" + for rx in (_SPEED_RE_REPLAY, SPEED_RE_TEXT): + m = rx.search(stdout) + if m: + return m + return None + + +# Public alias so existing imports keep working (tests reference SPEED_RE). +SPEED_RE = _SPEED_RE_REPLAY + + +# --- additional parsers (formats verified against glm.c printf strings) --- + +# "expert hit rate 88.1%" (summary line) | "expert hit 95.0%" (REPLAY line) +HIT_RE = re.compile(r"expert hit(?:\s+rate)?\s+([0-9.]+)%") + +# "[PROF] time shares: expert-I/O 3% | expert-matmul 12% | attention 56% | lm_head 2% | other 27%" +SHARES_RE = re.compile( + r"\[PROF\] time shares: expert-I/O\s+([0-9.]+)%\s*\|\s*expert-matmul\s+([0-9.]+)%\s*" + r"\|\s*attention\s+([0-9.]+)%\s*\|\s*lm_head\s+([0-9.]+)%\s*\|\s*other\s+([0-9.]+)%" +) + +# "[PROF] verdict: I/O-bound — 60% of the time ..." (also compute-bound / attention-bound / balanced) +VERDICT_RE = re.compile(r"\[PROF\] verdict:\s*(I/O-bound|compute-bound|attention-bound|balanced)") + +# "[PROF] expert I/O: ... hit 95.0% (76 hit / 4 load) | 4.0 loads/token" +LOADS_PER_TOK_RE = re.compile(r"\|\s*([0-9.]+)\s+loads/token") + +# "CUDA expert tier: 111 resident experts (2.36 GB) | 5400 calls served from VRAM" (stdout) +CUDA_TIER_RE = re.compile( + r"CUDA expert tier:\s+(\d+)\s+resident experts\s+\(([0-9.]+)\s+GB\)\s*\|\s+(\d+)\s+calls served from VRAM" +) + +# "[CUDA] device 0: NVIDIA ..., 14.4 GB VRAM, sm_120" (stderr, per device at init) +CUDA_DEVICE_RE = re.compile(r"\[CUDA\] device\s+\d+:") + +# "[CUDA] resident set: 12 tensors, 0.45 GB VRAM" (stderr, cuda_stats_print) +CUDA_RESIDENT_RE = re.compile( + r"\[CUDA\] resident set:\s+(\d+)\s+tensors,\s+([0-9.]+)\s+GB\s+VRAM" +) + +# "PREFILL (teacher-forcing) C vs oracle: 11/32 positions | 1700.4 pos/s" (TF=1 mode, stdout) +TF_MATCH_RE = re.compile(r"PREFILL \(teacher-forcing\).*:\s+(\d+)/(\d+)\s+positions") + +# --- the six deeper signals (added after "are you gathering everything?" audit) --- + +# "ATTENTION: projection/RoPE 0.050s | score-softmax-value 0.009s | output projection 0.011s" +# Sub-breakdown of the attention phase — answers "how is attention being read". +ATTN_BREAKDOWN_RE = re.compile( + r"ATTENTION: projection/RoPE\s+([0-9.]+)s\s*\|\s*score-softmax-value\s+([0-9.]+)s\s*" + r"\|\s*output projection\s+([0-9.]+)s" +) + +# "[PROF] decode forwards: 20 | latency p50 7.3 ms | p90 8.0 ms | p99 8.1 ms | max 8.2 ms | 1.00 tok/forward" +# Per-forward tail latency — a p99 >> p50 means decode stalls (I/O hiccups, KV grow). +LATENCY_RE = re.compile( + r"\[PROF\] decode forwards:\s+(\d+)\s*\| latency p50\s+([0-9.]+)\s*ms\s*\| " + r"p90\s+([0-9.]+)\s*ms\s*\|\s*p99\s+([0-9.]+)\s*ms\s*\|\s*max\s+([0-9.]+)\s*ms" +) + +# "[PROF] expert I/O: 0.004 GB fetched (0.2 MB/token, 0.03 GB/s over the run) | hit 95.0% ... | +# 4.0 loads/token | 0.0s read service / 0.0s felt wait" +# Absolute disk throughput + the felt-wait split (the [PROF] version, more detailed than PROFILE). +EXPERT_IO_RE = re.compile( + r"\[PROF\] expert I/O:\s+([0-9.]+)\s+GB fetched\s+\(([0-9.]+)\s+MB/token,\s*([0-9.]+)\s+GB/s" + r".*?\|\s*([0-9.]+)\s+loads/token\s*\|\s*([0-9.]+)s\s+read service\s*/\s*([0-9.]+)s\s+felt wait" +) + +# "speculation: 1.05 tokens/forward (19 forwards per 20 tokens) | MTP acceptance 44% (7/16)" +# Draft efficiency — is the speculative decoder pulling weight or dead overhead? +SPECULATION_RE = re.compile( + r"speculation:\s+([0-9.]+)\s+tokens/forward\s+\((\d+)\s+forwards per\s+(\d+)\s+tokens\)" + r"\s*\|\s*MTP acceptance\s+([0-9.]+)%" +) + +# "experts loaded/token: 450.0 (per-layer 56.25 across 8; baseline topk=8) | TOPK=0 TOPP=0.00" +# Fuller than loads_per_tok: includes the per-layer spread + the active topk/topp. +EXPERTS_LOADED_RE = re.compile( + r"experts loaded/token:\s+([0-9.]+)\s+\(per-layer\s+([0-9.]+)\s+across\s+(\d+);\s*baseline topk=(\d+)\)" +) + +# "[PROF] machine: Intel(...) | 22 cores (22 omp threads) | RAM 34.1 GB total, 27.1 GB available | backend CUDA" +# Provenance — makes a report reproducible across machines/runs. +MACHINE_RE = re.compile(r"\[PROF\] machine:\s*(.*)\|\s*(\d+)\s+cores.*backend\s+(\S+)") + +# "[PROF] config: RAM_GB=auto 23.9 CTX=4096 | expert cache cap 8/layer ... | DRAFT=0 PIPE=1 DIRECT=0 ..." +# Effective resolved config (after auto-budgeting). Answers "what config actually ran". +CONFIG_RE = re.compile(r"\[PROF\] config:\s*(.*)") + +# "[ORACLE] mismatch pos=7 expected=197 got=22" (TF=1 mode, stderr, per position) +# Captures the engine's *actual* argmax at each position, so two backends can be +# compared DIRECTLY (independent of how each relates to the oracle). +TF_MISMATCH_RE = re.compile(r"\[ORACLE\] mismatch pos=(\d+) expected=\d+ got=(\d+)") + +# --- routing-quality + disk-split + cuda-groups (the deep signals) --- + +# summary suffix: " | swap 3.1% (12/384)" (CACHE_ROUTE inline) +SWAP_RE = re.compile(r"swap\s+([0-9.]+)%\s+\((\d+)/(\d+)\)") + +# summary suffix: " | route_agree 85.0% | route_kl 0.0123" (CACHE_ROUTE/ROUTE_AGREE) +ROUTE_AGREE_RE = re.compile(r"route_agree\s+([0-9.]+)%\s*\|\s*route_kl\s+([0-9.]+)") + +# "disk-load split: draft 8 + absorb 0 + verify/main 3 misses | MTP-layer 0 loads 0.00 GB | +# main-layers 11 loads 0.01 GB (MTP 0.0% of bytes)" (DISK_SPLIT=1) +DISK_SPLIT_RE = re.compile( + r"disk-load split: draft\s+(\d+)\s+\+\s+absorb\s+(\d+)\s+\+\s+verify/main\s+(\d+)\s+misses" + r"\s*\|\s*MTP-layer\s+(\d+)\s+loads\s+([0-9.]+)\s+GB\s*\|\s*main-layers\s+(\d+)\s+loads\s+([0-9.]+)\s+GB" + r"(?:\s+\(MTP\s+([0-9.]+)%\s+of bytes\))?" +) + +# "[CUDA] expert groups: 120 call, 840 expert, 1200 righe (7.00 expert/call)" +CUDA_GROUPS_RE = re.compile( + r"\[CUDA\] expert groups:\s+(\d+)\s+call,\s+(\d+)\s+expert,\s+(\d+)\s+righe\s+\(([0-9.]+)\s+expert/call\)" +) + +# "[CUDA] expert groups timing: H2D 12.3 ms | kernel 45.6 ms | D2H 7.8 ms" (COLI_CUDA_PROFILE=1) +CUDA_GROUPS_TIME_RE = re.compile( + r"\[CUDA\] expert groups timing: H2D\s+([0-9.]+)\s+ms\s*\|\s*kernel\s+([0-9.]+)\s+ms\s*\|\s*D2H\s+([0-9.]+)\s+ms" +) + +# LOOKAHEAD recall block — 4 named rows. (name, pct, hit, tot) +LOOKAHEAD_RE = re.compile( + r"^\s*(.+?)\s+([0-9.]+)%\s+\((\d+)/(\d+)\)\s*$", re.MULTILINE +) + +# "loaded in 0.02s | resident dense: 0.21 MB | layers=5 experts=8 | MTP absent (draft=0)" +LOAD_BANNER_RE = re.compile( + r"loaded in\s+([0-9.]+)s\s*\|\s*resident dense:\s+([0-9.]+)\s+MB\s*\|" + r"\s*layers=(\d+)\s+experts=(\d+)\s*\|\s*MTP\s+(\w+)\s+\(draft=(\d+)\)" +) + + +# --- tunable floors ----------------------------------------------------------- +# These are deliberately generous so they catch *regressions* (broken builds, +# pathological configs, telemetry accounting bugs) without flapping on machine +# noise. The tiny model is fully resident at ~200 tok/s, so a 20 tok/s floor is +# a 10x margin. Tune per-host via env if needed (documented in README). +TINY_TOK_S_FLOOR = float(os.environ.get("COLI_TINY_TOK_S_FLOOR", "20.0")) +# On a fully-resident tiny model the expert-disk wait share should be tiny. +# If it exceeds this, something regressed in the I/O accounting or cache path. +MAX_DISK_WAIT_SHARE = float(os.environ.get("COLI_MAX_DISK_WAIT_SHARE", "0.20")) +# decode wall-time should be roughly the sum of PROFILE phases (other = residual). +PROFILE_SUM_TOLERANCE = float(os.environ.get("COLI_PROFILE_SUM_TOL", "0.05")) +# Minimum direct CPU-vs-CUDA argmax agreement on the tiny TF fixture. The two +# backends use different accumulation orders (SIMD dot vs CUDA kernel), so a +# few near-tied positions flip argmax — that's expected numeric divergence, not +# a kernel bug. Measured baseline ~84% (27/32) on this fixture; the 70% floor +# leaves headroom for machine noise while still catching a catastrophic kernel +# regression (e.g. a wrong GEMM would drop this to ~random = ~4%). +MIN_CPU_CUDA_AGREEMENT = float(os.environ.get("COLI_MIN_CPU_CUDA_AGREE", "0.70")) + + +def parse_run(stdout: str, stderr: str = "") -> dict: + """Parse one engine run's output into a telemetry dict. + + Returns keys: tok_s, hit_pct, profile (dict, seconds), profile_sum, + time_shares (dict, fractions 0..1), verdict, loads_per_tok, cuda (dict), + tf_match (tuple or None), parsed (set of field names found). + + Raises RuntimeError only if the core throughput line is missing — everything + else is optional and absent on some run modes (e.g. [PROF] needs PROF=1, + CUDA tier needs gpu_expert_count>0). + """ + out = dict( + tok_s=None, hit_pct=None, profile=None, profile_sum=None, + time_shares=None, verdict=None, loads_per_tok=None, + cuda=None, tf_match=None, stderr=stderr, + ) + parsed = set() + blob = stdout + "\n" + stderr # [PROF]/[CUDA] live on stderr; scan both. + + m = _first_speed(stdout) + if m: + out["tok_s"] = float(m.group(1)); parsed.add("tok_s") + + m = HIT_RE.search(blob) + if m: + out["hit_pct"] = float(m.group(1)); parsed.add("hit_pct") + + m = PROFILE_RE.search(stdout) + if m: + service, wait, emm, attn, head, other = (float(x) for x in m.groups()) + disk = service + (wait or 0.0) + out["profile"] = dict(zip(PROFILE_KEYS, (disk, emm, attn, head, other))) + out["profile_sum"] = disk + emm + attn + head + other + parsed.add("profile") + + # ATTENTION sub-breakdown: projection/RoPE | score-softmax-value | output. + m = ATTN_BREAKDOWN_RE.search(stdout) + if m: + out["attn_breakdown"] = dict(zip( + ("proj_rope", "score_sm_value", "out_proj"), + (float(x) for x in m.groups()))) + parsed.add("attn_breakdown") + + m = SHARES_RE.search(blob) + if m: + io, emm, attn, head, other = (float(x) / 100.0 for x in m.groups()) + out["time_shares"] = dict(io=io, matmul=emm, attention=attn, head=head, other=other) + parsed.add("time_shares") + + m = VERDICT_RE.search(blob) + if m: + out["verdict"] = m.group(1); parsed.add("verdict") + + # [PROF] decode forwards + latency p50/p90/p99/max. + m = LATENCY_RE.search(blob) + if m: + out["latency"] = dict(zip( + ("forwards", "p50_ms", "p90_ms", "p99_ms", "max_ms"), + (float(x) for x in m.groups()))) + parsed.add("latency") + + # [PROF] expert I/O throughput: GB fetched, MB/token, GB/s, service vs felt wait. + m = EXPERT_IO_RE.search(blob) + if m: + out["expert_io"] = dict(zip( + ("gb_fetched", "mb_per_tok", "gb_per_s", "loads_per_tok", + "read_service_s", "felt_wait_s"), + (float(x) for x in m.groups()))) + parsed.add("expert_io") + + m = LOADS_PER_TOK_RE.search(blob) + if m: + out["loads_per_tok"] = float(m.group(1)); parsed.add("loads_per_tok") + + # experts loaded/token with per-layer spread + baseline topk (run_text summary). + m = EXPERTS_LOADED_RE.search(stdout) + if m: + out["experts_loaded"] = dict( + per_tok=float(m.group(1)), per_layer=float(m.group(2)), + n_sparse_layers=int(m.group(3)), baseline_topk=int(m.group(4))) + parsed.add("experts_loaded") + + # speculation: