This website requires JavaScript.
4aca05900f
Merge #474 : LFRU-aware pilot eviction guard — a speculation never evicts a warm demand expert
2026-07-21 02:00:39 +02:00
eaa68f4005
pilot: LFRU-aware eviction guard — never evict a warm demand expert for a speculation (#441 )
2026-07-20 23:35:55 +02:00
1b8f3076b6
Merge #470 : respect explicit COLI_CUDA_MTP=1 — planner skips DRAFT=0 export
2026-07-21 01:18:03 +02:00
ac04b37af9
fix(plan): respect explicit COLI_CUDA_MTP=1 - skip the DRAFT=0 export so the engine's auto path can engage
2026-07-21 02:12:49 +03:00
e16ec68168
Merge #469 : COLI_MODEL_DIRS — split model shards across N drives (capacity aggregation)
2026-07-21 00:57:16 +02:00
4cc9885cb3
Merge #468 : honour explicit COLI_CUDA=0 over Windows auto-enable + drive-dependent DIRECT docs
2026-07-21 00:57:13 +02:00
f939191404
fix(coli): honour explicit COLI_CUDA=0 over the Windows bare-run auto-enable
2026-07-21 01:44:48 +03:00
0d2fb6f8a2
feat(st): COLI_MODEL_DIRS - split model shards across N drives, no duplication
2026-07-21 01:42:04 +03:00
153e6f4c89
site: fix license — Apache 2.0, not MIT
2026-07-20 23:40:20 +02:00
cbbe31094a
site: fix license — Apache 2.0, not MIT
2026-07-20 23:39:05 +02:00
68ac9ff696
Merge #421 : dual-SSD streaming (COLI_MODEL_MIRROR) — read the model from two drives at once
2026-07-20 22:08:16 +02:00
507c8a1808
fix(ci): remove accidentally-committed test_st_pread binary
2026-07-20 21:58:15 +02:00
87cd0aafed
Merge dev into dual-ssd-mirror: resolve #298/#192-era conflicts
2026-07-20 21:24:39 +02:00
880cfb4ec5
Merge #192 : serve stage 2 — response_format, per-request grammars, grammar-forced drafts
2026-07-20 21:09:04 +02:00
133d271bbf
Merge #458 : fmt=5 index codec tools — deployable bytes for the E8/IQ3 container
2026-07-20 21:07:39 +02:00
c98e5f8bf8
Merge #298 : per-group scales in CUDA dense/attention kernels (fmt=4 correctness)
2026-07-20 20:40:25 +02:00
7291fd19ac
tests: skip the fmt=5 codec tests where numpy is absent
2026-07-21 01:34:26 +08:00
8b69c89f26
fix(rebase): adapt to dev's split tensor_upload/_g API + fault-injection hook
2026-07-20 13:15:19 -04:00
f4409fd7ab
fix(rebase): adapt fmt=4 to dev's quant.h refactor + colibri.c rename
2026-07-20 12:39:46 -04:00
d96e4254a4
fix: apply per-group scales in CUDA attention kernels for fmt=4 (kv_b crash)
2026-07-16 20:07:54 -04:00
bfb80000be
fix: layer_cuda_shard_kvb scale offset for fmt=4 (multi-GPU, per JustVugg review)
2026-07-16 11:33:23 -04:00
6e29260635
cuda+engine: full fmt=4 (grouped int4 gs=64) support + diagnostic harness
2026-07-16 01:09:47 -04:00
86987154f5
eval_glm: stream results line-by-line + incremental CSV (no more lost runs)
2026-07-17 18:29:50 -04:00
00ed8d5aa4
docs: surface the website (badge + header link) in README
2026-07-20 18:49:16 +02:00
1a260a4605
docs: surface the website (badge + header link) in README
2026-07-20 18:46:42 +02:00
f42adfae76
Merge #168 : int3-g64 (fmt=5) per-group-scale 3-bit expert quant
2026-07-20 18:45:08 +02:00
932f3678b0
convert: --up-bits/--gate-bits/--down-bits per-projection expert quant
2026-07-20 17:31:05 +01:00
5e42e70ec3
int3-g64 (fmt=5): per-group-scale 3-bit weight format — engine, converter, tests
2026-07-20 17:31:05 +01:00
cccf8ec5cb
Merge pull request #463 from attilaolah/nix
2026-07-20 18:29:36 +02:00
84514a5f39
review: docs section, 1 MiB grammar pre-check, negative tests, measured compile overhead
2026-07-14 15:44:17 +01:00
d7b855f43f
serve stage 2: per-request grammars end-to-end — response_format -> SUBMIT -> grammar-forced drafts in the multiplexed server
2026-07-14 15:26:32 +01:00
