release 1.12.1: dev to main - #1663

Merged

JustVugg merged 364 commits into

mainfrom

Sep 24, 2026

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 21, 2026

feat(glm): continue an open assistant turn on the GLM-5.2 base renderer

enitimeago
and
claude
committed
c7445cc
feat(olmoe): continue an open assistant turn

enitimeago
and
claude
committed
f9efa23
feat(v4): continue an open assistant turn on deepseek_v4

enitimeago
and
claude
committed
e62dbed
feat(inkling): continue an open assistant turn

enitimeago
and
claude
committed
87c4156
test(serve): only the trailing assistant turn is opened, every family

enitimeago
and
claude
committed
82a3e5c
test(serve): continuation reaches the engine and returns as content, end-to-end

enitimeago
and
claude
committed
9b1151d
feat(kimi): continue an open assistant turn, framed in the engine

enitimeago
and
claude
committed
f461050
feat(v41): continue an open assistant turn on deepseek_v41

enitimeago
and
claude
committed
2327856
test(glm53): pin the continuation across every template-expressible effort

enitimeago
and
claude
committed
7d1fa95
tests(qwen38): normalize fallback test line endings
Stamina9
authored
e57b487
ci: upgrade Node.js 20 to 22 in release pipeline (Node 20 is EOL since Apr 30 2026)
Suraj2105-1
committed
1212b0b
ci: upgrade Alpine 3.21 to 3.24 in musl job (3.21 EOL Nov 1 2026)
Suraj2105-1
committed
e514e60
fix(qwen38): reject pins whose attention rows were overwritten
bokiko
committed
024bef5
fix: resolve static analysis bugs in coli, autotune, family_registry, and convert_qwen36
Suraj2105-1
committed
be958c1
qwen36: split added tokens before the pre-tokenizer, and the three whitespace rules (#1653)
JustVugg
committed
682f62d
deepseek_v4: the numeric channel (SUBMIT logprobs=k, pin=1), so brio works on the ninth engine (#1648)
JustVugg
committed
6f152cd
Merge pull request #1654 from JustVugg/fix/qwen36-special-after-punct
JustVugg
authored
d99eb3f
tok.h: the GPT-2 pre-tokenizer family, so OLMoE tokenizes like HF
JustVugg
committed
26e000b
Merge pull request #1655 from JustVugg/feat/deepseek-v4-brio
JustVugg
authored
2edf748
qwen36: offer the rest of the dense trunk to the VRAM placer (dnout, attnproj, shexp)
JustVugg
committed
6203a14
tests: two reads of available memory may differ, compare within a quarter of a GB
JustVugg
committed
6719b2c
Merge pull request #1658 from JustVugg/fix/test-mem-available-two-reads
JustVugg
authored
472450f
Merge pull request #1656 from JustVugg/fix/tok-gpt2-family
JustVugg
authored
4d96d7c
qwen36, qwen38: max_tokens is a ceiling, not a target (#1641)
JustVugg
committed
5c6a318
Merge pull request #1659 from JustVugg/fix/serve-max-tokens-ceiling
JustVugg
authored
12e7101
qwen36: measure the trunk placement at startup, withdraw it when the GPU loses
JustVugg
committed
0e84b37
fix(comment): stale "qdw entry" reference after the g_qdw removal
jtinbergen
committed
f9674b6
Merge pull request #1657 from JustVugg/feat/qwen36-trunk-vram
JustVugg
authored
3b2ebb3
release: 1.12.1
JustVugg
committed
da7bc4f
gateway: POST /v1/systemone, the request and reply of Jev, served by the brio channel
JustVugg
committed
48c974d
Merge pull request #1660 from JustVugg/release/1.12.1
JustVugg
authored
66ffa1d
Merge remote-tracking branch 'origin/dev' into feat/systemone-jev-api
JustVugg
committed
2a61e58
changelog: /v1/systemone under 1.12.1
JustVugg
committed
6954935
qwen36: integer dot products for the dense trunk and the routed experts
JustVugg
committed
dfec3a4
Merge pull request #1662 from JustVugg/feat/systemone-jev-api
JustVugg
authored
24dad80
Merge pull request #1664 from JustVugg/perf/qwen36-dense-idot
JustVugg
authored
391168c
changelog: 1.12.1 counts 32 pull requests, dated the day of the tag
JustVugg
committed
3c164f6