tokens/forward, forwards, tokens, MTP acceptance%. + m = SPECULATION_RE.search(stdout) + if m: + out["speculation"] = dict(zip( + ("tok_per_fw", "forwards", "tokens", "mtp_accept_pct"), + (float(m.group(1)), int(m.group(2)), int(m.group(3)), float(m.group(4))))) + parsed.add("speculation") + + # routing quality (CACHE_ROUTE inline suffixes on the summary line). + m = SWAP_RE.search(stdout) + if m: + out["swap"] = dict(pct=float(m.group(1)), swaps=int(m.group(2)), slots=int(m.group(3))) + parsed.add("swap") + m = ROUTE_AGREE_RE.search(stdout) + if m: + out["route_agree"] = dict(agree_pct=float(m.group(1)), kl=float(m.group(2))) + parsed.add("route_agree") + + # disk-load split by decode phase (DISK_SPLIT=1). + m = DISK_SPLIT_RE.search(stdout) + if m: + out["disk_split"] = dict(zip( + ("draft", "absorb", "verify_main", "mtp_loads", "mtp_gb", + "main_loads", "main_gb", "mtp_bytes_pct"), + (int(m.group(1)), int(m.group(2)), int(m.group(3)), int(m.group(4)), + float(m.group(5)), int(m.group(6)), float(m.group(7)), + float(m.group(8)) if m.group(8) else None))) + parsed.add("disk_split") + + # provenance: machine + resolved config. + m = MACHINE_RE.search(blob) + if m: + out["machine"] = dict(cpu=m.group(1).strip(), cores=int(m.group(2)), backend=m.group(3)) + parsed.add("machine") + m = CONFIG_RE.search(blob) + if m: + out["config_str"] = m.group(1).strip(); parsed.add("config") + + # load banner: load time, resident dense MB, layers, experts, MTP status. + m = LOAD_BANNER_RE.search(stdout) + if m: + out["load"] = dict(zip( + ("load_s", "resident_dense_mb", "layers", "experts", "mtp_status", "draft"), + (float(m.group(1)), float(m.group(2)), int(m.group(3)), + int(m.group(4)), m.group(5), int(m.group(6))))) + parsed.add("load") + + # LOOKAHEAD routing-recall block (LOOKA=1): list of {predictor, pct, hit, tot}. + la_block = re.search( + r"LOOKAHEAD routing.*?recall.*?:\n((?:^\s+.+?\s+[0-9.]+%\s+\(\d+/\d+\)\s*$\n?)+)", + blob, re.MULTILINE) + if la_block: + out["lookahead"] = [] + for row in LOOKAHEAD_RE.finditer(la_block.group(1)): + out["lookahead"].append(dict( + predictor=row.group(1).strip(), pct=float(row.group(2)), + hit=int(row.group(3)), tot=int(row.group(4)))) + parsed.add("lookahead") + + cuda = dict(enabled=False, expert_count=None, expert_gb=None, + calls_served=None, resident_tensors=None, resident_gb=None, + groups=None, groups_timing=None) + if CUDA_DEVICE_RE.search(stderr): + cuda["enabled"] = True + m = CUDA_TIER_RE.search(stdout) + if m: + cuda["expert_count"] = int(m.group(1)) + cuda["expert_gb"] = float(m.group(2)) + cuda["calls_served"] = int(m.group(3)) + m = CUDA_RESIDENT_RE.search(stderr) + if m: + cuda["resident_tensors"] = int(m.group(1)) + cuda["resident_gb"] = float(m.group(2)) + m = CUDA_GROUPS_RE.search(stderr) + if m: + cuda["groups"] = dict(zip( + ("calls", "experts", "rows", "experts_per_call"), + (int(m.group(1)), int(m.group(2)), int(m.group(3)), float(m.group(4))))) + m = CUDA_GROUPS_TIME_RE.search(stderr) + if m: + cuda["groups_timing"] = dict(zip( + ("h2d_ms", "kernel_ms", "d2h_ms"), + (float(m.group(1)), float(m.group(2)), float(m.group(3))))) + out["cuda"] = cuda + + m = TF_MATCH_RE.search(stdout) + if m: + out["tf_match"] = (int(m.group(1)), int(m.group(2))); parsed.add("tf_match") + + # Capture per-position argmax divergences from the oracle, keyed by position. + mismatches = {int(mm.group(1)): int(mm.group(2)) + for mm in TF_MISMATCH_RE.finditer(blob)} + if out["tf_match"] is not None: + out["tf_mismatches"] = mismatches + parsed.add("tf_mismatches") + + out["parsed"] = parsed + return out + + +def run_engine( + env_overlay: dict, + *, + engine: Optional[str] = None, + cap: int = 4, + ebits: int = 4, + dbits: int = 4, + timeout: float = 600.0, + snap: Optional[str] = None, +) -> tuple[dict, subprocess.CompletedProcess]: + """Run the engine with an env overlay; return (parsed_telemetry, proc). + + `engine` defaults to ./glm.exe (colibri's Windows host). `snap` defaults to + the bundled tiny model (glm_tiny) so callers can omit it for fast tests. + The positional argv is `cap ebits dbits`, matching the engine's main(). + """ + if engine is None: + engine = str(Path(__file__).resolve().parent.parent / "glm.exe") + env = os.environ.copy() + # Strip CUDA vars by default so a CPU run isn't accidentally GPU-accelerated + # by a leftover env; callers opt in by passing them in env_overlay. + for k in ("COLI_CUDA", "COLI_GPU", "COLI_GPUS", "CUDA_DENSE", "CUDA_EXPERT_GB"): + env.pop(k, None) + env.update(env_overlay) + if snap is not None: + env["SNAP"] = snap + elif "SNAP" not in env: + env["SNAP"] = str(Path(__file__).resolve().parent.parent / "glm_tiny") + + proc = subprocess.run( + [engine, str(cap), str(ebits), str(dbits)], + env=env, capture_output=True, text=True, timeout=timeout, + ) + telemetry = parse_run(proc.stdout, proc.stderr) + telemetry["returncode"] = proc.returncode + telemetry["env"] = {k: env_overlay[k] for k in env_overlay} + return telemetry, proc + + +def disk_wait_share(t: dict) -> Optional[float]: + """Fraction of decode wall-time spent waiting on expert disk reads. + + Preferred source: [PROF] time_shares (the engine's own accounting, which + separates felt-wait from read-service). Falls back to PROFILE disk / sum if + [PROF] wasn't emitted (PROF=0 runs). None if neither is available. + """ + if t.get("time_shares"): + return t["time_shares"]["io"] + if t.get("profile") and t.get("profile_sum"): + return t["profile"]["disk"] / t["profile_sum"] + return None + + +def tf_agreement(cpu: dict, cuda: dict, oracle: list[int]) -> tuple[float, list[int]]: + """Direct CPU-vs-CUDA argmax agreement on the TF fixture. + + Both runs prefilled the SAME oracle sequence; tf_mismatches holds each + backend's actual argmax where it diverged from the oracle. Where a backend + is ABSENT from the mismatch map, its prediction equals the oracle token at + that position. So the reconstructed per-position prediction is: + oracle[i] if i not in mismatches else mismatches[i] + and agreement is the fraction of positions where CPU and CUDA predictions + are identical — independent of how each relates to the oracle. + + `oracle` is ref_glm.json's tf_pred (the per-position oracle argmax). Pass + n_positions = len(oracle). + + Returns (agreement_fraction, list_of_differing_positions). + """ + cm = cpu.get("tf_mismatches") or {} + gm = cuda.get("tf_mismatches") or {} + diff = [] + for i, orc in enumerate(oracle): + cpu_tok = cm.get(i, orc) # matched oracle => oracle token + cuda_tok = gm.get(i, orc) + if cpu_tok != cuda_tok: + diff.append(i) + agree = (len(oracle) - len(diff)) / len(oracle) if oracle else 0.0 + return agree, diff diff --git a/c/tools/eval_glm.py b/c/tools/eval_glm.py index 3aa568e..55e12d6 100644 --- a/c/tools/eval_glm.py +++ b/c/tools/eval_glm.py @@ -22,7 +22,7 @@ USO: # leve di ricerca: passate al motore via env TOPP=0.9 python3 tools/eval_glm.py --snap /home/vincenzo/glm52_i4 --data ./bench --tasks mmlu --ram 15 """ -import os, sys, subprocess, argparse, random, json, tempfile, time +import os, sys, subprocess, argparse, random, json, tempfile, time, threading # mini-set OFFLINE per testare la meccanica (NON misura qualita': domande banali) SMOKE = [ @@ -114,6 +114,7 @@ def main(): ap.add_argument("--seed", type=int, default=1234) ap.add_argument("--dry", action="store_true", help="build requests and stop without running the engine") ap.add_argument("--selftest", action="store_true", help="verify the scoring calculations") + ap.add_argument("--out", default="", help="write incremental results CSV here (one row per request, flushed as it lands)") a = ap.parse_args() if a.selftest: # acc/acc_norm con logprob sintetici @@ -143,15 +144,62 @@ def main(): if a.ram: env["RAM_GB"] = str(a.ram) cmd = [a.glm, str(a.cap)] + a.bits.split() print("running:", " ".join(cmd), file=sys.stderr) + + # Stream results line-by-line so a crash at request N keeps 1..N-1 and shows + # exactly where it stopped. The engine prints " " per + # request to stdout and "[score N req | ...]" progress to stderr; buffering + # both until exit (the old subprocess.run) wastes the whole run on a crash. + out_f = open(a.out, "a") if a.out else None + if out_f: + out_f.write(f"# eval_glm snap={a.snap} tasks={a.tasks} limit={a.limit} seed={a.seed} started={time.strftime('%Y-%m-%dT%H:%M:%S')}\n") + out_f.write("req_idx,task,qi,oi,contlen,contchars,gold,logprob,greedy\n") + out_f.flush() t0 = time.time() - proc = subprocess.run(cmd, env=env, capture_output=True, text=True) - if proc.returncode != 0: - print("ENGINE ERROR:\n", proc.stderr[-2000:], file=sys.stderr); sys.exit(1) - lines = [l for l in proc.stdout.strip().splitlines() if l and l[0] in "-0123456789"] - if len(lines) != len(reqs): - print(f"WARNING: {len(lines)} outputs for {len(reqs)} requests", file=sys.stderr) - lp = [float(l.split()[0]) for l in lines] - print(f"(engine: {time.time()-t0:.0f}s){proc.stderr.strip().splitlines()[-1] if proc.stderr.strip() else ''}", file=sys.stderr) + proc = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.PIPE, + text=True, bufsize=1) # line-buffered + lp = [None] * len(reqs) + n_done = 0 + # Drain stderr (engine progress lines) to console live on a background thread + # so the [score N req] heartbeat is visible while stdout is consumed below. + def _drain_stderr(): + for line in proc.stderr: + print(f" [engine] {line.rstrip()}", file=sys.stderr) + threading.Thread(target=_drain_stderr, daemon=True).start() + for line in proc.stdout: + line = line.strip() + if not line or line[0] not in "-0123456789": continue + parts = line.split() + if n_done >= len(reqs): break + try: logprob = float(parts[0]) + except (ValueError, IndexError): continue + lp[n_done] = logprob + greedy = parts[2] if len(parts) > 2 else "?" + t, qi, oi, clen, cchars, gold = meta[n_done] + if out_f: + out_f.write(f"{n_done},{t},{qi},{oi},{clen},{cchars},{gold},{logprob:.6f},{greedy}\n") + out_f.flush() + n_done += 1 + if n_done % 5 == 0 or n_done == len(reqs): + elapsed = time.time() - t0 + rate = n_done / elapsed if elapsed > 0 else 0 + eta = (len(reqs) - n_done) / rate if rate > 0 else 0 + print(f"[progress] {n_done}/{len(reqs)} requests scored | {elapsed:.0f}s elapsed | " + f"{rate:.2f} req/s | ETA {eta:.0f}s | last: {t} q{qi} opt{oi} lp={logprob:.3f}", + file=sys.stderr) + proc.wait() + elapsed = time.time() - t0 + if out_f: + out_f.write(f"# finished: {n_done}/{len(reqs)} in {elapsed:.0f}s, exit={proc.returncode}\n") + out_f.close() + if proc.returncode != 0 and n_done == 0: + print(f"ENGINE ERROR (exit {proc.returncode})", file=sys.stderr); sys.exit(1) + if n_done != len(reqs): + print(f"WARNING: only {n_done}/{len(reqs)} requests scored (engine exited {proc.returncode}); " + f"scoring partial results.", file=sys.stderr) + # Fill any unscored slots with -inf so argmax never picks them + for i in range(len(lp)): + if lp[i] is None: lp[i] = float("-inf") + print(f"(engine: {elapsed:.0f}s, {n_done}/{len(reqs)} scored, exit {proc.returncode})", file=sys.stderr) score_accuracy(tasks, meta, perq, lp) print("\nNOTE: compare acc_norm with GLM-5.2's PUBLISHED model-card score. A close result" "\n indicates that int4 quantization preserved quality. (Fill REFERENCE in tools/eval_glm.py.)") diff --git a/c/tools/iq3_pack.py b/c/tools/iq3_pack.py new file mode 100644 index 0000000..03d139b --- /dev/null +++ b/c/tools/iq3_pack.py @@ -0,0 +1,170 @@ +#!