Merge v0.20.0 - #6

Open

billycao wants to merge 58 commits into

mainfrom

billy/synthetic-dev-v0.20.0

Commits on Jan 26, 2026

[Bugfix] Fix Dtypes for Pynccl Wrapper (#33030)

authored and
khluu
committed
cf1167e

Commits on Jan 28, 2026

[CI] Fix AssertionError: MCP tool call not found in output_messages (#33093)

chaunceyjiang
authored and
khluu
committed
afb390a
[Models] Kimi-K2.5 (#33131)

authored and
khluu
committed
5042815
[Bugfix] Disable CG for Whisper+FA2 (#33164)

NickLucche
authored and
khluu
committed
d51e1f8
[Bugfix] Fix DeepseekV32 `AssertionError: num_kv_heads == 1` (#33090)

NickLucche
authored and
khluu
committed
0d8ce32
[BugFix] Fix P/D with non-MoE DP (#33037)

njhill
authored and
khluu
committed
7779de3
Relax protobuf library version constraints (#33202)

jeffreywang88
authored and
khluu
committed
5f7f9ea
Revert "Enable Cross layers KV cache layout at NIXL Connector (#30207)" (#33241)

orozery
and
khluu
committed
fe18ce4

Commits on Jan 29, 2026

[Release] [CI] Optim release pipeline (#33156)

tjtanaa
authored and
khluu
committed
f176443

Commits on Feb 2, 2026

[Bugfix] Register fp8 cutlass_group_gemm as supported for only SM90+SM100 (#33285)

mgoin
authored and
khluu
committed
39e8b49
[Bugfix] Fix Qwen3-VL-Reranker load. (#33298)

authored and
khluu
committed
1ed963d
[Bugfix] Enable Triton MoE for FP8 per-tensor dynamic (#33300)

mgoin
authored and
khluu
committed
6ff16b7
[release] Minor fixes to release annotation and wheel upload (#33129)
khluu
committed
a2dba55
[Bugfix][ROCm] Fixing the skinny gemm dispatch logic from #32831 (#33366)

gshtras
authored and
khluu
committed
5f45b0b
[BugFix] Fix whisper FA2 + full cudagraphs (#33360)

authored and
khluu
committed
d984d66
[fix][torch.compile] Fix cold-start compilation time increase by adding kv cache update to splitting ops (#33441)

authored and
khluu
committed
2915268
fix: Add SM120 (RTX Blackwell) support for FlashInfer CUTLASS NVFP4 MoE kernels (#33417)

authored and
khluu
committed
15ebd0c
pin LMCache to v0.3.9 or greater with vLLM v0.15.0 (#33440)

authored and
khluu
committed
c7039a8
[Models] Step-3.5-Flash (#33523)

authored and
khluu
committed
8b45c58
[Nightly CI] Remove CT Model (#33530)

authored and
khluu
committed
94cbe0a
[Fix] prefix cache hit rate == 0 bug with gpt-oss style models (#33524)

ivanium
authored and
khluu
committed
f0d0058
[Release] patch step3p5 attention class in v0.15.1 release (#33602)

zhewenl
and
zhewenli
authored
57eae2f

Commits on Feb 3, 2026

[Release] Fix format and cherry-pick (#33618)

zhewenl
and
zhewenli
authored
31a64c6
Patch aiohttp for CVE-2025-69223 (#33621)

zaristei2
and
zaristei
authored
099a787
Patch Protobuf for CVE 2026-0994 (#33619)

zaristei2
and
zaristei
authored
7c023ba
[Misc][Build] Lazy load cv2 in nemotron_parse.py (#33189)

kiersten-stokes
authored and
khluu
committed
eec3546
[torch.compile] Speed up MOE handling in forward_context (#33184)

zou3519
authored and
khluu
committed
611b187
[torch.compile] Don't do the fast moe cold start optimization if there is speculative decoding (#33624)

authored and
khluu
committed
e4bf6ed

Commits on Feb 4, 2026

[Bugfix] Disable RoutingMethodType.[Renormalize,RenormalizeNaive] TRTLLM per-tensor FP8 MoE (#33620)

mgoin
authored and
Robert Shaw
committed
daa2784
cherry pick

mgoin
authored and
Robert Shaw
committed
7d98f09
[BugFix][Spec Decoding] Fix negative accepted tokens metric crash (#33729)

njhill
authored and
Robert Shaw
committed
1892993

Commits on Feb 6, 2026

Commit Synthetic Dockerfile
reissbaker
committed
75f6f61
Merge pull request #2 from synthetic-lab/kimi-autoclose-reasoning
reissbaker
authored
375df17
Add --max-request-secs to vllm serve (#1)
billycao
authored
5317c2c

Commits on Feb 7, 2026

Commits on Feb 10, 2026

Fix typo
reissbaker
committed
d95781a
cu130 Dockerfile
reissbaker
committed
b2c19f5

Commits on Feb 12, 2026

Fix tool call parsing for Kimi K2.5
reissbaker
committed
6ca6308
Merge pull request #5 from synthetic-lab/fix-tool-call-parsing
reissbaker
authored
6359dc3

Commits on Mar 13, 2026

Merge main into synthetic-dev: preserve timeout feature and Kimi K2.5 tool parser, adopt upstream reasoning parser
reissbaker
committed
abd72c1

Commits on Mar 14, 2026

Update dockerfiles
reissbaker
committed
eb8d783
Fix borked merge
reissbaker
committed
7ba3c3c

Commits on Apr 2, 2026

Merge commit 'c9a9db0e023a03f7e2c40e43be302fb07512946a' into synthetic-dev
reissbaker
committed
580b738

Commits on Apr 6, 2026

Update base image
reissbaker
committed
fb93877

Commits on Apr 8, 2026

Merge commit '2488d1dca2df05059fcfbad0a1612ef2a5202b47' into synthetic-dev
reissbaker
committed
5a0119a

Commits on Apr 27, 2026

Revert "[Frontend] Remove frontend pooling multi task support. (#37861)"
khluu
committed
5eb3657
[Bugfix] Cap SWA/chunked-local runtime admission to startup pool-sizing bound (#40946)

authored and
khluu
committed
84c276d
[DSV4] Add silu clamp limit to shared expert (#40950)

zyongye
authored and
khluu
committed
b816087
[Docker] Install numactl CLI in CUDA runtime image (#41032)

authored and
khluu
committed
88d34c6

Commits on May 1, 2026

update cu130.Dockerfile
billycao
committed
f6b36ed
Merge tag 'v0.20.0' into synthetic-dev
billycao
committed
d8980d4

Commits on May 6, 2026

Update Dockerfile
billycao
committed
2ae6e2d
Add test_kimi_k2_standalone.py
billycao
committed
e6dd372
Use v0.20.0 instead of nightly-01d4d1
billycao
committed
71fab0e