biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity
11,905 Commits 2 Branches 140 Tags
2902c348265639de300c95cbcae1c26486f57ac7
Commit Graph

1539 Commits

Author SHA1 Message Date
Michael Goin
e5f599d4d1 [Bugfix] Disable shared expert overlap if Marlin MoE is used (#28410) 
Signed-off-by: mgoin <mgoin64@gmail.com>
 2025-11-11 23:16:12 +00:00
Lukas Geiger
76e4dcf225 [Misc] Remove unused attention prefix prefill ops functions (#26971) 
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
 2025-11-11 18:26:04 +00:00
xuebwang-amd
5a1271d83a [Quantization] fix attention quantization of gpt_oss model (#27334) 
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
 2025-11-11 12:06:00 -05:00
xuebwang-amd
05576df85c [ROCm][Quantization] extend AMD Quark to support mixed-precision quantized model (#24239) 
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
Co-authored-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
 2025-11-11 12:05:22 -05:00
zhrrr
68c09efc37 [Kernel][Perf] fuse QK Norm and RoPE into one cuda kernel for Qwen Model (#27165) 
Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
 2025-11-11 12:00:31 -05:00
Michael Goin
f9a4087182 Remove weight_scale.T special case for SM90 Block FP8 CUTLASS kernel (#28431) 
Signed-off-by: mgoin <mgoin64@gmail.com>
 2025-11-11 11:46:04 -05:00
bnellnm
a1448b4b69 [Kernels] Split up fused_moe/layer.py, isolate more modular kernel code (#28064) 2025-11-11 07:29:02 -07:00
Robert Shaw
e605e8e323 [Bugfix] Fix Stream Sync for Shared Expert Overlap (#28430) 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
 2025-11-11 05:59:08 +00:00
Wentao Ye
de540c0354 [Feature] Add env var VLLM_MOE_USE_DEEP_GEMM (#28422) 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
 2025-11-11 02:29:48 +00:00
Wentao Ye
35d801f13f [Feature] Refactor batch invariant fp8 DeepGEMM (#27606) 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
 2025-11-11 00:08:40 +00:00
Ilya Markov
d17ecc6b19 [PERF] Allreduce fusion. Support torch native matching. Tuning of the thresholds (#24248) 
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
 2025-11-10 18:33:11 -05:00
Yong Hoon Shin
021143561f [ROCm] Add missing gemm_a8w8_blockscale import (#28378) 
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
 2025-11-10 23:13:36 +00:00
Lucas Wilkinson
6dec9f6109 [BugFix] Fix DeepGEMM over-allocating workspace (#28254) 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
 2025-11-10 17:01:17 -05:00
Sage Moore
40d33264c6 [Bugfix][EPLB] Disabled shared expert overlap when EPLB is enabled (#28377) 
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Signed-off-by: Sage Moore <sagemoore@utexas.edu>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
 2025-11-10 20:39:19 +00:00
jiahanc
34553b9d27 [Performance] Support FP8 flashinfer TRTLLM MOE on Qwen3 and Qwen-3next (#27492) 
Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
 2025-11-10 12:34:57 -05:00
Varun Sundar Rabindranath
b039bfda8f [Bugfix] Fix persistent_masked_m_silu_mul_quant tests (#28366) 
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
 2025-11-10 09:21:52 -08:00
vllmellm
f080a83511 [RFC][ROCm][AITER] Keep all AITER kernels in _aiter_ops class like _custom_ops and _ipex_ops (#24490) 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
 2025-11-10 08:20:53 -08:00
zejunchen-zejun
b06b9470ca [Rocm][fused_moe][fp4] view weight to torch.float4_e2m1fn_x2 when running aiter fused moe for fp4 model (#27474) 
Signed-off-by: zejunchen-zejun <zejun.chen@amd.com>
 2025-11-10 10:38:56 -05:00
Xiake Sun
03fa4d3fb3 [Hardware][AMD][Model] Add Triton MoE tuning support and optimized configs for Qwen3 omni for MI308X (#28373) 
Signed-off-by: Xiake Sun <xiake.sun@amd.com>
Signed-off-by: Xiake Sun <xisun@amd.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
 2025-11-10 04:53:40 +00:00
Yong Hoon Shin
de2b78305f [ROCm] Add env to enable/disable aiter triton gemm (#28321) 
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
 2025-11-08 22:27:00 -08:00
Robert Shaw
26990d25dc [Bugfix] Update device name for H200 detection (#28349) 
Signed-off-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
 2025-11-08 19:01:11 +00:00
Michael Goin
0852527647 [Perf][DeepSeek] Add sigmoid+bias fusion to fused_grouped_topk from TRTLLM (#28124) 
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
 2025-11-07 18:20:55 -08:00
Kunshang Ji
1aaecda078 [XPU] Enable Expert parallel for MoE models (#28263) 