/usr/bin/env python3 +"""fmt=5 (E8/IQ3 grouped container) index codec — #452 ladder step 2. + +The ablation (#453) proved the SCHEME: an IQ3_XXS-style codebook plus rotation +matches our simulated E8 ball (51.5% vs 51.5% on OLMoE). That code quantizes to +lattice points and keeps floats. This module produces the DEPLOYABLE bytes and +reads them back, so the container, the converter and the decode kernels all +agree on one layout. + +Layout — one 256-weight super-block, 98 bytes, 3.0625 bpw: + + [0 .. 63] uint8 grid index per 4-dim magnitude block (64 blocks) + [64 .. 95] uint32 x8, one per 32-weight sub-block: + bits 0..20 three 7-bit sign words (8 weights each, + bit i set => weight i negative; the 8th sign + is implied by odd parity) + bits 21..27 the fourth 7-bit sign word + bits 28..31 4-bit sub-scale code + [96 .. 97] fp16 super-scale d + + value(w) = d * (0.5 + code) * 0.5 * grid[idx][j] * 0.5 * sign + +The last 0.5 is the half-unit convention of the published grid (magnitudes are +stored doubled: 4,12,...,62 mean 2.0,6.0,...,31.0). + +Odd-parity signs: llama.cpp stores 7 of every 8 signs and derives the 8th so the +product of the eight is +1. The encoder therefore flips the smallest-magnitude +weight of any block whose true signs violate that — the same cost the ablation +priced in, now applied for real. +""" +import json +import os +import numpy as np + +QK = 256 # weights per super-block +SUB = 32 # weights per sub-block (one uint32 of signs+scale) +BLOCK_BYTES = QK // 4 + (QK // SUB) * 4 + 2 # 64 + 32 + 2 = 98 + +_GRID = None + + +def grid(): + """[256,4] float32 magnitudes in weight units (published table is doubled).""" + global _GRID + if _GRID is None: + path = os.path.join(os.path.dirname(__file__), "iq3xxs_grid.json") + _GRID = np.asarray(json.load(open(path)), dtype=np.float32) * 0.5 + return _GRID + + +def _nearest(mag4): + """[N,4] magnitudes -> [N] grid indices, argmin ||m-g||^2 without cdist.""" + g = grid() + g2 = (g * g).sum(1) + out = np.empty(len(mag4), dtype=np.uint8) + for i in range(0, len(mag4), 1 << 16): # bounded working set + c = mag4[i:i + (1 << 16)] + out[i:i + len(c)] = np.argmin(g2 - 2.0 * (c @ g.T), axis=1).astype(np.uint8) + return out + + +def encode(x): + """float32 [..., K] (K % 256 == 0) -> packed uint8 [..., K//256 * 98].""" + x = np.ascontiguousarray(x, dtype=np.float32) + K = x.shape[-1] + if K % QK: + raise ValueError(f"fmt=5 needs K % {QK} == 0, got {K}") + rows = x.reshape(-1, K) + nsb = K // QK + out = np.empty((len(rows), nsb * BLOCK_BYTES), dtype=np.uint8) + + for sb in range(nsb): + blk = rows[:, sb * QK:(sb + 1) * QK] # [R,256] + sign = np.where(blk < 0, -1.0, 1.0).astype(np.float32) + mag = np.abs(blk) + + # parity fix: flip the smallest magnitude of every 8 whose product is -1 + s8 = sign.reshape(len(rows), QK // 8, 8) + m8 = mag.reshape(len(rows), QK // 8, 8) + viol = s8.prod(-1) < 0 # [R,32] + amin = m8.argmin(-1) + r, b = np.nonzero(viol) + s8[r, b, amin[r, b]] *= -1.0 + sign = s8.reshape(len(rows), QK) + + base = sb * BLOCK_BYTES + # super-scale: RMS anchor, same statistic the ablation searches around + d = np.sqrt((mag * mag).mean(-1, keepdims=True)) / 20.0 + 1e-12 + out[:, base + 96:base + 98] = d.astype(np.float16).view(np.uint8) + d = d.astype(np.float16).astype(np.float32) # encode what we store + + g = grid() + for ib in range(QK // SUB): + m = mag[:, ib * SUB:(ib + 1) * SUB] # [R,32] + best_err = None + best = None + for code in range(16): + db = d * (0.5 + code) * 0.5 + q = (m / np.maximum(db, 1e-20)).reshape(-1, 4) + idx = _nearest(q) + rec = g[idx].reshape(len(rows), SUB) * db + err = ((rec - m) ** 2).sum(-1, keepdims=True) + if best_err is None: + best_err, best = err, (idx.reshape(len(rows), SUB // 4), code) + else: + take = (err < best_err)[:, 0] + if take.any(): + keep_idx, keep_code = best + ni = idx.reshape(len(rows), SUB // 4) + keep_idx = np.where(take[:, None], ni, keep_idx) + # per-row code: store alongside, resolved below + keep_code = np.where(take, code, keep_code) if isinstance( + keep_code, np.ndarray) else np.where( + take, code, np.full(len(rows), keep_code)) + best = (keep_idx, keep_code) + best_err = np.where(take[:, None], err, best_err) + bidx, bcode = best + if not isinstance(bcode, np.ndarray): + bcode = np.full(len(rows), bcode) + out[:, base + ib * 8:base + (ib + 1) * 8] = bidx.astype(np.uint8) + + # signs: four 7-bit words for this sub-block + the 4-bit code + s = sign[:, ib * SUB:(ib + 1) * SUB].reshape(len(rows), 4, 8) + neg = (s < 0).astype(np.uint32) + word = np.zeros(len(rows), dtype=np.uint32) + for l in range(4): + seven = np.zeros(len(rows), dtype=np.uint32) + for j in range(7): + seven |= neg[:, l, j] << j + word |= seven << (7 * l) + word |= (bcode.astype(np.uint32) & 0xF) << 28 + off = base + QK // 4 + ib * 4 + out[:, off:off + 4] = word.view(np.uint8).reshape(len(rows), 4) if False else \ + np.ascontiguousarray(word).view(np.uint8).reshape(len(rows), 4) + return out.reshape(*x.shape[:-1], nsb * BLOCK_BYTES) + + +def decode(packed, K): + """packed uint8 [..., K//256*98] -> float32 [..., K]. The kernels' reference.""" + packed = np.ascontiguousarray(packed, dtype=np.uint8) + nsb = K // QK + rows = packed.reshape(-1, nsb * BLOCK_BYTES) + out = np.empty((len(rows), K), dtype=np.float32) + g = grid() + for sb in range(nsb): + base = sb * BLOCK_BYTES + d = rows[:, base + 96:base + 98].copy().view(np.float16).astype(np.float32) + for ib in range(QK // SUB): + idx = rows[:, base + ib * 8:base + (ib + 1) * 8] # [R,8] + off = base + QK // 4 + ib * 4 + word = np.ascontiguousarray(rows[:, off:off + 4]).view(np.uint32).reshape(-1) + code = (word >> 28) & 0xF + db = d[:, 0] * (0.5 + code) * 0.5 # [R] + mag = g[idx].reshape(len(rows), SUB) # [R,32] + sgn = np.ones((len(rows), 4, 8), dtype=np.float32) + for l in range(4): + seven = (word >> (7 * l)) & 0x7F + par = 0 + for j in range(7): + bit = (seven >> j) & 1 + sgn[:, l, j] = np.where(bit == 1, -1.0, 1.0) + par ^= bit + sgn[:, l, 7] = np.where(par == 1, -1.0, 1.0) # odd parity closes the block + out[:, sb * QK + ib * SUB:sb * QK + (ib + 1) * SUB] = \ + mag * sgn.reshape(len(rows), SUB) * db[:, None] + return out.reshape(*packed.shape[:-1], K) + + +def bpw(): + return BLOCK_BYTES * 8 / QK diff --git a/c/tools/iq3xxs_grid.json b/c/tools/iq3xxs_grid.json new file mode 100644 index 0000000..3cde1a2 --- /dev/null +++ b/c/tools/iq3xxs_grid.json @@ -0,0 +1 @@ +[[4, 4, 4, 4], [20, 4, 4, 4], [36, 4, 4, 4], [12, 12, 4, 4], [28, 12, 4, 4], [62, 12, 4, 4], [4, 20, 4, 4], [20, 20, 4, 4], [12, 28, 4, 4], [20, 36, 4, 4], [28, 62, 4, 4], [44, 62, 4, 4], [12, 4, 12, 4], [28, 4, 12, 4], [4, 12, 12, 4], [20, 12, 12, 4], [12, 20, 12, 4], [44, 20, 12, 4], [4, 28, 12, 4], [20, 28, 12, 4], [12, 36, 12, 4], [36, 44, 12, 4], [4, 62, 12, 4], [4, 4, 20, 4], [20, 4, 20, 4], [36, 4, 20, 4], [12, 12, 20, 4], [4, 20, 20, 4], [20, 20, 20, 4], [12, 28, 20, 4], [28, 28, 20, 4], [62, 28, 20, 4], [12, 44, 20, 4], [62, 44, 20, 4], [44, 62, 20, 4], [12, 4, 28, 4], [62, 4, 28, 4], [4, 12, 28, 4], [20, 12, 28, 4], [44, 20, 28, 4], [4, 62, 28, 4], [28, 12, 36, 4], [62, 28, 36, 4], [36, 36, 36, 4], [62, 44, 36, 4], [28, 62, 36, 4], [44, 62, 36, 4], [12, 4, 44, 4], [62, 4, 44, 4], [20, 28, 44, 4], [20, 44, 44, 4], [44, 28, 52, 4], [36, 52, 52, 4], [4, 12, 62, 4], [36, 12, 62, 4], [52, 12, 62, 4], [28, 36, 62, 4], [12, 52, 62, 4], [12, 4, 4, 12], [28, 4, 4, 12], [4, 12, 4, 12], [20, 12, 4, 12], [12, 20, 4, 12], [28, 20, 4, 12], [4, 28, 4, 12], [20, 28, 4, 12], [36, 28, 4, 12], [62, 36, 4, 12], [4, 44, 4, 12], [4, 4, 12, 12], [20, 4, 12, 12], [12, 12, 12, 12], [4, 20, 12, 12], [20, 20, 12, 12], [12, 4, 20, 12], [28, 4, 20, 12], [4, 12, 20, 12], [20, 12, 20, 12], [12, 20, 20, 12], [4, 28, 20, 12], [20, 62, 20, 12], [4, 4, 28, 12], [20, 4, 28, 12], [4, 20, 28, 12], [12, 28, 28, 12], [52, 36, 28, 12], [52, 52, 28, 12], [12, 4, 36, 12], [44, 4, 36, 12], [4, 44, 36, 12], [4, 20, 44, 12], [36, 20, 44, 12], [52, 36, 44, 12], [12, 62, 44, 12], [44, 4, 52, 12], [20, 20, 62, 12], [4, 36, 62, 12], [4, 4, 4, 20], [20, 4, 4, 20], [12, 12, 4, 20], [28, 12, 4, 20], [4, 20, 4, 20], [20, 20, 4, 20], [52, 20, 4, 20], [12, 28, 4, 20], [20, 36, 4, 20], [12, 4, 12, 20], [28, 4, 12, 20], [44, 4, 12, 20], [4, 12, 12, 20], [20, 12, 12, 20], [12, 20, 12, 20], [4, 28, 12, 20], [28, 52, 12, 20], [62, 52, 12, 20], [4, 62, 12, 20], [4, 4, 20, 20], [20, 4, 20, 20], [12, 12, 20, 20], [62, 12, 20, 20], [4, 20, 20, 20], [20, 20, 20, 20], [62, 28, 20, 20], [4, 36, 20, 20], [44, 44, 20, 20], [12, 4, 28, 20], [4, 12, 28, 20], [36, 12, 28, 20], [4, 62, 28, 20], [36, 62, 28, 20], [44, 28, 36, 20], [28, 44, 36, 20], [28, 4, 44, 20], [62, 20, 44, 20], [12, 36, 44, 20], [36, 62, 44, 20], [12, 4, 62, 20], [28, 4, 62, 20], [52, 12, 62, 20], [44, 36, 62, 20], [12, 4, 4, 28], [4, 12, 4, 28], [20, 12, 4, 28], [12, 20, 4, 28], [28, 20, 4, 28], [4, 44, 4, 28], [44, 52, 4, 28], [20, 62, 4, 28], [4, 4, 12, 28], [20, 4, 12, 28], [4, 20, 12, 28], [12, 28, 12, 28], [36, 36, 12, 28], [52, 36, 12, 28], [12, 4, 20, 28], [28, 4, 20, 28], [4, 12, 20, 28], [44, 20, 20, 28], [20, 44, 20, 28], [20, 62, 20, 28], [12, 12, 28, 28], [28, 28, 28, 28], [4, 28, 36, 28], [62, 36, 36, 28], [20, 62, 36, 28], [4, 4, 44, 28], [52, 4, 44, 28], [20, 20, 44, 28], [44, 44, 44, 28], [36, 12, 52, 28], [52, 28, 52, 28], [28, 52, 52, 28], [28, 28, 62, 28], [4, 52, 62, 28], [36, 4, 4, 36], [62, 12, 4, 36], [44, 28, 4, 36], [62, 28, 4, 36], [28, 44, 4, 36], [62, 44, 4, 36], [36, 62, 12, 36], [4, 20, 20, 36], [62, 28, 20, 36], [4, 36, 20, 36], [4, 52, 20, 36], [52, 52, 20, 36], [62, 4, 28, 36], [44, 36, 28, 36], [36, 4, 36, 36], [12, 44, 36, 36], [36, 52, 36, 36], [44, 20, 44, 36], [28, 36, 44, 36], [4, 62, 44, 36], [44, 4, 62, 36], [4, 12, 62, 36], [20, 12, 62, 36], [4, 28, 62, 36], [20, 12, 4, 44], [12, 36, 4, 44], [4, 62, 4, 44], [4, 4, 12, 44], [52, 