Commits on Sep 22, 2026

Merge pull request #1665 from JustVugg/release/1.12.1-notes
JustVugg
authored
87eb980
fix(qwen38): merge dev while preserving prefix and dot-product dependencies
bokiko
committed
e958c0e
fix(oracle): reconcile validation with current dev
ekalb81
committed
4fb195c
qwen38: int8 trunk with integer dot products, vector FP8 expert kernel
JustVugg
committed
0137b8f
test_qwen38_native_weights: pin the FP8 dispatch to the table kernel
JustVugg
committed
b8e23ac
qwen38: rebuild the trunk table per load; the fixture checks run the int8 trunk too
JustVugg
committed
85b7943
qwen38: the trunk threshold and skip list are read per load, not cached forever
JustVugg
committed
dd10808
fix(v41): size gateway expert cache from the RAM budget
ZacharyZcR
committed
6c78834
fix(v41): clamp output budgets and reject oversized prompts
ZacharyZcR
committed
71622eb
Merge pull request #1650 from bokiko/fix/qwen38-stale-pin
JustVugg
authored
8d546fa
test(v41): give the protocol mock an explicit cache cap
ZacharyZcR
committed
42f640d
Merge origin/dev into perf/qwen38-idot-trunk
JustVugg
committed
0d8dd39
test(systemone): resolve the scoring fixture during discovery
ZacharyZcR
committed
e82567a
cuda: reuse streaming MXFP4 weight scratch
ZacharyZcR
committed
b6ac01e
qwen36: offload batched attention projections to CUDA
ZacharyZcR
committed
8e8447c
docs: qwen38 int8 trunk and vector FP8 kernel, measured; 1.12.1 notes for #1650 and #1668
JustVugg
committed
cde709c
cuda: fix resident MXFP4 scale layout and accounting
ZacharyZcR
committed
a65ea15
kimi: keep CUDA SiTU expert intermediates on device
ZacharyZcR
committed
cc27d98
Merge pull request #1668 from JustVugg/perf/qwen38-idot-trunk
JustVugg
authored
d99dc06
Merge remote-tracking branch 'upstream/dev' into feat/kimi-cuda-situ-pipeline
ZacharyZcR
committed
f2c305f
qwen36: batch resident DeltaNet input projections
ZacharyZcR
committed
0fc47cc
cuda: reuse SiTU expert staging across calls
ZacharyZcR
committed
ba5ca29
test: account for optional MXFP4 expert loader symbol
ZacharyZcR
committed
9d1c441
Merge dev and retain V4.1 budget and Qwen test targets
ZacharyZcR
committed
25507df
fix(cuda): release Qwen tier resources on shutdown
ZacharyZcR
committed
0977650
fix(cuda): reclaim partially uploaded expert tensors
ZacharyZcR
committed
81bbae1
fix(cuda): budget expert scale allocations separately
ZacharyZcR
committed
1d6bfe2
docs/qwen38: remove stale "no GPU backend" line (GPU tier merged in 1.12.0)
XBold
committed
b9401a9
README: align Qwen3.8 entries with the shipped CUDA tier (1.12.0)
XBold
committed
bff0f12
fix(cuda): propagate Qwen expert collection failures
ZacharyZcR
committed
e9c45f6
fix(cuda): unwind failed Qwen tier initialization
ZacharyZcR
committed
fe29def
qwen38: make the prefill chunk and its workspace runtime knobs

DebugSultan
and
claude
committed
8745651
qwen38: lift the top-k ceiling off the expert load batch

DebugSultan
and
claude
committed
9c0c6bc
qwen38: run QSA per position at prefill, and time it by the clock