d7211632b4
refactor: grammar-draft state into GrDraft struct (mechanical, no behavior change) — groundwork for per-request grammars in serve_mux
2026-07-14 15:14:03 +01:00
49c11539b2
nix: re-use existing pkgs binding
2026-07-20 18:22:58 +02:00
ebd85a781b
Merge pull request #426 from rgbkrk/quod/metal-residency-set-dev
2026-07-20 18:22:25 +02:00
31526ae619
fix(build): test_pipe_block references glm.c, renamed to colibri.c by #391
2026-07-20 18:15:39 +02:00
e48513c1c6
Merge pull request #338 from noobdev-ph/feat/gpu-backend-hardening
2026-07-20 18:12:00 +02:00
93cde63cd3
Merge pull request #409 from attilaolah/nix
2026-07-20 18:11:43 +02:00
c98b451738
Merge pull request #446 from cdhdt/perf/io-direct-pilot
2026-07-20 18:11:26 +02:00
04ed0049a4
Merge pull request #444 from cdhdt/perf/idot-avxvnni-accumulators
2026-07-20 18:11:09 +02:00
73badea697
Merge pull request #429 from ZacharyZcR/fix/numa-vma-419
2026-07-20 18:08:30 +02:00
aa825d361a
Merge pull request #447 from monotophic/metal/rtop8-parallel
2026-07-20 18:07:44 +02:00
06961728d8
Merge pull request #342 from ZacharyZcR/feat/expert-group-overlap
2026-07-20 18:07:02 +02:00
e78fcfcbca
Merge pull request #451 from ZacharyZcR/feat/cuda-grouped-g64
2026-07-20 18:06:45 +02:00
78a773e539
Merge pull request #432 from ZacharyZcR/feat/cuda-device-router
2026-07-20 18:06:30 +02:00
1aacfcff01
Merge pull request #433 from ZacharyZcR/fix/decode-grouped-kernels
2026-07-20 18:06:13 +02:00
e809171e62
Merge pull request #270 from ebootheee/pr-pipe-batch-sync
2026-07-20 17:52:42 +02:00
d34461c8d8
Merge remote-tracking branch 'origin/dev' into pr270
2026-07-20 17:51:50 +02:00
a9eb839868
Merge pull request #462 from JustVugg/site-to-main
2026-07-20 17:44:11 +02:00
c6801451bd
site: bring the official website + Pages deploy workflow to main (#424 )
2026-07-20 17:43:46 +02:00
d3362303f5
Merge pull request #424 from ZacharyZcR/feat/website
2026-07-20 17:42:08 +02:00
ff38be0fee
Merge pull request #330 from nbeerbower/tok-o200k
2026-07-20 17:41:40 +02:00
48d4af691d
Merge remote-tracking branch 'origin/dev' into pr330
2026-07-20 17:41:10 +02:00
f8e0612f26
Merge pull request #363 from woolcoxm/fix/win32-coli-chat-cuda-autoenable
2026-07-20 17:39:09 +02:00
ff93ae2bc1
Merge pull request #461 from JustVugg/docs-to-main
2026-07-20 17:36:39 +02:00
ec087f6bca
docs: bring the Quick Start guide (incl. Windows prebuilt steps) to main
2026-07-20 17:36:06 +02:00
b347092e5c
Merge remote-tracking branch 'origin/main' into dev
2026-07-20 17:29:44 +02:00
85e7dffb85
Merge pull request #460 from JustVugg/docs-windows-exe
2026-07-20 17:24:15 +02:00
7de49fa02d
docs: explain the prebuilt Windows binary — what the .exe is and how to run it (#450 )
2026-07-20 17:23:21 +02:00
505a0f69aa
Merge pull request #440 from okuvshynov/test_macpro_2019
2026-07-20 17:20:31 +02:00
86766bf989
Merge pull request #427 from DavidePapero/main
2026-07-20 17:20:26 +02:00
0337697023
Merge pull request #332 from woolcoxm/fix/auto-tier-single-core
2026-07-20 17:19:50 +02:00
af23f314fd
Merge pull request #360 from woolcoxm/test/efficiency-suite
2026-07-20 17:09:21 +02:00
1765ed80ac
Merge pull request #453 from ZacharyZcR/feat/ablation-iq3-scheme
2026-07-20 17:09:06 +02:00
bb2bae8904
Merge pull request #456 from monotophic/convert/resume-guard
2026-07-20 17:08:50 +02:00
3ddbc655ef
Merge pull request #437 from 4ny3l/fix-tool-calling-401
2026-07-20 17:07:34 +02:00
3455eeea0a
Merge remote-tracking branch 'origin/dev' into pr437
2026-07-20 17:04:18 +02:00
0b05f6a76b
win32: auto-enable the GPU in bare coli chat
2026-07-17 12:03:43 -04:00