4, 12, 44], [52, 20, 12, 44], [44, 44, 12, 44], [36, 12, 20, 44], [20, 28, 20, 44], [20, 62, 20, 44], [20, 4, 28, 44], [28, 44, 28, 44], [4, 12, 36, 44], [28, 20, 36, 44], [62, 20, 36, 44], [20, 62, 36, 44], [20, 4, 44, 44], [12, 28, 44, 44], [4, 44, 52, 44], [36, 20, 62, 44], [20, 36, 62, 44], [36, 20, 4, 52], [36, 36, 4, 52], [52, 36, 4, 52], [36, 52, 4, 52], [12, 20, 12, 52], [12, 52, 12, 52], [62, 12, 20, 52], [36, 52, 20, 52], [4, 28, 28, 52], [52, 28, 28, 52], [36, 36, 36, 52], [44, 4, 44, 52], [20, 44, 44, 52], [28, 28, 52, 52], [28, 4, 62, 52], [12, 20, 62, 52], [28, 4, 4, 62], [44, 4, 4, 62], [62, 4, 4, 62], [4, 12, 4, 62], [20, 28, 4, 62], [20, 44, 4, 62], [52, 20, 12, 62], [4, 36, 12, 62], [20, 12, 20, 62], [44, 36, 20, 62], [20, 44, 20, 62], [4, 4, 28, 62], [44, 12, 28, 62], [28, 28, 28, 62], [4, 52, 28, 62], [12, 20, 36, 62], [12, 36, 36, 62], [4, 4, 44, 62], [20, 4, 44, 62], [36, 20, 44, 62], [4, 28, 52, 62]] \ No newline at end of file diff --git a/c/tools/quant_ablation.py b/c/tools/quant_ablation.py index dd08715..a602ae8 100644 --- a/c/tools/quant_ablation.py +++ b/c/tools/quant_ablation.py @@ -117,11 +117,79 @@ def quantize_param(w, bits, group, rot=False, e8=""): def _grid_or_e8(x, bits, group, e8): + if e8 == "-iq3": + return _quant_iq3(x.float()) if e8: return _quant_e8(x.float(), group, bits, ball=(e8 == "-e8")) return _quant_last_dim(x, bits, group) +# -------------------------------------------------------------------------------------- +# IQ3_XXS-style codebook (#452 candidate (a)): llama.cpp's deployed 3.06-bpw scheme. +# 4-dim magnitude blocks quantized to a 256-entry lattice-subset grid (magnitudes on the +# odd ladder 4,12,..,62 in half-units), signs factored out per 8 weights with an odd-parity +# constraint (7 stored + 1 derived: a block whose true signs violate parity gets its +# smallest-magnitude sign flipped — modelled here so the ablation pays the real cost). +# Scales: fp16 super-scale per 256 + 4-bit sub-scale per 32, db = d*(0.5+s)*0.5. +# Grid extracted from ggml-common.h (MIT). +# -------------------------------------------------------------------------------------- +_IQ3_GRID = None +def _iq3_grid(device): + global _IQ3_GRID + if _IQ3_GRID is None or _IQ3_GRID.device != device: + import json, os + path = os.path.join(os.path.dirname(__file__), "iq3xxs_grid.json") + _IQ3_GRID = torch.tensor(json.load(open(path)), dtype=torch.float32, device=device) + return _IQ3_GRID # [256,4], half-unit magnitudes (value/2 = weight units) + +def _quant_iq3(x): + orig = x.shape + K = orig[-1] + assert K % 256 == 0, "iq3 needs multiples of 256 along the input dim" + xb = x.reshape(-1, 256) # super-blocks + grid = _iq3_grid(x.device) * 0.5 # weight units + out = torch.empty_like(xb) + signs = torch.sign(xb); signs[signs == 0] = 1.0 + mags = xb.abs() + for sb in range(8): # 8 sub-blocks of 32 + m = mags[:, sb*32:(sb+1)*32] # [N,32] + s = signs[:, sb*32:(sb+1)*32] + # per-8 sign parity: flip the smallest-|w| sign where the product is negative + s8 = s.reshape(-1, 4, 8) + m8 = m.reshape(-1, 4, 8) + viol = (s8.prod(-1) < 0) # odd number of minus signs + idxmin = m8.argmin(-1) + flip = torch.zeros_like(s8) + flip.scatter_(-1, idxmin[..., None], 1.0) + s8 = torch.where(viol[..., None].expand_as(s8) & (flip > 0), -s8, s8) + s = s8.reshape(-1, 32) + # sub-scale search: db candidates from the 4-bit code, super d from block RMS + d = m.pow(2).mean(-1, keepdim=True).sqrt() / 20.0 + 1e-12 # rough anchor + best = None + for code in range(16): + db = d * (0.5 + code) * 0.5 + q = m / db # [N,32] target magnitudes + q4 = q.reshape(-1, 4) # 4-dim grid blocks + # chunked argmin ||q-g||^2 = argmin(|g|^2 - 2 q.g): a full cdist on a + # 100M-param tensor materializes tens of GB — this stays at ~256 MB. + g2 = grid.pow(2).sum(-1) + idx = torch.empty(q4.shape[0], dtype=torch.long, device=q4.device) + CH = 1 << 18 + for i0 in range(0, q4.shape[0], CH): + cc = q4[i0:i0+CH] + idx[i0:i0+CH] = (g2 - 2.0 * (cc @ grid.T)).argmin(-1) + hit = grid[idx].reshape(-1, 8, 4) + rec = (hit.reshape(-1, 32) * db) + err = (rec - m).pow(2).sum(-1, keepdim=True) + if best is None: + best = (err, rec) + else: + take = err < best[0] + best = (torch.where(take, err, best[0]), torch.where(take, rec, best[1])) + out[:, sb*32:(sb+1)*32] = best[1] * s + return out.reshape(orig) + + def _rot_quant(x, bits, group, e8=""): """W -> Qn(W@Q) @ Q^T along the last (input) dim — see rotation() above.""" q = rotation(x.shape[-1], x.device) @@ -206,7 +274,7 @@ def _quant_e8(x, group, bits, ball): return best_out.reshape(shp) -SCHEME_RE = re.compile(r"^int(2|3|4|8)(?:-g(\d+))?(-e8u?)?(-rot)?(-nohead)?$") +SCHEME_RE = re.compile(r"^int(2|3|4|8)(?:-g(\d+))?(-e8u?|-iq3)?(-rot)?(-nohead)?$") def parse_scheme(name): diff --git a/docker/Dockerfile b/docker/Dockerfile new file mode 100644 index 0000000..fa639f9 --- /dev/null +++ b/docker/Dockerfile @@ -0,0 +1,15 @@ +FROM debian:stable-slim + +# We install the necessary packages to download and compile the code +RUN apt-get update && \ + apt-get install -y locales git build-essential python3 python3-pip && \ + rm -rf /var/lib/apt/lists/* && \ + mkdir /app + +# Let's clone the repository +WORKDIR /app +RUN git clone https://github.com/JustVugg/colibri.git . +WORKDIR /app/c +# It's time to compile +RUN ./setup.sh && ./coli build + diff --git a/docker/README.IT.md b/docker/README.IT.md new file mode 100644 index 0000000..43c3e31 --- /dev/null +++ b/docker/README.IT.md @@ -0,0 +1,454 @@ +_Read the read me in [English](README.md)._ + + +# Guida a Colibrì - Motore di Inferenza Locale + +Una guida semplice per eseguire **Colibrì**, un motore di inferenza locale basato su GLM 5.2, senza conoscenze di programmazione. Se hai già Docker installato, sei a buon punto. + +--- + +## 📋 Sommario + +- [Cosa è Colibrì?](#cosa-è-colibrì) +- [Cosa serve](#cosa-serve) + - [Hardware](#hardware) + - [Software](#software) +- [Come iniziare](#come-iniziare) + - [Passo 1: Scarica il modello](#passo-1-scarica-il-modello) + - [Passo 2: Scarica il Dockerfile di Colibrì](#passo-2-scarica-il-dockerfile-di-colibrì) + - [Passo 3: Compila l'immagine Docker](#passo-3-compila-limmagine-docker) + - [Passo 4: Avvia Colibrì](#passo-4-avvia-colibr%C3%AC) + - [Cosa significa quel comando?](#cosa-significa-quel-comando) + - [Usare Colibrì](#usare-colibrì) +- [Entrare in una console Linux dentro il container](#entrare-in-una-console-linux-dentro-il-container) +- [Risoluzione dei problemi](#risoluzione-dei-problemi) +- [Note tecniche](#note-tecniche) +- [Domande frequenti](#domande-frequenti) +- [Supporto e contributi](#supporto-e-contributi) +- [Test sul mio PC](#test-sul-mio-pc) + +--- + +## Cosa è Colibrì? + +Colibrì è un'applicazione che ti permette di eseguire un modello di intelligenza artificiale (GLM 5.2) direttamente sul tuo computer, senza connettersi a server esterni. È possbile anche farlo girare in Docker, che isola l'applicazione dal resto del sistema. + +> **Nota importante**: Il modello è molto grande. Attendi anche diversi minuti per una risposta a una domanda semplice, specialmente con poca RAM. Alla fine di questo readme vedrai il risultato sul mio PC (senza scheda grafica discreta) e arrivo a 0.01 token al secondo. + +--- + +## Cosa serve + +### Hardware + +| Memoria RAM | Funziona? | Note | +|:---:|:---:|---| +| < 16 GB | ❌ No | Memoria insufficiente | +| 24 GB | ⚠️ Forse | Possibile, da testare | +| 32 GB | ✅ Sì | Il minimo (ma vedi sezione memoria su Windows) | +| 48+ GB | ✅ Sì | Meglio | + +Inoltre: un **disco SSD veloce** è essenziale. Colibrì usa il disco come memoria aggiuntiva. Con una scheda grafica NVidia è ancora meglio. + +### Software + +- **Docker Desktop** (Windows, Mac, Linux) — [scarica qui](https://www.docker.com/products/docker-desktop/) +- **Python** (solo se vuoi scaricare il modello da casa tua) + - Windows: [python.org](https://www.python.org) oppure Microsoft Store + - Linux: `apt-get install python3 python3-pip` + - Mac: [python.org](https://www.python.org) oppure Homebrew + +Non serve nessun ambiente di compilazione. Tutto avviene dentro il container Docker. + +--- + +## Come iniziare + +### Passo 1: Scarica il modello + +Il modello GLM 5.2 è circa **360 GB**. Scegli uno di questi metodi: + +#### Metodo A: Con Python (consigliato) + +1. **Installa la libreria per Hugging Face:** + ```bash + python -m pip install -U huggingface_hub[cli] + ``` + Su Linux, usa `python3` al posto di `python`. + +2. **Scarica il modello** (apri il terminale nella cartella dove lo vuoi salvare): + ```bash + hf_download mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp --local-dir . + ``` + + **Esempio**: se vuoi salvarlo in `C:\LLM\models\glm-5.2` (Windows): + - Apri PowerShell in quella cartella + - Copia e incolla il comando sopra + - Attendi (molto) + +#### Metodo B: Senza Python (solo se necessario) + +Se sei su Windows e non riesci con Python: +- Scarica manualmente da [Hugging Face](https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp) +- Decomprimi in una cartella (es. `C:\LLM\models\glm-5.2`) + +--- + +### Passo 2: Scarica il Dockerfile di Colibrì + +1. Vai a: https://github.com/JustVugg/colibri/blob/main/docker/Dockerfile +2. Clicca il pulsante **Download** (icona ⬇️) in alto a destra +3. Salva il file in una cartella (es. `C:\LLM\Colibrì`) + +--- + +### Passo 3: Compila l'immagine Docker + +Apri il terminale (PowerShell su Windows, Terminal su Mac/Linux) **nella cartella dove hai salvato il Dockerfile** e digita: + +**Windows:** +```bash +docker build -t colibri-i . +``` + +**Linux/Mac:** +```bash +sudo docker build -t colibri-i . +``` + +Attendi che finisca (pochi minuti). Se tutto va bene, vedrai: `Successfully tagged colibri-i:latest` + +> **Se vuoi recepire gli aggiornamenti del repository**: Cancella prima l'immagine vecchia con `docker rmi colibri-i` e ricompila. + +--- + +### Passo 4: Avvia Colibrì + +Apri il terminale e digita il comando sottostante (sostituisci `C:\LLM\models\glm-5.2` con il percorso reale del tuo PC): + +**Windows** (PowerShell): +```bash +$MODEL_PATH="C:\LLM\models\glm-5.2" +docker run --rm -it --name colibri-c ` + -v "$MODEL_PATH`:/app/glm-5.2" ` + -e COLI_MODEL=/app/glm-5.2 ` + colibri-i ./coli chat +``` + +**Mac/Linux** (Terminal/Bash): +```bash +MODEL_PATH="/path/to/glm-5.2" +docker run --rm -it --name colibri-c \ + -v "$MODEL_PATH:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i ./coli chat +``` + +**Esempio per Linux:** +```bash +MODEL_PATH="/home/user/LLM/glm-5.2" +docker run --rm -it --name colibri-c \ + -v "$MODEL_PATH:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i ./coli chat +``` + +--- + +### Cosa