This website requires JavaScript.
59d74ae435
Merge pull request #340 from tonuonu/fix/fs-detect-9p-statfs
2026-07-17 20:42:22 +02:00
1af9435760
Merge pull request #328 from mohamedmastouri2000-boop/fix/olmoe-ebits
2026-07-17 20:42:18 +02:00
1223f9ca2e
cuda: batch ragged attention across independent streams
2026-07-18 01:54:31 +08:00
1552db2323
Merge pull request #347 from ZacharyZcR/tools/e8-rate-scaled
2026-07-17 19:17:33 +02:00
211d4488c3
win32: stop erasing the real ReadFile error — pread failures name their WinErr (#307 )
2026-07-17 18:49:28 +02:00
5d16368817
sampling: partial top-keep select in attention_rows DSA — O(nk) quickselect, not O(nk log nk) qsort (#356 )
2026-07-17 09:54:40 -04:00
37c96ee07a
bench: add bench_topp -- head-to-head old-qsort vs new-heap timing on V=151936 (#335 )
2026-07-17 09:22:26 -04:00
ba00889fb9
sampling: partial top-p select in dist_build — O(V) heapify + k pops, not O(V log V) qsort (#335 )
2026-07-17 09:01:23 -04:00
099e0c13c4
docs(env): refresh ENVIRONMENT.md — document all 34 drifted env vars
2026-07-17 08:28:41 -04:00
5e2be61a2c
fix(test): make test_stops build on Windows (mkdtemp compat shim)
2026-07-17 08:27:51 -04:00
411f237f94
fix(warnings): silence two -Wall/-Wextra warnings on the MinGW build
2026-07-17 08:26:18 -04:00
03e643f006
fix(flake): create flake.lock, use --set-default for COLI_ENGINE
2026-07-17 14:02:30 +02:00
d5327e2252
omp: cap LLVM libomp idle spin — a parked engine must not burn 3000% CPU (#341 )
2026-07-17 13:53:20 +02:00
617ae51631
Merge pull request #348 from woolcoxm/fix/nix-flake-packaging-345
2026-07-17 13:52:06 +02:00
7eb239328d
coli: serve pidfile + coli stop — one command to shut engine and server down
2026-07-17 13:26:48 +02:00
8bf4cb9a98
coli: chat attaches to a running serve — the engine survives the chat (local)
2026-07-17 02:07:09 +02:00
528b3cff37
fix(flake): package engine, support modules, and python for Nix builds (#345 )
2026-07-17 07:18:24 -04:00
3f239dbb92
tools: rate-scale the E8 lattice ball by bit-width — int3-e8 is now a real int3 codebook, not a fixed 2-bit ball
2026-07-17 18:39:53 +08:00
721bfcb325
docs: add Traditional Chinese README
2026-07-17 17:03:49 +08:00
dc196633f1
pipe: blocking pipe_wait (COLI_PIPE_BLOCK=1) + PIPE_WORKERS implies PIPE=1
2026-07-16 23:20:16 -06:00
81a56777df
glm.c: detect 9p by statfs f_type, not the "/mnt/" path prefix
2026-07-17 08:00:26 +03:00
288edd7190
fix: GPU backend failure-path hardening + tests
2026-07-17 12:39:59 +08:00
5ad4d540ab
test_tok_o200k: fgets instead of getline for the windows job
2026-07-16 16:33:22 -04:00
ed8dab4da4
test_st_pread: Windows-compatible — relative tmpdir, fork subtest gated POSIX
2026-07-16 16:32:43 -04:00
6507817d9f
st: chunked pread with EINTR retry and honest short-read errors
2026-07-16 15:20:43 -04:00
9e2d41567c
tests: o200k tokenizer coverage, no model download
2026-07-16 15:17:21 -04:00
364b9741e2
tok: o200k pre-tokenizer support, auto-detected from tokenizer.json
2026-07-15 23:10:10 -04:00
d45cfa268a
Merge branch 't321' into trial321b
2026-07-16 21:11:23 +02:00
69284facb8
docs: Windows 11 native install walkthrough - toolchain, SAC, CUDA DLL, failure index (#198 )
2026-07-16 21:59:20 +03:00
80f886cd22
convert_olmoe.py: wire --ebits through to quantization (#323 )
2026-07-16 21:59:16 +03:00
cead3d8bf3
Merge pull request #324 from dawnfield-institute/feat/olmoe-ppl-mode
2026-07-16 20:47:37 +02:00
57030ab231
Merge pull request #327 from lEWFkRAD/ci/windows-cuda-build
2026-07-16 20:47:33 +02:00
b75422d605
ci: build the Windows CUDA path (nvcc + MSVC host) — the one toolchain no job covers
2026-07-16 14:35:30 -04:00