Signed-off-by: Yan Ma <yan.ma@intel.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
 2025-11-08 00:33:11 +00:00
Pavani Majety
72b1c2ae2c [Bugfix] Use latency MOE backend as default for Flashinfer and other misc fixes (#27439) 
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
 2025-11-07 04:18:39 -08:00
smit kadvani
11fd69dd54 [amd][gptoss] Perf gain because of block alignment (#28024) 
Signed-off-by: Smit Kadvani <smit.kadvani@gmail.com>
Co-authored-by: Smit Shaileshbhai Kadvani <kadvani@meta.com>
 2025-11-07 05:27:42 +00:00
Aleksandr Malyshev
449de9001a [ROCm] triton fp8 kernel (#27058) 
Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com>
 2025-11-06 14:46:44 -05:00
Julien Denize
7a8375f8a0 Add llama 4 scaling support (#28145) 
Signed-off-by: Julien Denize <julien.denize@mistral.ai>
 2025-11-06 18:55:17 +00:00
Eric Yue
0370679ce9 [Kernel][Model] Tune fused_moe Triton configs for MiniMax-M2 on H100 (#28200) 
Signed-off-by: minatoaquaMK2 <jiacheng.yue@foxmail.com>
 2025-11-06 07:29:46 -08:00
xiangze-arm
c757a15f0f [CPU]Improve cpu fused moe perf (#27244) 
Signed-off-by: Zhang Xiangze <Xiangze.Zhang@arm.com>
 2025-11-06 11:04:18 +00:00
Xiaozhu Meng
e31946f86e [flashinfer] fix FI all2all with FI cutlass moe (#28166) 
Signed-off-by: Xiaozhu <mxz297@gmail.com>
 2025-11-06 05:52:16 +00:00
Wentao Ye
d71af5f502 [Feature] Enable TP + EP shared_experts overlap with router, 3.7% E2E performance improvement (#28164) 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
 2025-11-05 17:21:08 -08:00
Vadim Gimpelson
b6a248bdd7 [PERF] Decouple projections from GDN custom op. Attempt 2 (#28083) 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
 2025-11-05 17:01:12 -08:00
Paul Zhang
faedbb4d4f [Feature] Extend batch invariant torch.compile to B200 (#27856) 
Signed-off-by: PaulZhang12 <paulzhan@fb.com>
 2025-11-05 10:04:49 -08:00
Frost Mitchell
6e97eccf5d [XPU] Enable custom routing functions in IPEX for Llama4 (#28004) 
Signed-off-by: frost-intel <frost.mitchell@intel.com>
 2025-11-05 13:39:57 +00:00
amirkl94
6b7a81185d Bugfix: Cutlass FP8 FusedMoE bad scaling factors (#27255) 
Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
 2025-11-05 06:06:06 -05:00
Kunshang Ji
18b39828d9 [XPU] Add gpt-oss model support for Intel GPU (#27786) 
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
 2025-11-05 02:17:23 +00:00
tou
4ea62b77f5 [Qwen3-Next] MOE configs for A100-SXM4-80GB TP4 TP8 (#27740) 2025-11-05 09:25:09 +08:00
Vadim Gimpelson
d4e547bb7e Revert "[PERF] Decouple projections from GDN custom op" (#28080) 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
 2025-11-04 15:58:23 -08:00
Aleksandr Malyshev
2d977a7a9e [ROCm] gemm_a16w16 upstreaming (#26969) 
Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
 2025-11-04 16:01:00 -05:00
Vadim Gimpelson
5fd8f02ea9 [PERF] Decouple projections from GDN custom op (#27512) 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
 2025-11-04 08:11:41 -08:00
bnellnm
938772af03 [Kernels] Isolate modular kernel code from FusedMoEMethodBase subclasses. (#27123) 2025-11-04 21:59:45 +08:00
tomeras91
e4ee658672 [Model] add optimal triton fused moe configs for NemotronH MoE (#27967) 
Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
 2025-11-04 12:59:43 +00:00
Jerry Zhang
03c4c4aa9d Support using Int4PreshuffledTensor after loading (#26066) 
Signed-off-by: Jerry Zhang <jerryzh168@gmail.com>
 2025-11-04 06:00:57 -05:00
Varun Sundar Rabindranath
4022a9d279 [BugFix][Performance] Restore flashinfer autotuning for all scenarios (#27904) 2025-11-04 15:56:21 +08:00
Wentao Ye
7e4be74104 [Bug] Batch invariant: Fix flash attn MLA RuntimeError: scheduler_metadata must have shape (metadata_size) (#27884) 2025-11-04 14:05:55 +08:00
vllmellm
b13a447546 [Bugfix][ROCm] Fix ViT rotary embeddings for torch.compile compatibility on ROCm (#27748) 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
 2025-11-03 17:12:19 -08:00
Tyler Michael Smith
3758757377 [Bugfix] Fix MoE Routing Simulation (#28002) 
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
 2025-11-03 22:26:49 +00:00
Hank_
ccd3e55e51 [Bugfix][plugin] fla crash on plugin (#27322) 2025-11-04 05:27:03 +08:00
Asaf Joseph Gardin
00b31a36a2 [V1] [Hybrid] Mamba1 Automatic Prefix Caching (#26377) 
Signed-off-by: asafg <39553475+Josephasafg@users.noreply.github.com>
 2025-11-02 04:16:23 -08:00
Jee Jee Li
3a5de7d2d6 [Bugfix] Fix KDA output (#27905) 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
 2025-11-01 11:54:36 +08:00