significa quel comando? + +| Parte | Spiegazione | +|-------|-------------| +| `docker run` | Avvia un container | +| `--rm` | Cancella il container quando chiudi | +| `-it` | Modalità interattiva (puoi scrivere e leggere) | +| `-v "PERCORSO:/app/glm-5.2"` | Collega il tuo modello dentro il container | +| `-e COLI_MODEL=/app/glm-5.2` | Dice a Colibrì dove trovare il modello | +| `colibri-i` | Nome dell'immagine Docker | +| `./coli chat` | Avvia Colibrì in modalità chat | + +--- + +### Usare Colibrì + +Una volta avviato, vedrai un prompt come questo: + +``` + ────────────────────────────────────────────────────────── + type and press Enter · Ctrl-C stops the answer · :more continues · :reset clears memory · :q exits +``` + +**Comandi utili:** +- `Scrivi una domanda + Invio` → Ricevi la risposta +- `Ctrl + C` → Interrompi la risposta +- `:reset` → Cancella la memoria della conversazione +- `:q` → Esci + +**Esempio di uso:** +``` +› Quanti abitanti ha la Cina? + +La Cina è attualmente il paese più popoloso al mondo. +La popolazione è di circa 1,41 miliardi di abitanti. +``` + +Il modello capisce **italiano, inglese, cinese e altre lingue**, anche se è ottimizzato per inglese e cinese. + +--- + +## Entrare in una console Linux dentro il container + +Se vuoi esplorare il container come fosse una macchina Linux normale: + +```bash +docker run --rm -it --name colibri-c \ + -v "PERCORSO_MODELLO:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i /bin/bash +``` + +Ora sei dentro Linux. Digita `exit` per uscire. + +--- + +## Risoluzione dei problemi + +### ❌ "Docker non trovato" + +**Causa**: Docker non è installato o il terminale non lo riconosce. + +**Soluzione**: +1. Reinstalla [Docker Desktop](https://www.docker.com/products/docker-desktop/) +2. Riavvia il computer +3. Apri un nuovo terminale e riprova + +--- + +### ❌ "Out of memory" (memoria insufficiente) o container che si chiude subito + +**Causa**: Il tuo computer non ha abbastanza RAM, oppure su Windows, WSL usa meno memoria di quella disponibile. + +**Soluzione per Windows (WSL):** + +1. Apri PowerShell e controlla la memoria disponibile a WSL: + ```bash + wsl + cat /proc/meminfo | grep MemTotal + exit + ``` + Dividi il numero per 1.073.741.824 (è 1024³) per averlo in GB. + +2. Se WSL usa meno di quello che hai, crea un file di configurazione: + - Apri un editor di testo (Notepad va bene) + - Copia questo: + ```ini + [wsl2] + memory=24GB + processors=12 + swap=16GB + ``` + - Salva il file con il nome: `.wslconfig` (con il punto) + - Posizionalo in: `C:\Users\TuoNomeUtente\` + +3. Riavvia WSL da PowerShell: + ```bash + wsl --shutdown + wsl + ``` + +4. Controlla di nuovo: + ```bash + # cat /proc/meminfo | grep MemTotal + # exit + ``` + +**Soluzione per Mac/Linux**: Aumenta la RAM disponibile a Docker dalle impostazioni di Docker Desktop, oppure aggiungi più RAM al computer. + +--- + +### ❌ La risposta è molto lenta + +**Cause possibili**: +1. Il disco è lento +2. Hai poca RAM +3. Colibrì sta usando il disco come memoria aggiuntiva (normale) + +**Come controllare la velocità del disco:** + +**Windows** (PowerShell da amministratore): +```bash +winsat disk -drive C +``` +Cambia `C` con la lettera del tuo disco. + +**Linux/Mac** (Terminal): +```bash +sudo hdparm -Tt /dev/sda +``` +Cambia `/dev/sda` con il tuo disco (vedi con `lsblk` per Linux). + +Un **SSD NVMe moderno** arriva a 15 GB/sec. Se il tuo è sotto 2-3 GB/sec, è lento. + +--- + +### ❌ "Permission denied" su Linux + +**Causa**: Docker richiede permessi da amministratore. + +**Soluzione - Opzione 1** (rapida): +```bash +sudo docker build -t colibri-i . +sudo docker run ... (come sopra, con sudo davanti) +``` + +**Soluzione - Opzione 2** (permanente): +```bash +sudo usermod -aG docker $USER +# Riavvia il computer +docker run ... (senza sudo) +``` + +--- + +### ❌ "Image not found" o errore durante il build + +**Causa**: Il Dockerfile è corrotto o non nella cartella giusta. + +**Soluzione**: +1. Verifica che il Dockerfile sia nella cartella dove apri il terminale: + ```bash + ls Dockerfile # Mac/Linux + dir Dockerfile # Windows + ``` +2. Riscarica il Dockerfile dal repository GitHub +3. Elimina l'immagine vecchia: `docker rmi colibri-i` +4. Riprova il build + +--- + +### ❌ "hf_download: command not found" + +**Causa**: La libreria Hugging Face non è installata correttamente. + +**Soluzione**: +```bash +pip install -U huggingface_hub[cli] +# oppure su Linux/Mac: +pip3 install -U huggingface_hub[cli] +``` + +Poi riprova il comando `hf_download`. + +--- + +### ❌ Il modello non si scarica (timeout o errori di rete) + +**Cause**: Connessione lenta o instabile, Hugging Face temporaneamente non disponibile. + +**Soluzione**: +1. Attendi e riprova il comando `hf_download` +2. Se continua, scarica manualmente da [qui](https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp) +3. Decomprimi il file ZIP nella cartella desiderata + +--- + +## Note tecniche + +### Perché il disco è importante? + +Colibrì usa il disco come "RAM aggiuntiva" virtuale (paging). Un disco **veloce** è cruciale per prestazioni decenti. + +- **SSD NVMe** (consigliato): 1-15 GB/sec +- **SSD SATA**: 0.5-1 GB/sec +- **Hard disk meccanico**: 0.05-0.1 GB/sec ❌ (troppo lento) + +Se il tuo disco è lento, le risposte saranno molto lente anche con molta RAM. + +--- + +### Configurazione di default consigliata per WLS su Windows + +Se hai **esattamente 32 GB di RAM** e usi Windows, è molto probabile che WLS di default sia settato per non consumare più di 16 GB di RAM. Bisogna aumentare questo limite [Risoluzione dei problemi](#risoluzione-dei-problemi) . Nel mio caso ho adottato questa configurazione: +```ini +[wsl2] +memory=24GB +processors=12 +swap=16GB +``` + +Ovvero, nel mio caso, ho lasciato 8 GB di RAM e 4 CPU a Windows e dato 24 GB e 12 processori a WSL + Linux. + +--- + +## Domande frequenti + +**D: E se ho meno di 32 GB di RAM?** +R: Probabile che non funzioni bene. Puoi provare se hai 24 GB, ma non è garantito. + +**D: Posso aumentare la velocità di risposta?** +R: Sì, in parte: +- Usa un SSD NVMe veloce +- Aumenta la RAM +- Riduci la complessità delle domande +- Usa `:reset` per cancellare la memoria e alleggerire il carico + +**D: Posso usare Colibrì senza Docker?** +R: Colibrì è nato così, ma questa guida assume Docker. Per compilare da sorgente, vedi il repository GitHub. + +**D: Quanta connessione internet mi serve dopo aver scaricato il modello?** +R: Zero. Colibrì funziona completamente offline. + +--- + +## Supporto e contributi + +Se trovi errori o hai suggerimenti per migliorare questa guida, aprici una issue o una pull request sul repository GitHub di Colibrì. + +Buon divertimento! 🐦 + +--- + +## Test sul mio PC + +Nel primo caso ho fatto una domanda in italiano, nel secondo in giapponese, e nel terzo ho rifatto la domanda in giapponese ma ho richiesto una risposta in italiano. + +``` +PS C:\quack\llm\colibri\docker> docker run --rm -it --name colibri-c -v "C:\quack\llm\models\glm-5.2:/app/glm-5.2" -e COLI_MODEL=/app/glm-5.2 colibri-i ./coli chat + + ▄▀▀▀▄ ▄ colibrì v1.0 + ▄▄▄▄▀▀▀▀▄▀▀ tiny engine, immense model + ▀▀▀▀▀▀▀ GLM-5.2 · 744B MoE · int4 · streaming CPU + ▀▀▀▀ chat · glm-5.2 · ram -GB · topp off + ▀ + ────────────────────────────────────────────────────────── + type and press Enter · Ctrl-C stops the answer · :more continues · :reset clears memory · :q exits + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › Quanti abitanti ha la Cina? │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + La Cina è attualmente il paese più popoloso al mondo (sebbene, secondo alcune stime recenti, sia stata ormai superata dall'India). + + La popolazione totale della Repubblica Popolare Cinese è di circa 1,41 miliardi di abitanti (dati del 2020-2022 circa). + └─ 76 tok · 0.04 tok/s · hit 3% · RSS 15.9 GB · 2012s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › 漫画「ワンピース」の主人公の名前を教えてください。名前だけで、それ以上のコメントはありません。 │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + ルフィ + └─ 2 tok · 0.01 tok/s · hit 1% · RSS 16.7 GB · 260s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › 漫画「ワンピース」の主人公の名前を教えてください。名前だけで、それ以上のコメントはありません。イタリア語で返信 │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + Il nome del protagonista di One Piece è Monkey D. Luffy. + └─ 14 tok · 0.02 tok/s · hit 2% · RSS 17.3 GB · 593s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › + ``` \ No newline at end of file diff --git a/docker/README.md b/docker/README.md new file mode 100644 index 0000000..44d8a19 --- /dev/null +++ b/docker/README.md @@ -0,0 +1,488 @@ +_Leggi il leggimi in [Italiano](README.IT.md)._ + + +# Colibrì Guide - Local Inference Engine + +A simple guide to running **Colibrì**, a local inference engine based on GLM 5.2, without needing programming knowledge. If you already have Docker installed, you are well on your way. + +--- + +## 📋 Table of Contents + +* [What is Colibrì?](https://www.google.com/search?q=#what-is-colibr%C3%AC) +* [Requirements](https://www.google.com/search?q=#requirements) + * [Hardware](https://www.google.com/search?q=#hardware) + * [Software](https://www.google.com/search?q=#software) +* [How to get started](https://www.google.com/search?q=#how-to-get-started) + * [Step 1: Download the model](https://www.google.com/search?q=#step-1-download-the-model) + * [Step 2: Download the Colibrì Dockerfile](https://www.google.com/search?q=#step-2-download-the-colibr%C3%AC-dockerfile) + * [Step 3: Build the Docker image](https://www.google.com/search?q=#step-3-build-the-docker-image) + * [Step 4: Start Colibrì](https://www.google.com/search?q=#step-4-start-colibr%C3%AC) + * [What does that command mean?](https://www.google.com/search?q=#what-does-that-command-mean) + * [Using Colibrì](https://www.google.com/search?q=#using-colibr%C3%AC) +* [Entering a Linux console inside the container](https://www.google.com/search?q=#entering-a-linux-console-inside-the-container) +* [Troubleshooting](https://www.google.com/search?q=#troubleshooting) +* [Technical notes](https://www.google.com/search?q=#technical-notes) +* [Frequently Asked Questions](https://www.google.com/search?q=#frequently-asked-questions) +* [Support and contributions](https://www.google.com/search?q=#support-and-contributions) +* [Tests on my PC](https://www.google.com/search?q=#tests-on-my-pc) + +--- + +## What is Colibrì? + +Colibrì is an application that allows you to run an artificial intelligence model (GLM 5.2) directly on your computer, without connecting to external servers. It is also possible to run it in Docker, which isolates the application from the rest of the system. + +> **Important note**: The model is