DebugSultan
and
claude
committed
c7f5b7f
fix(cuda): unwind partially initialized device contexts
ZacharyZcR
committed
33900c2
feat(serve): expose Prometheus scheduler metrics
ZacharyZcR
committed
90ff72b
feat(serve): measure first output and engine call durations
ZacharyZcR
committed
6567e76
feat(tools): add reproducible HTTP serving benchmark
ZacharyZcR
committed
9885e89
fix(cuda): stop parked issuers before teardown
ZacharyZcR
committed
510cfa5
test(cuda): group tier regression rules with related tests
ZacharyZcR
committed
2eb1975
test(cuda): group tier regression rules with related tests
ZacharyZcR
committed
f5ec92f
test(cuda): group tier regression rules with related tests
ZacharyZcR
committed
39adcc6
test(cuda): group tier regression rules with related tests
ZacharyZcR
committed
09f1403
test(cuda): group tier regression rules with related tests
ZacharyZcR
committed
127b077
test(cuda): cover FP8 and per-row int4 upload rollback
ZacharyZcR
committed
b9c679a
feat(bench): report latency-constrained request goodput
ZacharyZcR
committed
0e939fc
feat(bench): schedule fixed-rate arrivals and account for client backlog
ZacharyZcR
committed
df6cf53
fix(cuda): preserve active contexts on repeated initialization
ZacharyZcR
committed
f0f8dfc
fix(cuda): reject initialization of an active Qwen tier
ZacharyZcR
committed
cba5f31
fix(bench): reject malformed choices and unsuccessful stream endings
ZacharyZcR
committed
2c442da
test(qwen36): verify each attention projection failure with poisoned output
ZacharyZcR
committed
f17038d
bench(cuda): measure paired resident projection batching
ZacharyZcR
committed
aee54ce
test(qwen36): verify split prefill state and continued decode
ZacharyZcR
committed
35c39ca
fix(serve): check cancellation before admitting an available slot
ZacharyZcR
committed
24788ac
fix(serve): enforce zero queue limit for busy pinned slots
ZacharyZcR
committed
5ef694d
feat(bench): separate warmup from measured serving requests
ZacharyZcR
committed
282b743
build: avoid conflicts between CUDA benchmark and test targets
ZacharyZcR
committed
44150b6
fix(serve): enforce queue deadlines before slot admission
ZacharyZcR
committed
de27ba0
fix(bench): reject choice chunks after stream completion
ZacharyZcR
committed
03f6aa8
test(bench): cover partial engine failure and recovery
ZacharyZcR
committed
3059afb
test(cuda): verify attention prefill continuation and KV state
ZacharyZcR
committed
48795d4
fix(serve): admit requests to unreserved free slots
ZacharyZcR
committed
1a05695
fix(bench): account for legacy function call output
ZacharyZcR
committed
b1c9d52
Add reproducible Poisson arrivals to HTTP benchmark
ZacharyZcR
committed
965d175
Admit unreserved free slots when the waiting queue is full
ZacharyZcR
committed
2608b0f
Stop stream keepalive threads on generation errors and cancellation
ZacharyZcR
committed
0d91cdc
Merge pull request #1669 from ZacharyZcR/fix/systemone-test-discovery
JustVugg
authored
73f6a6b
Merge pull request #1670 from ZacharyZcR/fix/1666-v41-ram-budget
JustVugg
authored
9f49a80
Merge pull request #1671 from ZacharyZcR/fix/v41-output-budget
JustVugg
authored
f24abbc
Merge pull request #1673 from ZacharyZcR/perf/cuda-mxfp4-scratch
JustVugg
authored
3c0b387
Merge pull request #1675 from ZacharyZcR/fix/cuda-mxfp4-scale-layout
JustVugg
authored
8ca90cd
Merge pull request #1679 from ZacharyZcR/fix/qwen-cuda-upload-rollback
JustVugg
authored
ccf32ec
Merge pull request #1680 from ZacharyZcR/fix/qwen-cuda-scale-budget
JustVugg
authored
86f4567
Merge pull request #1682 from ZacharyZcR/fix/qwen-cuda-take-error
JustVugg
authored
62f6a47
Merge pull request #1683 from ZacharyZcR/fix/qwen-cuda-init-cleanup
JustVugg
authored
6a25042
Merge pull request #1684 from ZacharyZcR/fix/cuda-context-init-unwind
JustVugg
authored
bc54327
Merge pull request #1678 from ZacharyZcR/fix/qwen-cuda-tier-shutdown
JustVugg
authored
f969118
Merge pull request #1674 from ZacharyZcR/perf/qwen36-batch-dense
JustVugg
authored
ae48037
Merge pull request #1677 from ZacharyZcR/perf/qwen36-dnproj-prefill
JustVugg
authored
94cef8c
Merge dev into feat/kimi-cuda-situ-pipeline: keep both MXFP4 scratch sets, both cuda-test additions
JustVugg
committed
7edf6eb
CHANGELOG 1.12.1: the eighteen pull requests merged today
JustVugg
committed
31b0905
coli: resolve the launcher's own symlink before deriving libexec (#1689)
JustVugg
committed
b197c56
CHANGELOG 1.12.1: #1691
JustVugg
committed
b5944eb
Merge pull request #1686 from DebugSultan/prefill-batching
JustVugg
authored
52be8e1
Merge pull request #1687 from ZacharyZcR/feat/serve-prometheus-metrics
JustVugg
authored
a4e18f9
Merge pull request #1688 from ZacharyZcR/feat/http-serving-benchmark
JustVugg
authored
ed7f535
Merge pull request #1681 from XBold/docs/qwen38-remove-stale-gpu-backend-claims
JustVugg
authored
3961574
tools(qwen36): --down-bits, the mixed expert layout (int4 gate/up, int8 down in one slab)

