3339 Commits
Author SHA1 Message Date
Jia GuoandClaude Opus 4.6 ec85340531 [Quantization] Support FP8 MoE bias for models like GPT-OSS (#34906)
Signed-off-by: jasperjiaguo <jasperg662@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-02-23 19:07:47 -08:00
2ff4e51152 [ROCm] AITER fused RoPE+KVCache (#33443)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Signed-off-by: charlifu <charlifu@amd.com>
Signed-off-by: Rohan Potdar <66227218+Rohan138@users.noreply.github.com>
Co-authored-by: charlifu <charlifu@amd.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
2026-02-23 19:06:00 -08:00
Michael GoinandRobert Shaw 22a97e6613 [Perf] Improve default triton fused moe configs (#34846)
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
2026-02-23 16:01:28 -08:00
haosdent a7f341c323 [Bugfix] Fix MRotaryEmbedding missing truncate attr with YaRN scaling (#35080)
Signed-off-by: haosdent <haosdent@gmail.com>
2026-02-23 16:05:52 +00:00
Eldar KurtićandYour Name b95bb6927f [kv-cache, ct] Use compressed-tensors as a source of ground-truth for quant strategies (#34254)
Signed-off-by: Your Name <you@example.com>
Co-authored-by: Your Name <you@example.com>
2026-02-23 07:37:55 -07:00
Yan Ma b1b5e045df [XPU] allow TORCH_SDPA/TRITON_ATTN as XPU vit Backend (#35010)
Signed-off-by: Yan Ma <yan.ma@intel.com>
2026-02-23 05:06:44 -08:00
Wentao Ye 7f40e9e516 [Refactor] Remove dead private func _fp8_perm and _extract_mask_for_item (#35068)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-02-23 05:05:20 -08:00
qizixiandLu Fang 2bcf71b9c0 [Spec Decode] Reduce TP communication for speculative decoding draft token generation (#34049)
Signed-off-by: qizixi <qizixi@meta.com>
Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>
2026-02-22 14:59:16 -08:00
b7892a3bef [Model] Add NVFP4 quantization support for Step3.5-Flash (#34478)
Signed-off-by: tacos8me <ian@cloudhabit.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
2026-02-22 12:30:46 -07:00
Vadim Gimpelson 74d90b1ce4 [Model Bash][DSR1] Add selective dynamic shape marking for CustomOp (#34900)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
2026-02-21 19:28:01 -05:00
BADAOUI Abdennacer 8dc8a99b56 [ROCm] Enable bitsandbytes quantization support on ROCm (#34688)
Signed-off-by: badaoui <abdennacerbadaoui0@gmail.com>
2026-02-21 00:34:55 -08:00
Rohan Potdar ded333fb9b [ROCm][Bugfix]: Only save unpadded sizes for shared_experts in MoERunner to fix rmsnorm pad fusion (#34636)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
2026-02-20 19:56:16 -08:00
Lucas Wilkinson aaefc58ee0 [CI] Revert PRs 34818 and 33600 (#34979) 2026-02-20 13:25:50 -08:00
Frank WangandWentao Ye 059779231f [Minor] Add logging when using MXFP4 MXFP8 TRTLLM backend (#34916)
Signed-off-by: frankwang28 <frank.wbb@hotmail.com>
Signed-off-by: Frank Wang <41319051+frankwang28@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
2026-02-19 22:07:57 -08:00
81bfc21a6a [Model Bash]: Improve FP8 Oracle for Config Specific Kernel Selection (#34260)
Signed-off-by: Elizabeth Thomas <email2eliza@gmail.com>
Signed-off-by: Robert Shaw <robertgshaw2-redhat@h100-02.nemg-001.lab.rdu2.dc.redhat.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <robertgshaw2-redhat@h100-02.nemg-001.lab.rdu2.dc.redhat.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
2026-02-19 21:29:08 -08:00
Bowen Bao d9e62c03eb [Quark] Fix MoE fp8 activation scale handling on mi300 (#34386)
Signed-off-by: Bowen Bao <bowenbao@amd.com>
2026-02-19 21:27:14 -08:00
Michael Goin f72061a19a [UX] More descriptive reasons in is_supported_config for MoE (#34908)
Signed-off-by: mgoin <mgoin64@gmail.com>
2026-02-19 15:20:52 -08:00
Matthew BonanniandRobert Shaw 662205d34e [Bugfix] Fix Basic Models Test (#34818)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
2026-02-19 14:49:07 -08:00
Roger Wang 4fb8beefaa [Bugfix] Fix cutlass fp8 kernel on hopper for Qwen3.5 (#34914)
Signed-off-by: Roger Wang <hey@rogerw.io>
2026-02-19 13:34:55 -08:00
Wentao Ye c683d11c94 [Refactor] Deprecate head_first for chunk_gated_delta_rule (#34263)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-02-19 13:23:49 -05:00
Robert ShawandRobert Shaw 4685a630a2 [Model Bash][DeepSeekR1] Remove Shared Expert Clone (#34344)
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
2026-02-19 07:56:14 -08:00
Manrique Vargas ad5aa6bd9f fix(docs): fix typos in comments and docstrings (#34836)
Signed-off-by: machov <mv1742@nyu.edu>
2026-02-18 23:17:41 -08:00
Yongye ZhuandRobert Shaw 40da9625a1 [MoE Refactor] Convert mxfp4 marlin into modular kernel format (#34588)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
2026-02-18 14:37:14 -08:00
847a57cd12 [Bugfix][MoE Kernel] Fix incorrect routing selection for models without expert groups (e.g., MiniMax-M2.1) (#34673)
Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
2026-02-18 13:03:24 -08:00
Jaden Mathias 0e60c925cf [Bugfix] Remove assert causing hipErrorStreamCaptureUnsupported (#34455)
Signed-off-by: Jaden Mathias <jaden.mathias@amd.com>
2026-02-18 18:54:54 +00:00
Isotr0pyandJJJYmmm c0bd8b13da [Bugfix] Redo Qwen3.5/Qwen3-Next GDN projector fusion (#34697)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com>
2026-02-18 09:46:53 -08:00
Michael Goin caeb887bf6 [Bugfix] Fix NVFP4 TRTLLM MoE non-gated support; add gsm8k for Nemotron-3-Nano FP8+NVFP4 (#34725)
Signed-off-by: mgoin <mgoin64@gmail.com>
2026-02-18 09:39:22 -08:00
Robert ShawandRobert Shaw 6874638bc4 [Model Bash] DeepSeek R1 BF16 Min Latency QKV A GEMM (0.5% E2E Speedup) (#34758)
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
2026-02-18 07:42:36 -08:00
Asaf Joseph Gardin 1faa8cb73c [Quantization] - Added uses_meta_device_weights to quant config (#34645)
Signed-off-by: Josephasafg <ajgard7@gmail.com>
2026-02-17 23:43:44 -08:00
Marek Michalowski e89a91d927 [Bugfix] fix activation in cpu_fused_moe_torch call (#34696)
Signed-off-by: Marek Michalowski <marek.michalowski@arm.com>
2026-02-17 23:39:46 -08:00
ElizaWszolaandLuka Govedič a88b3be7c4 [Bugfix] Fix quant RMS norm fusion for quantization with TMA-aligned scales (#33255)
Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
2026-02-17 23:35:04 -08:00
Tim DettmersandClaude Opus 4.6 d44a5b6c47 Remove dead bitsandbytes CxB code from 8-bit inference path (#34633)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-17 01:49:14 -08:00
Jiangyun Zhu 1d65283e95 Revert "[Models] Fuse Qwen3.5 GDN's qkvz_proj and ba_proj" (#34683) 2026-02-17 01:29:27 -08:00
haosdentandMichael Goin b68fd899d1 [Bugfix] Fix fused MoE int32 overflow in stride*offset without perf regression (#34507)
Signed-off-by: haosdent <haosdent@gmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
2026-02-16 17:58:49 -08:00
roikoren755 3b30e61507 [NemotronH] Do not force router to run in fp32 (#34582)
Signed-off-by: Roi Koren <roik@nvidia.com>
2026-02-16 10:15:32 -08:00
danisereb ec7d9e6745 Fix call to moe_mk in modelopt MoE modules (required for LoRA) (#34575)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
2026-02-16 07:33:09 -08:00
Isotr0py 3bb4e4311c [Models] Fuse Qwen3.5 GDN's qkvz_proj and ba_proj (#34492)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
2026-02-16 07:32:51 -08:00
Samu TamminenandWentao Ye a5ccc85c8c [Bugfix] Fix Dynamo unexpected keyword argument (#34320)
Signed-off-by: Samu Tamminen <stammine@amd.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
2026-02-16 01:32:30 -08:00
bnellnm 5bff999d12 [Bugfix] Add method to swap quant_method on FusedMoE to fix LoRA issues (#34453)
Signed-off-by: Bill Nell <bnell@redhat.com>
2026-02-15 20:10:50 -08:00
Isotr0py 71cd89264f [MM Encoder] Add Triton ViT attention backend (#32183)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
2026-02-15 06:32:47 -08:00
haosdent 79f3fab05a [Bugfix] Handle num_expert_group=None in flashinfer block-scale FP8 MoE (#34494)
Signed-off-by: haosdent <haosdent@gmail.com>
2026-02-14 23:25:46 -08:00
Vadim Gimpelson 604b9eaec5 [BUGFIX] Fix accuracy regression for NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 with TP>1 (#34476)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
2026-02-14 23:25:17 -08:00
Michael Goin bfaa559305 Revert "[Bugfix] Fix fused MoE IMA (sans chunking) by using int64 for strides" (#34530) 2026-02-13 10:35:29 -08:00
Matthias GehreandTJian 934acddef9 [Perf] fused_moe: add int4_w4a16 benchmark support and tuning config (#34130)
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
2026-02-13 00:14:27 -08:00
Roger Wang eea3024f43 [Bugfix] Fix mamba state dtype setting for Qwen3-Next and Qwen3.5 (#34489)
Signed-off-by: Roger Wang <hey@rogerw.io>
2026-02-12 22:48:42 -08:00
bnellnm 04ea31baab [Bugfix] Remove assert that's no longer valid (#34443)
Signed-off-by: Bill Nell <bnell@redhat.com>
2026-02-12 18:18:15 -08:00
amitz-nv f120bd42d3 [Kernel] Support Flashinfer trtllm fused MoE non gated FP8 & NVFP4 (#33506)
Signed-off-by: amitz-nv <203509407+amitz-nv@users.noreply.github.com>
2026-02-12 13:06:58 -08:00
Michael Goin 6d4e27ce29 [Bugfix] Enforce DeepGEMM when using sparse_attn_indexer on CUDA (#34374)
Signed-off-by: mgoin <mgoin64@gmail.com>
2026-02-12 12:08:06 -08:00
xuebwang-amdandTJian 766e167821 [ROCm][quantization] improve OCP weight quant parser robust (#34431)
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
2026-02-12 09:40:19 -08:00
f2c47886fd [Attention] Add FlashInfer Sparse MLA backend (#33451)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
2026-02-12 17:21:54 +00:00