very large. Expect to wait several minutes for an answer to a simple question, especially with low RAM. At the end of this readme, you will see the result on my PC (without a discrete graphics card), and I reach 0.01 tokens per second. + +--- + +## Requirements + +### Hardware + +| RAM Memory | Works? | Notes | +| --- | --- | --- | +| < 16 GB | ❌ No | Insufficient memory | +| 24 GB | ⚠️ Maybe | Possible, needs testing | +| 32 GB | ✅ Yes | The minimum (but see memory section for Windows) | +| 48+ GB | ✅ Yes | Better | + +Additionally: a **fast SSD** is essential. Colibrì uses the disk as additional memory. Having an NVidia graphics card is even better. + +### Software + +* **Docker Desktop** (Windows, Mac, Linux) — [download here](https://www.google.com/search?q=https://www.docker.com/products/docker-desktop/) +* **Python** (only if you want to download the model yourself) +* Windows: [python.org](https://www.google.com/search?q=https://www.python.org) or Microsoft Store +* Linux: `apt-get install python3 python3-pip` +* Mac: [python.org](https://www.google.com/search?q=https://www.python.org) or Homebrew + +No build environment is needed. Everything happens inside the Docker container. + +--- + +## How to get started + +### Step 1: Download the model + +The GLM 5.2 model is approximately **360 GB**. Choose one of these methods: + +#### Method A: Using Python (recommended) + +1. **Install the library for Hugging Face:** +```bash +python -m pip install -U huggingface_hub[cli] +``` + +On Linux, use `python3` instead of `python`. +2. **Download the model** (open the terminal in the folder where you want to save it): +```bash +hf_download mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp --local-dir . +``` + +**Example**: if you want to save it in `C:\LLM\models\glm-5.2` (Windows): +* Open PowerShell in that folder +* Copy and paste the command above +* Wait (a long time) + +#### Method B: Without Python (only if necessary) + +If you are on Windows and cannot get it to work with Python: + +* Download manually from [Hugging Face](https://www.google.com/search?q=https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp) +* Unzip into a folder (e.g., `C:\LLM\models\glm-5.2`) + +--- + +### Step 2: Download the Colibrì Dockerfile + +1. Go to: [https://github.com/JustVugg/colibri/blob/main/docker/Dockerfile](https://www.google.com/search?q=https://github.com/JustVugg/colibri/blob/main/docker/Dockerfile) +2. Click the **Download** button (⬇️ icon) in the top right +3. Save the file in a folder (e.g., `C:\LLM\Colibrì`) + +--- + +### Step 3: Build the Docker image + +Open the terminal (PowerShell on Windows, Terminal on Mac/Linux) **in the folder where you saved the Dockerfile** and type: + +**Windows:** + +```bash +docker build -t colibri-i . +``` + +**Linux/Mac:** + +```bash +sudo docker build -t colibri-i . +``` + +Wait for it to finish (a few minutes). If everything goes well, you will see: `Successfully tagged colibri-i:latest` + +> **If you want to receive repository updates**: First delete the old image with `docker rmi colibri-i` and rebuild. + +--- + +### Step 4: Start Colibrì + +Open the terminal and type the command below (replace `C:\LLM\models\glm-5.2` with the actual path on your PC): + +**Windows** (PowerShell): + +```bash +$MODEL_PATH="C:\LLM\models\glm-5.2" +docker run --rm -it --name colibri-c ` + -v "$MODEL_PATH`:/app/glm-5.2" ` + -e COLI_MODEL=/app/glm-5.2 ` + colibri-i ./coli chat +``` + +**Mac/Linux** (Terminal/Bash): + +```bash +MODEL_PATH="/path/to/glm-5.2" +docker run --rm -it --name colibri-c \ + -v "$MODEL_PATH:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i ./coli chat +``` + +**Example for Linux:** + +```bash +MODEL_PATH="/home/user/LLM/glm-5.2" +docker run --rm -it --name colibri-c \ + -v "$MODEL_PATH:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i ./coli chat +``` + +--- + +### What does that command mean? + +| Part | Explanation | +| --- | --- | +| `docker run` | Starts a container | +| `--rm` | Deletes the container when you close it | +| `-it` | Interactive mode (you can write and read) | +| `-v "PATH:/app/glm-5.2"` | Mounts your model inside the container | +| `-e COLI_MODEL=/app/glm-5.2` | Tells Colibrì where to find the model | +| `colibri-i` | Name of the Docker image | +| `./coli chat` | Starts Colibrì in chat mode | + +--- + +### Using Colibrì + +Once started, you will see a prompt like this: + +``` + ────────────────────────────────────────────────────────── + type and press Enter · Ctrl-C stops the answer · :more continues · :reset clears memory · :q exits + +``` + +**Useful commands:** + +* `Write a question + Enter` → Receive the answer +* `Ctrl + C` → Stop the answer +* `:reset` → Clear conversation memory +* `:q` → Exit + +**Usage example:** + +``` +› How many inhabitants does China have? + +China is currently the most populous country in the world. +The population is approximately 1.41 billion people. + +``` + +The model understands **Italian, English, Chinese, and other languages**, although it is optimized for English and Chinese. + +--- + +## Entering a Linux console inside the container + +If you want to explore the container as if it were a normal Linux machine: + +```bash +docker run --rm -it --name colibri-c \ + -v "MODEL_PATH:/app/glm-5.2" \ + -e COLI_MODEL=/app/glm-5.2 \ + colibri-i /bin/bash +``` + +Now you are inside Linux. Type `exit` to leave. + +--- + +## Troubleshooting + +### ❌ "Docker not found" + +**Cause**: Docker is not installed or the terminal does not recognize it. + +**Solution**: + +1. Reinstall [Docker Desktop](https://www.google.com/search?q=https://www.docker.com/products/docker-desktop/) +2. Restart your computer +3. Open a new terminal and try again + +--- + +### ❌ "Out of memory" or container closes immediately + +**Cause**: Your computer does not have enough RAM, or on Windows, WSL is using less memory than available. + +**Solution for Windows (WSL):** + +1. Open PowerShell and check the memory available to WSL: +```bash +wsl +cat /proc/meminfo | grep MemTotal +exit +``` + + +Divide the number by 1,073,741,824 (which is 1024³) to get it in GB. +2. If WSL uses less than what you have, create a configuration file: +* Open a text editor (Notepad is fine) +* Copy this: +```ini +[wsl2] +memory=24GB +processors=12 +swap=16GB + +``` + +* Save the file with the name: `.wslconfig` (with the dot) +* Place it in: `C:\Users\YourUsername\` + +3. Restart WSL from PowerShell: +```bash +wsl --shutdown +wsl +``` + +4. Check again: +```bash +# cat /proc/meminfo | grep MemTotal +# exit +``` + +**Solution for Mac/Linux**: Increase the RAM available to Docker from the Docker Desktop settings, or add more RAM to the computer. + +--- + +### ❌ The answer is very slow + +**Possible causes**: + +1. The disk is slow +2. You have low RAM +3. Colibrì is using the disk as additional memory (normal) + +**How to check disk speed:** + +**Windows** (PowerShell as administrator): + +```bash +winsat disk -drive C +``` + +Change `C` with your disk letter. + +**Linux/Mac** (Terminal): + +```bash +sudo hdparm -Tt /dev/sda +``` + +Change `/dev/sda` with your disk (check with `lsblk` on Linux). + +A **modern NVMe SSD** reaches 15 GB/sec. If yours is under 2-3 GB/sec, it is slow. + +--- + +### ❌ "Permission denied" on Linux + +**Cause**: Docker requires administrator permissions. + +**Solution - Option 1** (quick): + +```bash +sudo docker build -t colibri-i . +sudo docker run ... (as above, with sudo in front) +``` + +**Solution - Option 2** (permanent): + +```bash +sudo usermod -aG docker $USER +# Restart the computer +docker run ... (without sudo) +``` +--- + +### ❌ "Image not found" or error during build + +**Cause**: The Dockerfile is corrupted or not in the right folder. + +**Solution**: + +1. Verify that the Dockerfile is in the folder where you open the terminal: +```bash +ls Dockerfile # Mac/Linux +dir Dockerfile # Windows +``` + +2. Redownload the Dockerfile from the GitHub repository +3. Delete the old image: `docker rmi colibri-i` +4. Retry the build + +--- + +### ❌ "hf_download: command not found" + +**Cause**: The Hugging Face library is not installed correctly. + +**Solution**: + +```bash +pip install -U huggingface_hub[cli] +# or on Linux/Mac: +pip3 install -U huggingface_hub[cli] +``` + +Then retry the `hf_download` command. + +--- + +### ❌ The model does not download (timeout or network errors) + +**Causes**: Slow or unstable connection, Hugging Face temporarily unavailable. + +**Solution**: + +1. Wait and retry the `hf_download` command +2. If it continues, download manually from [here](https://www.google.com/search?q=https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp) +3. Unzip the ZIP file into the desired folder + +--- + +## Technical notes + +### Why is the disk important? + +Colibrì uses the disk as "additional virtual RAM" (paging). A **fast** disk is crucial for decent performance. + +* **NVMe SSD** (recommended): 1-15 GB/sec +* **SATA SSD**: 0.5-1 GB/sec +* **Rotational hard disk**: 0.05-0.1 GB/sec ❌ (too slow) + +If your disk is slow, the answers will be very slow even with a lot of RAM. + +--- + +### Recommended default configuration for WLS on Windows + +If you have **exactly 32 GB of RAM** and are using Windows, it is very likely that WLS by default is set to consume no more than 16 GB of RAM. We need to increase this limit [Troubleshooting](https://www.google.com/search?q=#troubleshooting) . In my case I adopted this configuration: + +```ini +[wsl2] +memory=24GB +processors=12 +swap=16GB + +``` + +That is, in my case, I left 8 GB of RAM and 4 CPUs to Windows and gave 24 GB and 12 processors to WSL + Linux. + +--- + +## Frequently Asked Questions + +**Q: What if I have less than 32 GB of RAM?** + +A: It is likely that it will not work well. You can try if you have 24 GB, but it is not guaranteed. + +**Q: Can I increase the response speed?** + +A: Yes, partly: + +* Use a fast NVMe SSD +* Increase RAM +* Reduce the complexity of questions +* Use `:reset` to clear memory and lighten the load + +**Q: Can I use Colibrì without Docker?** + +A: Colibrì was born that way, but this guide assumes Docker. To build from source, see the GitHub repository. + +**Q: How much internet connection do I need after downloading the model?** + +A: Zero. Colibrì works completely offline. + +--- + +## Support and contributions + +If you find errors or have suggestions for improving this guide, open an issue or a pull request on the Colibrì GitHub repository. + +Have fun! 