Bingxu Chen ab936ce694 Revert "perf: optimize PCG inductor path for FP8 models (#21734)" (#23159) ...

Feel free to PR again.

2026-04-19 23:32:50 -07:00

..

deepseek_common

GLM-5/5.1 MXFP4 Checkpoint Inference Compatibility Fix (#22543)

2026-04-13 23:56:48 -07:00

afmoe.py

refactor(moe): de-duplicate triton MoE runner path into shared helpers (#23019)

2026-04-17 17:05:13 -07:00

apertus.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

arcee.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

baichuan.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

bailing_moe_linear.py

[CPU] Add FP8 Bmm support (#9744)

2026-03-18 22:19:48 -07:00

bailing_moe_nextn.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

bailing_moe.py

fix fused_set_kv_buffer for rope with Ling-v2 (#20316)

2026-03-23 19:20:40 +08:00

bert.py

[NPU] bugfix with Kimi-k2 and bge-reranker-v2 model (#17478)

2026-01-22 22:02:05 +08:00

chatglm.py

…

clip.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

commandr.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

dbrx.py

refactor(moe): de-duplicate triton MoE runner path into shared helpers (#23019)

2026-04-17 17:05:13 -07:00

deepseek_janus_pro.py

[AMD] Fix Janus-Pro crash and add Kimi-K2.5 nightly test (#18269)

2026-02-10 22:33:13 -08:00

deepseek_nextn.py

[Refactor] Deduplicate NSA utils.py into cp_utils.py for context parallel (#22914)

2026-04-19 21:35:35 -07:00

deepseek_ocr.py

[CPU] Fix MoE layer support for DeepSeek-OCR models (#12555)

2026-03-18 22:57:55 -07:00

deepseek_v2.py

[Refactor] Deduplicate NSA utils.py into cp_utils.py for context parallel (#22914)

2026-04-19 21:35:35 -07:00

deepseek_vl2.py

[VLM] Chunk-aware ViT encoding with per-image cache and lazy device transfer (#22038)

2026-04-04 16:55:17 +08:00

deepseek.py

refactor(moe): de-duplicate triton MoE runner path into shared helpers (#23019)

2026-04-17 17:05:13 -07:00

dflash.py

[AMD] Enable DFLASH speculative decoding on ROCm (#22342)

2026-04-17 13:10:14 -07:00

dots_ocr.py

…

dots_vlm_vit.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

dots_vlm.py

feat: support EPD disaggregation (#12263)

2025-12-14 22:30:08 +08:00

ernie4_eagle.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

ernie4.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

ernie45_moe_vl.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

ernie45_vl.py

[VLM] Support cos sin cache for Ernie4.5-VL (#19743)

2026-03-04 10:54:23 +08:00

exaone4.py

[Model] Add support for EXAONE-4.0 Model (#8205)

2026-01-27 14:08:24 +08:00

exaone_moe_mtp.py

[Model] Add K-EXAONE model support (#16294)

2026-01-30 20:01:14 +08:00

exaone_moe.py

[Model] Add K-EXAONE model support (#16294)

2026-01-30 20:01:14 +08:00

exaone.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

falcon_h1.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

gemma2_reward.py

[Score API] Add return_pooled_hidden_states to Scoring API for SequenceClassification / RewardModel (#22427)

2026-04-15 14:58:56 -07:00

gemma2.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

gemma3_causal.py

[New Model] Gemma 4 (#21952)

2026-04-06 20:24:44 -07:00

gemma3_mm.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

gemma3n_audio.py

…

gemma3n_causal.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

gemma3n_mm.py

Fix nightly VLM accuracy: gemma3n TP fixes + removal, latency thresholds (#19401)

2026-02-27 14:24:02 -08:00

gemma4_audio.py

[New Model] Gemma 4 (#21952)

2026-04-06 20:24:44 -07:00

gemma4_causal.py

[New Model] Gemma 4 (#21952)