e7188df16a
tests: test_stops must not assume /tmp exists (fixes the windows job)
2026-07-16 20:07:44 +02:00
ee5d273cd4
Merge pull request #232 from nbeerbower/profiling-upstream
2026-07-16 19:58:31 +02:00
98f7c88ca9
numa: parenthesise the page-align arithmetic (#313 )
2026-07-16 19:50:20 +02:00
3cec90cd45
Merge pull request #313 from ZacharyZcR/feat/numa-expert-interleave
2026-07-16 19:49:17 +02:00
2dca36a8a0
Merge pull request #314 from mohamedmastouri2000-boop/fix/windows-cuda-dll-build
2026-07-16 19:49:14 +02:00
644e0d16c6
Merge pull request #302 from alekseysorokin68/main
2026-07-16 19:49:11 +02:00
59988c41c8
Merge pull request #143 from lEWFkRAD/ci/make-check-matrix
2026-07-16 19:46:27 +02:00
545d9d0ab5
Merge pull request #317 from ZacharyZcR/docs/serve-protocol
2026-07-16 19:46:24 +02:00
b04c2ca50e
Merge pull request #322 from woolcoxm/fix/oracle-transformers-pin
2026-07-16 19:46:20 +02:00
c686f6bd51
Merge pull request #320 from Magnet-js/docs/winget-python-readme
2026-07-16 19:46:17 +02:00
95b059b15b
Merge pull request #318 from tt1203/fix/oracle-mkdir-glm-tiny
2026-07-16 19:46:13 +02:00
72d3d37231
ci: run the checks on main too (#144 )
2026-07-16 19:39:40 +02:00
78c77bf6c8
ci: make the CUDA job able to fail, and able to run (#144 )
2026-07-16 19:36:31 +02:00
5272d19abc
Merge pull request #144 from ZacharyZcR/ci/github-actions
2026-07-16 19:32:18 +02:00
7bc9f6e2e9
docs/media: refreshed dashboard shots (full-residency 4 tok/s, disk 0), new Atlas galaxy, phone + sidebar shots
2026-07-17 00:29:50 +08:00
c5f2027fa1
olmoe: PPL=1 teacher-forced perplexity mode (loss meter for throughput experiments)
2026-07-16 13:10:50 -03:00
c70d94368e
Merge branch 'main' into fix/windows-cuda-dll-build
2026-07-16 19:05:43 +03:00
fa821a15a2
oracle: hard-fail on transformers < 5.11.0 (interleaved-RoPE floor, #281 )
2026-07-16 12:03:15 -04:00
6de32c55f6
Don't trust a converted checkpoint's config for the stop set
2026-07-16 17:59:34 +02:00
ac7103fe9c
tests: extend the fmt=4 oracle to the fused gate+up kernel (#298 )
2026-07-16 17:27:06 +02:00
64291238da
README: restructure around the technical path — parameters and full results move to docs/, four new figures
2026-07-16 23:24:37 +08:00
05f9bd4b77
docs: add winget python one-liner to Windows install block (#310 )
2026-07-16 16:43:32 +02:00
2a5961a01b
tests: exactness oracle for the grouped-int4 kernel (fmt=4)
2026-07-16 16:37:39 +02:00
c769e04d13
fix(prefetch): address fourth round of copilot review comments
2026-07-16 15:58:38 +02:00
a3e942516c
fix(oracle): create glm_tiny/ before saving so a fresh checkout works
2026-07-16 14:52:16 +01:00
b6bae91b66
fix(prefetch): address third round of copilot review comments
2026-07-16 15:46:26 +02:00
2d8d2951ee
fix(prefetch): address second round of copilot review comments
2026-07-16 15:37:20 +02:00
deafa72b95
docs: serve protocol reference — mux wire format, telemetry lines, HTTP surface (#310 )
2026-07-16 21:25:45 +08:00
36394bc317
Profiling: fix the phase accounting so 'Other' stops swallowing the disk stall
2026-07-15 01:17:52 +00:00
6afffbcbf2
Profiling page: per-turn phase timings, live in the web dashboard
2026-07-14 23:48:54 +00:00
63a6824881
PROF=1: per-request reports in mux serve mode too (the web-dashboard path)
2026-07-14 23:18:03 +00:00
09bf17f001
PROF=1: opt-in performance profile (latency percentiles, expert I/O, tuning verdict)
2026-07-14 23:13:11 +00:00
1ac2e7b487
fix(prefetch): address copilot code quality reviews on safety and concurrency
2026-07-16 14:36:53 +02:00
2b182ae7f6
glm: COLI_NUMA=1 interleaves resident weights across NUMA nodes — +13% decode on 2S, raw mbind, no libnuma (#82 )