🐦 + +--- + +## Tests on my PC + +In the first case, I asked a question in Italian; in the second, in Japanese; and in the third, I repeated the question in Japanese but requested an answer in Italian. + +``` +PS C:\quack\llm\colibri\docker> docker run --rm -it --name colibri-c -v "C:\quack\llm\models\glm-5.2:/app/glm-5.2" -e COLI_MODEL=/app/glm-5.2 colibri-i ./coli chat + + ▄▀▀▀▄ ▄ colibrì v1.0 + ▄▄▄▄▀▀▀▀▄▀▀ tiny engine, immense model + ▀▀▀▀▀▀▀ GLM-5.2 · 744B MoE · int4 · streaming CPU + ▀▀▀▀ chat · glm-5.2 · ram -GB · topp off + ▀ + ────────────────────────────────────────────────────────── + type and press Enter · Ctrl-C stops the answer · :more continues · :reset clears memory · :q exits + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › Quanti abitanti ha la Cina? │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + La Cina è attualmente il paese più popoloso al mondo (sebbene, secondo alcune stime recenti, sia stata ormai superata dall'India). + + La popolazione totale della Repubblica Popolare Cinese è di circa 1,41 miliardi di abitanti (dati del 2020-2022 circa). + └─ 76 tok · 0.04 tok/s · hit 3% · RSS 15.9 GB · 2012s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › 漫画「ワンピース」の主人公の名前を教えてください。名前だけで、それ以上のコメントはありません。 │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + ルフィ + └─ 2 tok · 0.01 tok/s · hit 1% · RSS 16.7 GB · 260s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › 漫画「ワンピース」の主人公の名前を教えてください。名前だけで、それ以上のコメントはありません。イタリア語で返信 │ + ╰────────────────────────────────────────────────────────────────────────────────────────────────╯ + + ◆ colibrì + Il nome del protagonista di One Piece è Monkey D. Luffy. + └─ 14 tok · 0.02 tok/s · hit 2% · RSS 17.3 GB · 593s + + ╭────────────────────────────────────────────────────────────────────────────────────────────────╮ + │ › + +``` + +[source: 1] \ No newline at end of file diff --git a/docs/ENVIRONMENT.md b/docs/ENVIRONMENT.md index 394a58f..ce44fce 100644 --- a/docs/ENVIRONMENT.md +++ b/docs/ENVIRONMENT.md @@ -114,6 +114,15 @@ Per-drive byte counts are reported in a `MIRROR:` stats line. Combine with `DIRE | `COLI_CUDA_SHARED_W4A16_MIN_ROWS` | `32` | Min row count to engage the shared-MLP W4A16 kernel. | | `COLI_METAL_UNTRACKED` | off (Metal only) | `=1` sets `MTLResourceHazardTrackingModeUntracked` on Metal buffers (reduces hazard-tracking overhead). | +> **Windows note.** On Windows, a bare `coli chat` / `coli run` / `coli serve` +> (no `--gpu`/`--vram`/`--auto-tier`) **auto-enables the GPU** when it detects a +> CUDA build (`coli_cuda.dll` next to the engine) and at least one GPU via +> `nvidia-smi`. The expert-tier VRAM budget is then sized automatically from the +> card's free VRAM (same computation as `--auto-tier`). If `nvidia-smi` is not on +> `PATH` the run falls back to CPU with a warning — pass `--vram N` (or add +> `nvidia-smi` to `PATH`) to enable CUDA in that case. `--gpu none` forces +> CPU-only. (Linux/macOS behaviour is unchanged: pass a flag to enable CUDA.) + --- ## Advanced / experimental / debug diff --git a/docs/grammar-draft.md b/docs/grammar-draft.md index 707fe5c..6140cbb 100644 --- a/docs/grammar-draft.md +++ b/docs/grammar-draft.md @@ -83,3 +83,27 @@ this implementation adds: forced spans as a **draft source verified in the targe own forward** (lossless even under a wrong grammar, composes with MTP/n-gram in one union batch), deployed where the win is denominated in expert I/O rather than forward passes. + +## Server usage: `response_format` (OpenAI API) + +The gateway (`openai_server.py`) turns `response_format` into a **per-request** +grammar carried to the engine over the `SUBMIT` protocol (optional 7th header +field, `gbytes`; 6-field headers from older clients remain valid — the field is +additive and back-compatible in both directions): + +```jsonc +{"response_format": {"type": "json_object"}} // generic JSON grammar +{"response_format": {"type": "json_schema", + "json_schema": {"schema": { ... }}}} // compiled by schema_gbnf.h +{"response_format": {"type": "gbnf", "grammar": "root ::= ..."}} // raw GBNF (extension) +``` + +Semantics are identical to `GRAMMAR=`/`SCHEMA=`: a **draft source, never a +sampling constraint**. A schema outside the supported subset, or malformed GBNF, +costs the speedup — never the request, never the output. Drafting engages for +greedy requests (`temperature: 0`); sampled requests run undrafted. Compile +overhead is negligible: ~8 µs/request for a typical schema, ~18 µs at the +32-level nesting cap (measured, M3 Max). Grammar payloads are capped at 1 MiB. +As with MTP (#100), a drafted greedy run may differ from an undrafted one in +near-tie tokens (the verify forward has a different batch shape); each output +is a valid greedy stream of its own forward shapes. diff --git a/docs/quickstart.md b/docs/quickstart.md index 7d6a9bf..60d4877 100644 --- a/docs/quickstart.md +++ b/docs/quickstart.md @@ -42,11 +42,31 @@ engine needs. You have two options. **Option A — download a prebuilt binary (no compiler needed).** -Grab the latest `colibri--windows-x86_64.zip` from the -[Releases page](https://github.com/JustVugg/colibri/releases), unzip it, and -skip to [step 3](#3-get-the-model). Python 3 (from -[python.org](https://www.python.org/downloads/)) is still needed if you want to -convert a model yourself. +Grab `colibri--windows-x86_64.zip` from the +[Releases page](https://github.com/JustVugg/colibri/releases) and unzip it. +Inside you'll find: + +| File | What it is | +|---|---| +| `colibri--windows-x86_64.exe` | **the engine** — the C program that actually runs the model | +| `coli` | the command-line launcher (`chat`, `serve`, `convert`, `doctor`, …) | +| `openai_server.py`, `resource_plan.py`, `doctor.py` | Python support for the API server and placement planner | + +Two setup steps: + +1. **Rename the engine to `glm.exe`** so the launcher can find it (it looks for a + binary named `glm`): + ```powershell + Rename-Item colibri-*-windows-x86_64.exe glm.exe + ``` +2. **Install Python 3** from [python.org](https://www.python.org/downloads/) — the + `coli` launcher and the API gateway are Python scripts (the engine itself is + pure C and needs nothing). + +Then continue to [step 3](#3-get-the-model). Prefer to skip the launcher? You can +run the engine directly — `.\glm.exe` reads the model path from the `SNAP` +environment variable (see [docs/windows.md](windows.md)) — but `coli chat` is the +easy path. **Option B — build from source with MSYS2.** Install [MSYS2](https://www.msys2.org/), open the **UCRT64** shell, and run: diff --git a/flake.lock b/flake.lock index 9b788fa..a89029e 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1784160687, - "narHash": "sha256-iYL/bixrb6FlHFu/gIuBYzq6c6lM5AAXsXNSWXtIgQc=", + "lastModified": 1784280462, + "narHash": "sha256-DtoqIqM7VkR6NxAkcLpMwmi02USwWb3JdmNGLyhthc0=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "4382ed2b7a6839d4280a9b386db49cbc5907414d", + "rev": "293d6abedf0478e681a4dfcfcb35b30fc796a32f", "type": "github" }, "original": { diff --git a/flake.nix b/flake.nix index 584a088..a408386 100644 --- a/flake.nix +++ b/flake.nix @@ -6,37 +6,49 @@ flake-utils.url = "github:numtide/flake-utils"; }; - outputs = { self, nixpkgs, flake-utils }: - flake-utils.lib.eachDefaultSystem (system: - let - pkgs = import nixpkgs { inherit system; }; + outputs = { + self, + nixpkgs, + flake-utils, + }: + flake-utils.lib.eachDefaultSystem ( + system: let + pkgs = import nixpkgs {inherit system;}; # Python with the packages needed by the offline converter tools - pythonEnv = pkgs.python3.withPackages (ps: with ps; [ - torch - safetensors - huggingface-hub - numpy - tokenizers - datasets - ]); + pythonEnv = pkgs.python3.withPackages ( + ps: + with ps; [ + torch + safetensors + huggingface-hub + numpy + tokenizers + datasets + ] + ); colibri = pkgs.stdenv.mkDerivation { pname = "colibri"; version = "1.0"; src = ./.; - # python3 is needed by checkPhase: `make test-c` shells out to - # `python3 tools/run_tests.py` (see c/Makefile, PYTHON ?= python3). - nativeBuildInputs = [ pkgs.makeWrapper pkgs.python3 ]; + nativeBuildInputs = with pkgs; [makeWrapper]; - buildInputs = [ - pkgs.gcc - pkgs.gmp + buildInputs = with pkgs; [ + gcc + gmp ]; + # python3 is needed by checkPhase: `make test-c` shells out to + # `python3 tools/run_tests.py` (see c/Makefile, PYTHON ?= python3). + nativeCheckInputs = with pkgs; [python3]; + # Use x86-64-v3 (AVX2) for a portable binary; override with ARCH=native for local builds - ARCH = "x86-64-v3"; + ARCH = + if pkgs.stdenv.hostPlatform.isx86_64 + then "x86-64-v3" + else "native"; buildPhase = '' runHook preBuild @@ -56,7 +68,8 @@ cp c/glm $out/lib/colibri/glm cp c/coli $out/lib/colibri/coli chmod +x $out/lib/colibri/coli - cp c/openai_server.py c/resource_plan.py c/doctor.py $out/lib/colibri/ + cp c/openai_server.py c/resource_plan.py c/doctor.py c/version.py \ + $out/lib/colibri/ cp -r c/tools/* $out/lib/colibri/tools/ # $out/bin holds the user-facing entry points. @@ -86,12 +99,11 @@ description = "Run GLM-5.2 (744B MoE) on a consumer machine with ~25 GB RAM"; homepage = "https://github.com/JustVugg/colibri"; license = licenses.asl20; - platforms = platforms.linux; - mainProgram = "glm"; + platforms = with platforms; linux ++ darwin; + mainProgram = "coli"; }; }; - in - rec { + in { packages = { default = colibri; inherit colibri; @@ -100,23 +112,25 @@ apps = { default = { type = "app"; - program = "${colibri}/bin/glm"; + program = pkgs.lib.getExe colibri; }; - coli = { + glm = { type = "app"; - program = "${colibri}/bin/coli"; + program = "${colibri}/share/colibri/glm"; }; }; - devShells.default = pkgs.mkShell { - inputsFrom = [ colibri ]; + formatter = pkgs.alejandra; - packages = [ + devShells.default = pkgs.mkShell { + inputsFrom = [colibri]; + + packages = with pkgs; [ pythonEnv - pkgs.gcc - pkgs.gnumake - pkgs.clang-tools # clangd / clang-tidy for IDE support - pkgs.pkg-config + gcc + gnumake + clang-tools # clangd / clang-tidy for IDE support + pkg-config ]; shellHook = '' diff --git a/site/colibri-icon.svg b/site/colibri-icon.svg new file mode 100644 index 0000000..76838b9 --- /dev/null +++ b/site/colibri-icon.svg @@ -0,0 +1,50 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/site/colibri.svg b/site/colibri.svg new file mode 100644 index 0000000..4aed0bb --- /dev/null +++ b/site/colibri.svg @@ -0,0 +1,55 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + colibrì + tiny engine, immense model + GLM-5.2 · 744B MoE · int4 · streaming CPU + diff --git a/site/index.html b/site/index.html new file mode 100644 index 0000000..dd6dac1 --- /dev/null +++ b/site/index.html @@ -0,0 +1,698 @@ + + + + + +colibrì — tiny engine, immense model + + + + + + + + + + + + +
+ +
tiny engine · immense model
+