171aa69fcb
fix(resource_plan): take last two lscpu fields, not [1]/[2] (#332 review)
2026-07-17 15:23:51 -04:00
bd3efb1c97
resource_plan: stop setting OMP_PROC_BIND/OMP_PLACES from --auto-tier (#325 )
2026-07-16 16:57:38 -04:00
f27a89ea82
resource_plan: fix --auto-tier pinning decode to one core (#325 )
2026-07-16 16:23:29 -04:00
cf126cbcf9
fix(tests): skip efficiency tests when glm_tiny fixture is absent (CI #360 )
2026-07-17 11:05:25 -04:00
d43b54534f
tests: efficiency suite — tiny-model regression gates + full-model optimization dossier
2026-07-17 10:52:41 -04:00
e48346162a
tools: fmt=5 index codec — deployable bytes for the E8/IQ3 container (#452 step 2)
2026-07-20 22:22:46 +08:00
42a5417c13
convert: mirror the --indir resume params guard onto --repo download loops
2026-07-20 00:19:41 -04:00
ae4e31a15a
tools: IQ3_XXS-codebook scheme in the quant ablation (#452 step 1)
2026-07-20 12:55:44 +08:00
57e67d6c2f
numa: skip binding the GPU-prefix staging slabs
2026-07-20 01:21:21 +08:00
536d8bfd1a
numa: per-layer pin arenas — fix the PIN_GB=all VMA explosion (#419 )
2026-07-20 00:47:57 +08:00
a74e3e0c3a
cuda: grouped-int4 (fmt=4) support in the expert-group kernels (#334 )
2026-07-20 08:56:49 +08:00
e9b36141a4
Merge pull request #449 from cdhdt/fix/rss-guard-uaf
2026-07-20 02:29:06 +02:00
9c5ab39b62
rss_guard: free the evicted expert slab under g_pilot_mx (use-after-free)
2026-07-20 01:27:28 +02:00
4b704823db
Merge pull request #445 from ZacharyZcR/fix/pin-budget-release-host
2026-07-20 00:03:13 +02:00
9f30916003
metal: parallel top-8 expert selection (r_top8_par), default ON
2026-07-19 16:56:14 -04:00
85107f8ce6
io: skip the wasted WILLNEED on expert weights under DIRECT=1
2026-07-19 23:21:10 +02:00
453d1401ba
pin: with CUDA_RELEASE_HOST the VRAM prefix must not consume the RAM pin budget
2026-07-20 05:15:45 +08:00
c4b2cf6db0
idot: independent accumulators on the AVX-VNNI int8/int4 dot kernels
2026-07-19 23:09:51 +02:00
9c698b29a6
bench: microbench for AVX-VNNI idot accumulator A/B (not a gate)
2026-07-19 23:09:51 +02:00
17fbdfa8f8
Merge pull request #393 from ZacharyZcR/feat/auto-tune
2026-07-19 22:40:53 +02:00
a7ef058fc0
Merge pull request #438 from monotophic/e7/disk-class-instr
2026-07-19 22:40:22 +02:00
5c84b25645
fix: old x86 Mac scalar -> vnni
2026-07-19 16:06:49 -04:00
fbabaa4544
Merge pull request #422 from nbeerbower/fix-mtp-probe-expert-count
2026-07-19 22:39:44 +02:00
1f00142b25
PROF: DISK-CLASS — per-load cold/warm disk instrumentation (re-derived onto #391 split)
2026-07-19 15:42:57 -04:00
26bd8b403a
fix(engine): filter non-EOS stop tokens in serve mode to prevent premature halt on tool calls
2026-07-19 13:39:11 -06:00
70fb5b00f3
fix(api): pass tools and tool_choice as parameters to generation() to prevent duplicate extraction
2026-07-19 13:38:57 -06:00
34f6e50091
cuda: decode expert groups take the grouped kernels even under TC_W4A16 (#431 )
2026-07-20 03:35:46 +08:00
113ece3bc7
cuda: COLI_CUDA_ROUTER=1 — route the decode row on the layer's home device (#431 PR-A)
2026-07-20 02:53:43 +08:00
e135119f19
Document current Metal warning baseline
2026-07-19 09:41:50 -07:00
042fd64a7c
Align residency-set docs with split runtime
2026-07-19 09:19:22 -07:00
8efa9ec6c3
E5: hazard-audit round-2 fixes (re-register set hygiene, SUMMARY mutex ref)
2026-07-18 15:57:12 -04:00
e011092ce1
E5: carry the fslab-OOM unwind fix (unregister before free)
2026-07-18 15:35:12 -04:00
82d06000b5
E5: validator round-1 fixes (resset mutex split, flush visibility, citations)
2026-07-18 14:36:00 -04:00
Vincenzo Fornaro