Repository navigation

release: v1.6.0 — the regression fixed, and prefill learns to read each expert once - #943

Merged

JustVugg merged 123 commits into

mainfrom

Aug 12, 2026

Commits on Aug 1, 2026

fix(tune): derive throughput from replay timing
dyKiU
committed
a570e86
fix(tune): retain legacy throughput parsing
dyKiU
committed
8430c38
fix(tune): minimize replay compatibility change
dyKiU
committed
c215880

Commits on Aug 4, 2026

docs(fmt2): M1 Ultra fmt=2 benchmark context — issue/PR summaries, plans, scripts
boblaublaw
committed
1743054

Commits on Aug 5, 2026

wip(fmt2): snapshot fmt=2 benchmark workspace — logs, plans, harness, converter
boblaublaw
committed
1ccdd58
Run the test suite under ASan+UBSan, and fix the clean that hid it
ZacharyZcR
committed
ca13cce
docs(fmt2): M1 Ultra fmt=2 report + benchmarks row + fmt4->fmt2 converter
boblaublaw
committed
f2f745c
Decode MXFP4 (fmt=7) on CUDA, so Kimi K3 has a CUDA expert tier
ZacharyZcR
committed
db0c80f
Move the no-CUDA-flags guard from kimi_k3 to olmoe
ZacharyZcR
committed
53b9959
ci(release): attach to an existing release instead of failing on it

JustVugg
and
claude
committed
256b8fb
fix(metal-test): rename test-local fp8_nblk, which quant.h now also declares

JustVugg
and
claude
committed
c7aaaf7
fix(api): reasoning_effort rendered Max for every level except high

JustVugg
and
claude
committed
3c6788a
Merge pull request #845 from JustVugg/fix/metal-test-fp8-nblk
JustVugg
authored
552c2eb
Merge pull request #846 from JustVugg/fix/reasoning-effort
JustVugg
authored
9bdb0d4
Merge pull request #844 from JustVugg/fix/release-idempotent
JustVugg
authored
c7ae787
fix(doctor): name the shard that failed to read

JustVugg
and
claude
committed
e64b483
Merge pull request #847 from JustVugg/fix/doctor-names-shard
JustVugg
authored
a54b4da

Commits on Aug 6, 2026

Commits on Aug 7, 2026

Commits on Aug 8, 2026

metal: adapt #587's fmt=4 kv_b + moe_block coverage onto our #457 base

monotophic
authored and
RDouglasSharp
committed
19d4004
moe: gs-consistency guard for fmt=4 GPU batches (review F2)

monotophic
authored and
RDouglasSharp
committed
95fdfda
engine: kv_b fmt-gate notice, re-derived for post-#587 two-format gate (r3)

monotophic
authored and
RDouglasSharp
committed
3be4753
probe: env-gated top-2 logit-gap dump at greedy selection (COLI_LOGIT_GAP)

monotophic
authored and
RDouglasSharp
committed
32dc3f9
metal: COLI_METAL_MOE_EXACT — opt-in CPU-exact fmt=4 experts, fast batched default (#587 gate-2)
RDouglasSharp
committed
4427dab
metal: COLI_METAL_MOE_EXACT covers all fmt=4 GPU paths (attn/layer decode + dense)
RDouglasSharp
committed
36a555a
metal: scope COLI_METAL_MOE_EXACT by path, not format

monotophic
authored and
RDouglasSharp
committed
e3bccc4
test: reconcile test_backend_metal with post-rebase dev (qgs=0 at fp8/e8 call sites, rename local fp8_nblk to avoid quant.h clash)
RDouglasSharp
committed
184b3a5
Merge pull request #836 from boblaublaw/fmt_2_m1u
JustVugg
authored
a126067
Merge pull request #886 from shujaanazhar/olmoe-serve-mode
JustVugg
authored
77619b0
Merge pull request #587 from RDouglasSharp/metal-attn-fmt4
JustVugg
authored
98cab3b
ci: build and RUN the Vulkan backend, under Lavapipe

JustVugg
and
claude
committed
0e5510b
web: render markdown in the chat pane, and stop clamping it to 1024 tokens

JustVugg
and
claude
committed
6e90242
deepseek_v4: size the OpenMP team around the expert loaders
Blakeolson21
committed
1684e89
plan: print every placement term, and a switch to A/B the slab shrink (#885)

JustVugg
and
claude
committed
591eace
tune: dispatch per architecture, and give every engine a throughput line (#898)