colibrì

+
a 744-billion-parameter mind, on your desk
+

Run GLM-5.2 — 744B Mixture-of-Experts — on hardware you already own. + Pure C, zero dependencies, experts staged just-in-time across VRAM, RAM and disk. + The galaxy behind this text is real: the model's measured expert atlas.

+
+ ★ Star on GitHub + Quick start → +
+
+
744B
parameters
+
19,456
experts
+
0
dependencies
+
25 GB
min RAM
+
9.2
tok/s measured
+
+
scroll
+
+ +
+ +
+
the vision
+

Frontier models should not be sealed inside datacenters.
+ Anyone curious enough should be able to open one up —
+ run it, watch every expert fire, and make it better.

+
+
⚡

Run it

+

A 744B frontier-class MoE answering on a 25 GB machine — the proven floor — up to + 9+ tok/s on a workstation. No cluster, no cloud bill, no API key.

+
🔬

Study it

+

The engine is a microscope: live routing telemetry, per-expert heat, and the first + published expert atlas of a 700B-class model — measured on machines like yours.

+
🛠

Improve it

+

One C file, readable in an afternoon. Every optimisation this project found came from + someone measuring something — the next one can come from you.

+
+
+ +
+
01 — the algorithm
+

A JIT, but for weights

+
+
+

A compiler JIT never compiles the whole program — it watches what actually runs + and compiles the hot paths, just in time. colibrì applies the same bet to a + 744B parameter space: a token activates only ~40B parameters, and just + ~11 GB of those change from token to token — the routed experts.

+
So parameters are not resident state to be held — they are + data to be staged, exactly when the router proves they are needed.
+

The engine keeps the dense weights resident and treats the 19,456 experts as a + heterogeneous storage hierarchy: measured routing heat decides which experts earn + VRAM, which earn pinned RAM, and which stream from NVMe. The router runs a layer + ahead, so prefetch hides the staging latency — and because routing has + measurable structure (the atlas below), the hierarchy keeps getting the answer right. + Like a JIT, it learns your workload: the more you run, the hotter the right experts get.

+

Everything is validated token-exact against the reference transformers + implementation — speed never buys drift.

+
⚗ Active research: the next algorithmic step is under way — + smarter placement and scheduling beyond LRU + pin, overlap of CPU and GPU expert execution, + and routing-aware speculation. Measured, reviewed, and merged in the open.
+
+
+
+ VRAMhottest experts · grouped GPU matmuls · + six 5090s hold the entire routed set — disk reads hit zero
+
↑ promote by measured heat ↓ demote
+
+ RAMwarm set · pinned & wired, learned from + your usage · AVX-512/NEON int4 kernels + LRU behind it
+
↑ promote ↓ demote
+
+ NVMecold tail · streamed on demand with io_uring, + prefetched a layer ahead — how 744B fits a 25 GB machine at all
+
token → router picks 8 of 256 experts × 75 layers → + gather across tiers → int4 matmul → next token
+
+
+
+ +
+
02 — the engine, live
+

Watch it think

+

A replay of the engine at work, paced to measured decode speeds from real + community hardware. The grid is all 19,456 experts (colour = storage tier, brightness = + routing heat, routed = white flash); the galaxy below it is the measured atlas — + watch different questions light different regions of the mind.

+
+
+
+
+ coli chat —
+
+
+
+
+ the brain — 75 MoE layers × 256 experts + MTP
+
+ +
+
0.0 tok/s
decode
+
—
ttft (measured)
+
—
resident hit
+
+
+ VRAM tier + RAM tier + disk tier + routed now +
+
+
+
+
+ the atlas, live — measured specialists routed by the current topic
+
+ +
routing: —
+
+
+
+

Simulation replays a fixed transcript at each profile's measured decode rate + (#82, #389, + #387, #273); + “—” means that machine's report did not publish the number — nothing here is invented. + Atlas flashes sample real measured affinity (atlas v1) for the active topic.

+
+ +
+
+
03 — the atlas
+

Measured, not assumed

+

Every point below is a real measured expert from the published + canonical atlas v1: + 721 canonical specialists confirmed across three independent kernel families + (AVX-512 VNNI · NEON · AVX2, engine-pinned, replication-gated) plus 637 gate-sensitive ones. + Position is the measured 10-topic affinity vector — a specialist sits at its topic's anchor, + a generalist drifts to the centre. Colour = top topic, large points = canonical. Drag to spin.

+
+
+ +
+
drag to spin · scroll past to release
+
+
+
+ +
+
04 — the ladder
+

Same engine, any hardware

+

Measured decode on real community machines — the hardware only changes + where the experts live. Every row links to its public report.

+ + + + + + + + + + + + + + + + + + + + +
hardwaredecodewhere the experts live
6× RTX 5090 · full residency + selective NUMA (#82)9.0–9.2 tok/s
all in VRAM+RAM · disk 0
2× Xeon Gold 6430 · 1 TB DDR5 · CPU-only (#389)5.42 tok/s
every expert pinned in RAM
DGX Spark GB10 · 121 GB unified (#161)3.33 tok/s
unified memory + CACHE_ROUTE
MacBook Pro M5 Max · 128 GB · Metal (#387)2.0 tok/s
Metal tier + warm pin
Ryzen AI Max+ 395 · 128 GB (#200)1.83 tok/s
hot set pinned in RAM
Ryzen 9 9950X3D2 · RTX 5090 · Gen5 NVMe (#120)1.23 tok/s
28 GB VRAM tier + 11.5 GB/s disk
single RTX 5070 Ti laptop-class (#273)1.07 tok/s
GPU-resident pipeline
25 GB dev box · cold — the proven floor0.05–0.1 tok/s
streamed from NVMe
+

…and 15+ more community machines, from an M4 Pro Mac Mini to a 430 GB EPYC server. + Full tables, methodology and quality ablations: docs/benchmarks.md

+
+ +
+
05 — the models
+

One engine, more minds coming

+

The tiering algorithm is model-agnostic: any Mixture-of-Experts with routed + experts can be staged the same way. GLM-5.2 is live today; support for more open-weight + families is on the roadmap.

+
+
live

GLM-5.2

Z.ai · 744B MoE
+

The flagship target: int4 container, token-exact vs reference, full atlas published.

+
live

OLMoE

Allen AI · 7B MoE
+

The small research workhorse — quantization A/Bs and quality ablations run here first.

+
planned

Kimi K2

Moonshot AI · 1T MoE
+

The next scale step: a trillion-parameter mind on the same tiered engine.

+
planned

Qwen3 MoE

Alibaba Qwen
+

The most widely-deployed open family — broad hardware coverage meets broad adoption.

+
planned

MiniMax

MiniMax · MoE
+

Long-context specialists — a different routing profile for the atlas to map.

+
+
+

Thank you to the teams whose open weights make this project possible — + Z.ai (GLM), Moonshot AI (Kimi), Alibaba Qwen, MiniMax, + Allen AI (OLMoE), and everyone releasing frontier-class models in the open. + An engine is nothing without a mind to run.

+
+
+ +
+
06 — join in
+

Come build the microscope

+

Every optimisation in this project started with someone measuring something on + their own machine — a disk swap, a NUMA experiment, an atlas replication on a Mac Mini. + You don't need a datacenter to move frontier-model research. That is the whole point.

+
+

▸ run a benchmark

+

Any hardware is a datapoint — the community table grows one issue at a time.

 +

▸ pick an issue

+

Kernels, schedulers, the atlas, the dashboard — measured problems, waiting.

 +

▸ read the engine

+

One C file plus headers, zero dependencies. Readable in an afternoon — really.

 +
+
+ +
+ +
+ 🐦 colibrì — MIT license + GitHub + API + Tuning + the hummingbird: tiny, fast, precise. +
+ + + +