kreuzzelg
and
claude
committed
5fbf3a6
qwen36: read the mixed expert layout (int4 gate/up, int8 down) on the CPU path

kreuzzelg
and
claude
committed
87d4139
ci(qwen36), docs: the mixed expert layout on the tiny fixture, and the #1370 numbers in "Which container?"

kreuzzelg
and
claude
committed
e1d51be
CHANGELOG 1.12.1: #1559
JustVugg
committed
2b35814
fix(serve): check every server->engine stdin write
monotophic
committed
40abd32
Merge pull request #1691 from JustVugg/fix/launcher-realpath
JustVugg
authored
d1c2bca
Merge pull request #1676 from ZacharyZcR/feat/kimi-cuda-situ-pipeline
JustVugg
authored
a926d1b
Merge pull request #1690 from JustVugg/release/1.12.1-notes-batch
JustVugg
authored
9ddd00a
Merge pull request #1559 from kreuzzelg/qwen36-down-bits
JustVugg
authored
2051c18
Merge pull request #1651 from Suraj2105-1/fix/pyflakes-code-bugs
JustVugg
authored
9e2555a
Merge pull request #1649 from Suraj2105-1/fix/alpine-3.21-eol-upgrade-to-3.24
JustVugg
authored
ac94bb0
Merge dev into fix/serve-tool-call-arguments-non-object: keep both tool-call regressions
JustVugg
committed
5480ea5
Merge pull request #1647 from Suraj2105-1/fix/node20-eol-release-pipeline
JustVugg
authored
96ef114
Merge dev into fix/windows-qwen36-cuda-dll: keep dev's qwen36 prerequisites plus the alias and .build-config
JustVugg
committed
9fc4ca3
Merge dev into fix/macos-metal-gpu-discovery: keep the int8 trunk line in the VRAM branch
JustVugg
committed
232b282
Merge pull request #1624 from namespaceMarcello/fix/build-warnings
JustVugg
authored
11106c9
Merge pull request #1618 from bherald/docs/qwen38-vision-overview-20260918
JustVugg
authored
0eb3a87
CHANGELOG 1.12.1: the third wave of contributor pull requests
JustVugg
committed
9e85af7
Merge pull request #1568 from Yoruxyv/feat/indonesian-dashboard-translation
JustVugg
authored
b6955b6
Merge pull request #1622 from Frank-zhu0404/fix/issue-1615-per-matrix-direct-read
JustVugg
authored
4efbb8d
Merge pull request #1605 from ekalb81/fix/glm-oracle-validation
JustVugg
authored
9e152d4
refactor(quant): move FP8_BLOCK/fp8_nblk into fp8_format.h
monotophic
committed
6826c00
feat(colibri): add fmt=8 arms to the CPU absorb path
monotophic
committed
0ad3eef
feat(cuda): add fmt=8 (fp8-e4m3) decode to the absorb kernels
monotophic
committed
9c7ae38
test(fp8): e2e serve-batch pin, refusal canary, and mint-to-load regression
monotophic
committed
ea8a7e4
feat(api): accept and ignore a per-request seed
monotophic
committed
7d79ad1
Merge pull request #1597 from kevin9327/fix/serve-tool-call-arguments-non-object
JustVugg
authored
d8c4209
fix(olmoe): wait for an expert already being read instead of reading it twice