2026-04-06 20:24:44 -07:00

gemma4_mm.py

[New Model] Gemma 4 (#21952)

2026-04-06 20:24:44 -07:00

gemma4_vision.py

[New Model] Gemma 4 (#21952)

2026-04-06 20:24:44 -07:00

gemma.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

glm4_moe_lite.py

fix[glm4.7 flash]: properly detect gfx95_quant_format (#22720)

2026-04-13 13:10:07 -07:00

glm4_moe_nextn.py

[Bugfix] Preserve auto-detected quant_config for GLM NextN draft model (#22823)

2026-04-15 13:25:36 -07:00

glm4_moe.py

[AMD] Fuse RMSNorm + FP8 per-token quant for GLM-4.7-FP8 (#21403)

2026-04-10 22:45:31 -07:00

glm4.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

glm4v_moe.py

[Bug Fix] GLM-V / GLM-OCR: field detection for transformers 5.x and MTP omission fix (#21134)

2026-03-23 13:19:48 -07:00

glm4v.py

[bugfix]GLM-4V model (#17122)

2026-04-01 10:37:40 +08:00

glm_ocr_nextn.py

[GLM-OCR] Support GLM-OCR Model (#17582)

2026-01-26 22:24:00 -08:00

glm_ocr.py

[Bug Fix] GLM-V / GLM-OCR: field detection for transformers 5.x and MTP omission fix (#21134)

2026-03-23 13:19:48 -07:00

glmasr.py

fix: change class name of GLM-ASR (#15772)

2025-12-26 00:15:16 +08:00

gpt2.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

gpt_bigcode.py

…

gpt_j.py

[Fix] Remove unused Type import in gpt_j.py (#17975)

2026-01-29 21:11:11 -08:00

gpt_oss.py

[CI] Add per-job uv venv isolation and upgrade CI version to Cuda 13 (#23119)

2026-04-19 05:32:36 -07:00

granite.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

granitemoe.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

granitemoehybrid.py

Model: Support IBM Granite (Dense/Mamba + MoE) (#18040)

2026-02-15 11:24:41 +08:00

grok.py

fix(grok): adapt huihui-ai/grok-2 (#21522)

2026-04-07 10:04:41 +08:00

hunyuan.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

idefics2.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

internlm2_reward.py

[Score API] Add return_pooled_hidden_states to Scoring API for SequenceClassification / RewardModel (#22427)

2026-04-15 14:58:56 -07:00

internlm2.py

…

interns1.py

[Bug] Fix Intern-S1 model accuracy and support /generate interface with input_ids (#12367)

2025-11-02 20:22:33 -08:00

interns1pro.py

model: support interns1-pro (#18145)

2026-02-05 00:22:44 +08:00

internvl.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

iquest_loopcoder.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

jet_nemotron.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

jet_vlm.py

model: support JetVLM (#13289)

2025-11-18 12:02:03 +08:00

kimi_k25.py

[VLM] Enable per-image ViT cache and avoid TP CUDA context creation for Kimi-K2.5 (#22858)

2026-04-16 01:14:24 +08:00

kimi_linear.py

perf(kimi_linear): replace einops rearrange with native torch ops in Kimi-Linear KDA path (#20396)

2026-03-20 10:38:12 +08:00

kimi_vl_moonvit.py

Add SGLang CUDA crash API logging inspired by FlashInfer (#20910)

2026-03-22 16:39:40 +08:00

kimi_vl.py

[EPD][VLM] Support Kimi VL EPD (#22490)

2026-04-16 12:40:02 +08:00

lfm2_moe.py

Add LMF2 MoE model architecture (#17997)

2026-02-12 01:03:43 +08:00

lfm2_vl.py

[VLM] fix LFM2-VL offline inference and GPU JPEG decode (#22448)

2026-04-15 09:13:25 +08:00

lfm2.py

model: support LFM2-VL (Liquid Foundation Model 2 Vision-Language) (#21230)