2026-07-16 20:32:03 +08:00
9cf97a4f0c
Fix Windows cuda-dll build: quote NVCC, gate MSVC warnings, add config stamp (#306 )
2026-07-16 15:27:08 +03:00
6ade4093de
refactor(prefetch): clean up unused variables, dead functions, and hardcoded paths
2026-07-16 14:22:04 +02:00
6728029555
fix: set stdout to O_BINARY on Windows to prevent READY sentinel corruption
2026-07-16 14:10:27 +03:00
54cfe56324
Merge remote-tracking branch 'origin/main' into dev
2026-07-16 12:38:42 +02:00
16b253ae43
Merge pull request #308 from fermionoid/fix-metal-clang16-std
2026-07-16 12:38:12 +02:00
0d297ee85b
Merge pull request #301 from NeuralNotwerk/pr/pin-auto
2026-07-16 12:38:09 +02:00
613a13020f
Merge pull request #309 from bokiko/fix/bench-fetch-robust
2026-07-16 12:38:05 +02:00
35f90b9e76
Quarantine EXPERT_BUDGET: the operating window is measured empty (#303 )
2026-07-16 12:26:17 +02:00
d57955e95a
Say why the engine died instead of dying mute (#305 )
2026-07-16 12:25:53 +02:00
f86fc6860d
bench: survive transient hub errors — retry with backoff, per-task isolation, atomic writes (#304 )
2026-07-16 13:20:35 +03:00
85c7857fe2
Makefile: add -std=gnu++17 to METALXX for Apple clang 16
2026-07-16 18:11:45 +08:00
81c0398877
PIN=auto — seed the pin from the live usage history
2026-07-16 06:57:36 +00:00
d4b4f33f22
Merge main into dev: Windows fixes (#275-279, #290 ) alongside engine work (#274 , #293 , #294 , #297 )
2026-07-16 07:59:53 +02:00
3074c7d9d0
Merge pull request #279 from KingIcyCreamProjects/pr/win32-launcher-defaults
2026-07-16 07:56:43 +02:00
12d3bd5140
Merge pull request #290 from NeuralNotwerk/fix/mlock-under-mmap
2026-07-16 07:51:48 +02:00
e8479da97d
Merge pull request #278 from KingIcyCreamProjects/pr/win32-hwinfo
2026-07-16 07:51:44 +02:00
2f9ca3143a
Merge pull request #277 from KingIcyCreamProjects/pr/readme-mechanical-fixes
2026-07-16 07:51:41 +02:00
965e8a80aa
Merge pull request #276 from KingIcyCreamProjects/pr/bench-parser-fix
2026-07-16 07:51:37 +02:00
1a8371631b
Merge pull request #275 from KingIcyCreamProjects/pr/win32-physical-cores
2026-07-16 07:51:34 +02:00
a6a99925a2
Merge pull request #294 from bokiko/fix/spec-kernel-pin
2026-07-16 07:37:32 +02:00
c7531c43d1
Merge pull request #297 from RonitBStudent/fix/portable-target-architecture
2026-07-16 07:35:51 +02:00
6d190d91e1
Merge pull request #293 from woolcoxm/fix/pipe2-profiling-and-cuda-mtp-optin
2026-07-16 07:35:48 +02:00
5f167e5964
Windows: pipe2 resident pipeline at decode + EXPERT_BUDGET ws_b cache fix (#274 )
2026-07-16 07:29:34 +02:00
53324fdcf0
fix(build): handle portable detection edge cases
2026-07-15 22:19:25 -05:00
c82593ad87
fix(build): make portable checks target-aware
2026-07-15 21:13:05 -05:00
37da111bb4
glm: pin draft+verify to one kernel family during speculation; soften MTP guard (#163 )
2026-07-16 01:36:15 +03:00
78c675eb8b
cuda: fix pipe2 profile double-count + add COLI_CUDA_MTP opt-in (#292 )
2026-07-15 18:33:25 -04:00
97fa52698f
glm: make MLOCK=1 actually wire pinned experts under COLI_MMAP
2026-07-15 22:12:31 +00:00
550ddcba83
glm(win): ignore cmd.exe's built-in PROMPT template; add COLI_PROMPT (#271 )
2026-07-15 23:11:35 +02:00
6b4f5e5fd0
glm: bilingual (EN+IT) oracle-mismatch message, lead with English (#271 )
2026-07-15 23:02:35 +02:00
a2391587c7
Makefile(win): link -lpsapi so the MinGW build resolves GetProcessMemoryInfo
2026-07-15 22:52:17 +02:00
e56d483a15
run_text: TOKENS=1 env to dump generated token ids for path A/B comparison
2026-07-15 16:31:18 -04:00
a37dda9eae
Merge pull request #289 from woolcoxm/fix/grouped-quant-detect-arbitrary-gs
2026-07-15 22:01:20 +02:00
Vincenzo