namespaceMarcello
and
claude
committed
2020aec
Merge pull request #1580 from LTCjRet/fix/windows-qwen36-cuda-dll
JustVugg
authored
5f1cb69
Merge pull request #1556 from karlem/fix/macos-metal-gpu-discovery
JustVugg
authored
2cc1040
Merge pull request #1692 from JustVugg/release/1.12.1-notes-wave3
JustVugg
authored
b94accc
Merge dev into feat/glm53-assistant-continuation: keep the max_tokens ceiling paragraph next to the continuation one
JustVugg
committed
836310f
CHANGELOG 1.12.1: #1402
JustVugg
committed
7dbe7e3
perf(sse41): vectorize grouped-int4 GEMV

codex 20260829-144356-267014324
authored and
cameron
committed
4d35744
fix(sse41): preserve grouped-int4 scalar order

codex 20260830-005353-2548412838
authored and
cameron
committed
e09b3db
fix(build): track SSE4.1 kernel header dependencies

codex 20260920-012741-906614622
authored and
cameron
committed
a3ec030
qwen36: keep the prefill echo out of the segment build

namespaceMarcello
and
claude
committed
8212a39
Merge pull request #1694 from JustVugg/release/1.12.1-notes-1402
JustVugg
authored
6497713
Merge pull request #1402 from enitimeago/feat/glm53-assistant-continuation
JustVugg
authored
b02b801
fix(web): restore the reasoning stream the redesign dropped
kevin9327
committed
11eb32e

Commits on Sep 23, 2026

docs(fp8): correct stale anchors, an unreachable test rule and a wrong test comment
monotophic
committed
b95b942
docs(fp8): FORMATS.md names symbols, so stop promising line numbers
monotophic
committed
4159a23
Opt-in exact verify mode for speculative batches (COLI_EXACT_VERIFY=1, #689)
rybruscoe
committed
049d444
Makefile: list exact_dot.h as a colibri prerequisite
rybruscoe
committed
78018c3
exact verify (#689): document the quantised-KV fallback and the untested near-tie flip
rybruscoe
committed
a13b3cd
CHANGELOG 1.12.1: #1693, #1695, #1697
JustVugg
committed
6035bcd
Merge pull request #1695 from namespaceMarcello/fix/qwen36-echo-guard
JustVugg
authored
bf7c767
Merge pull request #1697 from kevin9327/fix/web-reasoning-stream-dropped
JustVugg
authored
4399eef
Merge pull request #1693 from namespaceMarcello/fix/olmoe-inflight-dedup
JustVugg
authored
aa2d295
feat(tools): add three-engine serving baseline campaigns
ZacharyZcR
committed
577d8eb
Merge pull request #1704 from JustVugg/release/1.12.1-notes-wave4
JustVugg
authored
5d6671f
CHANGELOG 1.12.1: #1705
JustVugg
committed
1f82285
Merge pull request #1705 from ZacharyZcR/feat/three-engine-serving-baseline
JustVugg
authored
926be8e
Merge pull request #1706 from JustVugg/release/1.12.1-notes-1705
JustVugg
authored
30c70ad
fix(v4): rebuild the unit objects when the build flags change

namespaceMarcello
and
claude
committed
4e8ef20
fix(glm53): close the Vulkan status fprintf after the ternary

