[Kimi] Support kimi-k3 - #32541

Merged

hnyls2002 merged 92 commits into

mainfrom

Aug 4, 2026

Commits on Jul 27, 2026

[Kimi] Support kimi-k3

committed
4051b19
[Kimi-K3] Clone branch source in Docker images (#32545)
Fridge003
authored
87dd459
Fix S3 speculative draft loading (#32563)

ByronHsu
and
Byron Hsu
authored
15fa94d

Commits on Jul 28, 2026

perf(kimi-k3): KDA MTP verify -8.6% at production, retire padding blocks (#32571)

zcnrex
and
claude
authored
ac6d795
[AMD] Fix top_k/p sampling issue on AMD code path (#32621)

1am9trash
and
RolaoDenthu
authored
4b6e6fc
perf: avoid redundant Kimi vision projection copies
mickqian
committed
3c2beaf
fix: preserve Kimi projector output contract
mickqian
committed
30b65a9
perf(vlm): keep packed MoonViT grid metadata on CPU
mickqian
committed
6144e9c
perf(vlm): fuse MoonViT rotary embedding on CUDA
mickqian
committed
fe7454f
perf(vlm): compute MoonViT max sequence length on host
mickqian
committed
c49fda5
fix: make image benchmark seeds reproducible
mickqian
committed
4c76511
perf(vlm): enable CUDA IPC pool handle cache by default
mickqian
committed
7935796
fix(vlm): return bad media input as client error
mickqian
committed
8b17f3e
refactor(vlm): share fused vision RoPE with Kimi K2.5
mickqian
committed
2e35ae3
fix(vlm): preserve host-precomputed vision metadata
mickqian
committed
16bf96a
perf(vlm): share fastest Kimi vision RoPE paths
mickqian
committed
9cb0351
[Kimi K3] Support parser auto-detection (#32617)
JustinTong0323
authored
f748ae3
fix(kimi-k3): named persistent symmetric buffers for the fused collectives (#32664)

DarkSharpness
and
claude
authored
b5607ec

Commits on Jul 29, 2026

perf(kimi-k3): fuse MoE-front prep (route + trtllm pack + mxfp8 quant) into one launch (#32699)
BBuf
authored
bd51cab
Fix Kimi K3 reasoning leak in the Responses API (#32757)
yhyang201
authored
c6ad1f2
chore: move Kimi K3 Docker changes to a separate PR
Fridge003
committed
6dd5b24
Revert "chore: move Kimi K3 Docker changes to a separate PR"
Fridge003
committed
cf56487
update sgl-deep-gemm to include changes in MegaMoE
Fridge003
committed
bd4828e
perf(kimi-k3): free the gemm_ag producer grid from the AR push-counter array (#32764)
BBuf
authored
40feea2
perf(kimi-k3): optimize KDA MTP ns0 memory path (#32823)
zcnrex
authored
c4d9f5f
DCP tests clean up (#32820)

kpham-sgl
and
claude
authored
bc928b7
feat(kda): optimize PTX prefill kernel (#32819)

zcnrex
and
pranjalssh
authored
f37baeb
Merge commit '61057bda6c4b8cda1117d74f100d0735645c0cfb' into kimi-k3 (07/25)
Fridge003
committed
19ba402

Commits on Jul 30, 2026

Commits on Jul 31, 2026

Commits on Aug 1, 2026

Merge commit '4f51dad1da320faf7489594ec001779a5affff66' into kimi-k3
Fridge003
committed
ac77803
fix(kimi-k3): align kimi-k3 branch with DCP + DSpark fix and cleanup (#33149)
kpham-sgl
authored
db5c256
Merge commit 'e1964da45f4a3018da01a86f30f7f61475b0e937' into kimi-k3

ch-wan
and
claude
committed
49b58a0
Merge remote-tracking branch 'origin/main' into HEAD
BBuf
committed
6d0adfc
chore: satisfy lint on the kimi-k3 branch
BBuf
committed
3b003e7
clean unit dcp test
Fridge003
committed
99b392b
test: prune the kernel-side tests this PR adds, keeping all-reduce
BBuf
committed
ef5c414
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
056971c
remove duplicate kimi k3 detector tests
hnyls2002
committed
ee560a2

Commits on Aug 2, 2026

Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
aa03024
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
9a0efa4
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
70bf37b
fix stale unified_cache_components import in streaming session test
hnyls2002
committed
4df65c9
Merge branch 'main' into kimi-k3
hnyls2002
authored
dd1b683
Fix ComponentType import in hybrid_pool_assembler
hanming-lu
committed
38aeeeb

Commits on Aug 3, 2026

Update runtime context usage and test fixtures
mmangkad
committed
872bbd2
Update encoder tests to publish runtime context
mmangkad
committed
d96faae
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
363629c
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
0001380
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
4c4bfdb
keep aiter swiglu translation for non-situ activations
hnyls2002
committed
9b34836
pass request token ids to fast_load_mm_data; skip the redundant rebuild
hnyls2002
committed
921603d
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
1aebf2f
drop vacuous tests; remove dead broadcast override
hnyls2002
committed
e7362bc
restore coalesce yield test
hnyls2002
committed
2306380
drop vacuous ar-fusion gate tests; trim tautological asserts
hnyls2002
committed
e6311f7
Merge branch 'main' into kimi-k3
JustinTong0323
authored
46f96d1
[Kimi K3] Use runtime config accessors and update linear attention tests (#33405)
JustinTong0323
authored
1f9d689
small cleaning unit tests
Fridge003
committed
d4c9a27

Commits on Aug 4, 2026

merge main
hnyls2002
committed
338dca7
drop banner dividers; class docstrings instead
hnyls2002
committed
b557c69
remove unreachable gdn fast path
hnyls2002
committed
eafa1ef
Merge remote-tracking branch 'origin/main' into kimi-k3
hnyls2002
committed
f0718b1
[CI] Add 8-GPU B300 Kimi-K3 accuracy coverage (#33482)
Fridge003
authored
ed447b6
Merge branch 'main' into kimi-k3
hnyls2002
authored
6422ff4
drop redundant codespell ignore; inline directive covers it
hnyls2002
committed
92ee870
Merge branch 'main' into kimi-k3
hnyls2002
authored
373196c
[CI] Install Kimi-K3 FlashInfer assets on B300 (#33494)
Fridge003
authored
381f833
Merge branch 'main' into kimi-k3
hnyls2002
authored
0c1e51c
fix encoder grid fallback; stub moe_tp_size on inkling loader; int32 gate for single-token align; update hybrid cpu-seq-lens test
hnyls2002
committed
2434eaf
bump kimi-vl test mem fraction off the 0.4008 load minimum
hnyls2002
committed
2a67bc5
Merge branch 'main' into kimi-k3
hnyls2002
authored
03f8cc0
rename enable_gdn_replayssm_spec to enable_linear_replayssm_spec
ispobock
committed
3a98c0b
rename the replayssm spec flag, deprecate the old one, allow extra_buffer_lazy
ispobock
committed
89561a9
rank-guard the padded moe copy; drop obsolete second warmup forward
hnyls2002
committed
c0db43f
give the mxfp4 sm90 mock layer a runner config
ispobock
committed
c33c577
Merge branch 'main' into kimi-k3
ispobock
authored
36e31ee
[AMD] Fix Kimi-K3 SiTU JIT compile failure on ROCm (#33447)
RolaoDenthu
authored
ba7b810