JustVugg
and
claude
committed
bac9bd7
experiment: add offline residency simulator
terrizoaguimor
committed
cc18bfb
fix: make placement accounting backend neutral
terrizoaguimor
committed
39e5f0c
deepseek_v4: honor CTX in the generate CLI
Blakeolson21
committed
71ef303

Commits on Aug 9, 2026

Commits on Aug 10, 2026

Merge pull request #914 from bherald/fix/mixed-width-pin-budget-885
JustVugg
authored
b347ba4
Merge pull request #902 from terrizoaguimor/pr/cnre-simulator-clean
JustVugg
authored
b2364e5
Merge pull request #915 from mgua/dev
JustVugg
authored
72497b0
Merge pull request #904 from Blakeolson21/v4-cli-ctx
JustVugg
authored
1e2b293
Merge pull request #909 from SulimanAbdulrazzaq/fix/907-multiline-prompts
JustVugg
authored
ed6b475
Merge pull request #903 from terrizoaguimor/pr/placement-accounting-clean
JustVugg
authored
afc771a
Merge pull request #897 from Blakeolson21/v4-omp-loader-reservation
JustVugg
authored
e3f4f7f
quant: AVX2 kernel for int3-g64 matmul (was scalar on x86)

steve-m
authored and
JustVugg
committed
982b827
v4: enable the expert-loader pool and make its depth runtime-tunable

JustVugg
and
claude
committed
48db957
fix(cli): activate readline so UP/DOWN recall previous prompts

JustVugg
and
claude
committed
360da8c
Merge pull request #926 from JustVugg/quant/avx2-matmul-i3
JustVugg
authored
cab0108
Merge pull request #927 from JustVugg/v4/loader-lanes
JustVugg
authored
114c5ad
Merge pull request #928 from JustVugg/fix/chat-history
JustVugg
authored
5fd53a9
fix CNRE cost and capacity accounting
terrizoaguimor
committed
b93e524

Commits on Aug 11, 2026

fix(cli): close the markdown stream before stopping the spinner
dcutugno
committed
e6f014d
v4: expert-major prefill — read each distinct expert once per chunk

JustVugg
and
claude
committed
b00e353
Merge pull request #941 from JustVugg/v4/expert-union
JustVugg
authored
4b9b47d
Merge pull request #930 from terrizoaguimor/fix/cnre-simulator-accounting
JustVugg
authored
2c8ce27
Merge pull request #942 from dcutugno/fix/cli-attached-reply-tail
JustVugg
authored
c40924e
fix: reject GLM top-k larger than expert count
benmaster82
committed
639d0c0
style: add trailing newline to top-k test
benmaster82
committed
11fe383
chore: version 1.6.0

JustVugg
and
claude
committed
52482d7
v4: support arm64 macOS — one mach branch, the rest was already portable

JustVugg
and
claude
committed
57f0abd
fixup: the Makefile.deepseek-v4 internal gate never got the Darwin arm

JustVugg
and
claude
committed
25bd8b7
fixup: the tiny oracle regenerates its fixture — give macOS the pinned deps

JustVugg
and
claude
committed
a169bec
Merge pull request #944 from benmaster82/fix/glm-topk-bounds
JustVugg
authored
54c970f
fixup: torch +cpu wheels are Linux-only — split the pin by platform

JustVugg
and
claude
committed
59e40e2
fixup: keep Windows in the torch pin — +cpu wheels exist there too

JustVugg
and
claude
committed
f03bcea
Merge pull request #946 from JustVugg/v4/macos
JustVugg
authored
33302b8
fix(api): cancel a generation that disconnects before the first frame (#908)
gouravkargwal
committed
aa4c51a
fix(omp): skip the spin-wait tuning on macOS — measured slower (#707)
gouravkargwal
committed
a20c7aa
convert_inkling: support output locking on Windows
Blakeolson21
committed
1b10ddc
coli: select python.exe from Windows project venvs
Blakeolson21
committed
e47bf00
convert: make multistream shard writes portable on Windows
Blakeolson21
committed
d19b161
Merge pull request #962 from Blakeolson21/blake/convert-windows-positioned-write
JustVugg
authored
1d21eaa
Merge pull request #961 from Blakeolson21/blake/coli-windows-venv-python
JustVugg
authored
f158906
Merge pull request #960 from Blakeolson21/blake/inkling-windows-lock
JustVugg
authored
58ac189
Merge pull request #956 from gouravkargwal/fix/server-pre-first-frame-cancel
JustVugg
authored
600c1a0
Merge pull request #957 from gouravkargwal/fix/omp-macos-tuning
JustVugg
authored
34513cc

Commits on Aug 12, 2026