wittchen
and
claude
committed
1165bcd
fix(web): regenerate the last turn with its pictures
kevin9327
committed
4653c34
fix(serve): a tool whose function is not an object answered 500
kevin9327
committed
c64c8d6
fix(coli): Windows GPU probe missed Kimi CUDA_DLL and HIP hosts
kevin9327
committed
6d616c5
fix(serve): kimi, inkling, olmoe clamp max_tokens to the context
kevin9327
committed
1627377
fix(planner): export K3_EXPERT_GB and GLM53_EXPERT_GB from the plan
kevin9327
committed
b1b9c60
test(glm53): rename the harnesses out of the unittest glob, run the tiny oracles in CI

namespaceMarcello
and
claude
committed
69c3b0f
fix(brio): pin the shared state in the options form so per-question requests reuse it
benmaster82
committed
e1feb15
Merge branch 'dev' into sse41-tier-upstream-next-v2 (+25 commits)
jtinbergen
committed
dd1cf4e
feat(brio): let the web cancel a scoring run, dedupe the CLI options
benmaster82
committed
d13dec7
kimi_k3: keep K3_CUDA the only engine switch; coli maps --gpu onto it
JustVugg
committed
2bc08d0
Merge pull request #1710 from kevin9327/fix/serve-tool-function-non-object
JustVugg
authored
b8869a2
Merge pull request #1712 from kevin9327/fix/k3-ink-olmoe-max-tokens-ceiling
JustVugg
authored
3431e87
Merge pull request #1713 from kevin9327/fix/planner-kimi-expert-gb
JustVugg
authored
72e49d2
Merge pull request #1707 from namespaceMarcello/fix/v4-unit-flags-stamp
JustVugg
authored
c0a3fb4
Merge pull request #1708 from wittchen/fix/glm53-vk-fprintf
JustVugg
authored
c1746e0
Merge dev: keep dev's build-flags stamp (#1707) and serve_budget.h (#1712) next to this branch's header
JustVugg
committed
77e6a0d
Merge dev: keep dev's build-flags stamp (#1707) and serve_budget.h (#1712) next to this branch's header
JustVugg
committed
bf4ac29
Merge dev: fallback diagnostics on the heap-allocated batch of #1686
JustVugg
committed
c07f62c
Merge pull request #1711 from kevin9327/r8b-win-cuda-dll-probe
JustVugg
authored
f89040b
Merge pull request #1709 from kevin9327/fix/web-regenerate-drops-images
JustVugg
authored
7a82be4
Merge pull request #1395 from rybruscoe/exact-verify-689
JustVugg
authored
d9e7bea
Merge dev: colibri prerequisites keep exact_dot.h (#1395) and fp8_format.h
JustVugg
committed
b4e51d5
Merge pull request #1102 from monotophic/f8/absorb-fmt8
JustVugg
authored
1d7defa
Merge dev: keep #1102's fp8_format.h, #1395's exact_dot.h and the new CPU-only test rules next to this branch's sse41_kernels.h and SSE4.1 gates
JustVugg
committed
621ae67
Merge pull request #1646 from Stamina9/codex/fix-qwen38-parallel-fallback
JustVugg
authored
66167b6
Merge pull request #1721 from monotophic/serve/checked-engine-writes
JustVugg
authored
97622ad
Merge pull request #1720 from monotophic/api/seed-accepted
JustVugg
authored
de563da
Merge pull request #1719 from benmaster82/feat/brio-web-cancel-and-cli-dedup
JustVugg
authored
017fe1b
Merge pull request #1714 from benmaster82/fix/brio-options-state-pin
JustVugg
authored
e2de147
Merge pull request #1715 from namespaceMarcello/test/glm53-harness-discovery
JustVugg
authored
7a5bdee
CHANGELOG 1.12.1: sixteen more contributor pull requests
JustVugg
committed
8db487f
refactor(sse41): collect grouped-int4 kernels in the shared header

codex 20260923-120015-1331522267
authored and
cameron
committed
3b24221
Merge pull request #1286 from cameron/perf/sse41-grouped-int4-upstream
JustVugg
authored
4aac292
Merge pull request #1722 from JustVugg/release/1.12.1-notes-wave5
JustVugg
authored
33422c3
qwen36: decode added tokenizer tokens

GenericRikka
authored and
JustVugg
committed
9342b47
qwen36: decode only the non-special added tokens
JustVugg
committed
93bc084
Merge pull request #1724 from JustVugg/fix/qwen36-decode-added-tokens
JustVugg
authored
20183ac
CHANGELOG 1.12.1: #1724
JustVugg
committed
09b78b1
Merge pull request #1725 from JustVugg/release/1.12.1-notes-1724
JustVugg
authored
a2e578c
fix(inkling): measure RAM on Windows so auto-cap is not 16 experts
kevin9327
committed
c703d26

Commits on Sep 24, 2026

tests: link backend_vulkan.o into the engine-including tests under VK=1
crichalchemist
committed
37d9367
Merge branch 'dev' into sse41-tier-upstream-next-v2 (+214 commits)
jtinbergen
committed
2c7979f
feat(web): continue a trailing assistant turn from the chat UI

enitimeago
and
claude
committed
f4ee2e0
feat(web): show Continue only on an incomplete assistant turn

enitimeago
and
claude
committed
6923b7d
feat(api): report continue_assistant in the authed /health

enitimeago
and
claude
committed
eb592cc
feat(web): offer Continue only when /health reports continue_assistant

enitimeago
and
claude
committed
7743fb4
feat(web): keep each reply's finish reason on the message itself

enitimeago
and
claude
committed
96467ec
tools: add FP4 expert matmul microbench (SIMD vs scalar arm)

mfethe1
and
claude
committed
fa86687
dsv4: NEON arm for coli_fp4_matmul_batch_rows16_order (#1696)

mfethe1
authored and
mfethe1
committed
da68961
fix(v4): rebuild backend_cuda_dsv4.o when the nvcc command changes
bokiko
committed
9ba9ed8
dsv4: NEON arm for coli_fp4_matvec_rows16_order (decode path, #1696)
mfethe1
committed
75d2677
serve: give the engine its graceful exit — stdin drain, SIGBREAK, group isolation
tarazum
committed
49e72cd
make: write .build-config through printf on GNU Make 3.x (#1732)
JustVugg
committed
b6fbb65
ci: prove the NEON FP4 arms bit-exact against the scalar arm on the ARM runner
JustVugg
committed
4087d5a
Merge pull request #1717 from enitimeago/feat/web-continue
JustVugg
authored
6b5671c
Merge pull request #1726 from kevin9327/fix/inkling-windows-ram-probe
JustVugg
authored
a8de807
Merge pull request #1728 from crichalchemist/vk-link-colibri-tests
JustVugg
authored
d23e362
Merge pull request #1731 from bokiko/fix/v4-cuda-arch-stamp
JustVugg
authored
2516472
close(): decide the hard-stop ladder by poll(), not by TimeoutExpired
tarazum
committed
f5497c8
Merge pull request #1735 from JustVugg/fix/build-config-make381
JustVugg
authored
78d56c2
Merge pull request #1734 from tarazum/fix/heat-save-stdin-drain
JustVugg
authored
50c9e16
Merge pull request #1730 from mfethe1/perf/dsv4-fp4-neon-arm
JustVugg
authored
20e4582
Merge pull request #1716 from jtinbergen/sse41-tier-upstream-next-v2
JustVugg
authored
53ca388
CHANGELOG 1.12.1: eight more pull requests
JustVugg
committed
5e11f46
ci: give the macOS make check 25 minutes
JustVugg
committed
4f119c0
ci: the Linux make check needs the same room
JustVugg
committed
73fcbcf
docs(changelog): date 1.12.1 to the release day
JustVugg
committed
25b4a9f
Merge pull request #1737 from JustVugg/ci/macos-timeout
JustVugg
authored
a44865f
Merge branch 'dev' into release/1.12.1-notes-wave6
JustVugg
authored
384beb4
Merge pull request #1736 from JustVugg/release/1.12.1-notes-wave6
JustVugg
authored
1f42b14