2026-04-04 16:36:04 +08:00

lightonocr.py

Support LightOnOCR-2-1B (#17806)

2026-01-29 23:03:41 +08:00

llada2.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

llama4.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

llama_classification.py

[Score API] Add SequenceClassification Model support (#22118)

2026-04-08 01:30:58 -07:00

llama_eagle3.py

Bug fix for llama eagle3 (#21397)

2026-04-01 15:01:53 +08:00

llama_eagle.py

…

llama_embedding.py

…

llama_reward.py

[Score API] Add return_pooled_hidden_states to Scoring API for SequenceClassification / RewardModel (#22427)

2026-04-15 14:58:56 -07:00

llama.py

[sgl] fix using symmetric memory issues for attention_tp (#22286)

2026-04-11 00:26:18 +08:00

llava.py

[VLM] Enable per-image MM splitting by default and remove MULTI_IMAGES modality (#21899)

2026-04-03 11:04:41 +08:00

llavavid.py

…

longcat_flash_nextn.py

[CPU] Add FP8 Bmm support (#9744)

2026-03-18 22:19:48 -07:00

longcat_flash.py

[CPU] Add FP8 Bmm support (#9744)

2026-03-18 22:19:48 -07:00

midashenglm.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mimo_mtp.py

…

mimo_v2_flash_nextn.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mimo_v2_flash.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mimo.py

…

mindspore.py

[NPU]MindSpore backend support eagle3 (#17098)

2026-03-11 09:11:19 +03:00

minicpm3.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

minicpm.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

minicpmo.py

[vlm][minicpm] support input formats of processor output and embedding (#19614)

2026-03-04 12:11:12 -05:00

minicpmv.py

[VLM] Enable per-image MM splitting by default and remove MULTI_IMAGES modality (#21899)

2026-04-03 11:04:41 +08:00

minimax_m2.py

[AMD] Add MiniMax-M2.7 accuracy and performance nightly tests (#22722)

2026-04-14 00:30:11 -07:00

ministral3.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mistral_large_3_eagle.py

Add Mistral Small 4 (Pixtral) support (#20708)

2026-03-18 14:15:32 -07:00

mistral_large_3.py

Mistral Large 3 NVFP4 support (#14485)

2025-12-13 13:34:42 +08:00

mistral.py

…

mixtral_quant.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mixtral.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mllama4.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

mllama.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

nano_nemotron_vl.py

EVS Framework: Support NemotronH_Nano_VL_V2 (#14051)

2026-01-05 16:18:07 +08:00

nemotron_h_mtp.py

Fix/nemotron mtp quantaized (#19433)

2026-03-03 01:07:46 -08:00

nemotron_h.py

[Model] Fix NemotronH OOM on unified-mem systems: stream weights + safetensors cleanup (#20580)

2026-03-17 09:47:58 -07:00

nemotron_nas.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

nvila_lite.py

…

nvila.py

…

olmo2.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

olmo.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

olmoe.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

opt.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

orion.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

paddleocr_vl.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

persimmon.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

phi3_small.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

phi4mm_audio.py

…

phi4mm_utils.py

…

phi4mm.py

fix: add missing f-string prefixes in warning and assert messages (#22067)

2026-04-04 17:20:16 -07:00

phi.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

phimoe.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

pixtral.py

Add Mistral Small 4 (Pixtral) support (#20708)

2026-03-18 14:15:32 -07:00

points_v15_chat.py

…

qwen2_5_vl.py

[NPU] Convert cu_window_seqlens to CPU for npu_flush_attention_unpad operator (#20328)

2026-03-12 09:08:43 +08:00

qwen2_audio.py

update pre-commit config (#18860)

2026-02-16 00:18:31 +08:00

qwen2_classification.py

[Score API] Add SequenceClassification Model support (#22118)

2026-04-08 01:30:58 -07:00

qwen2_eagle.py

…

qwen2_moe.py

refactor(moe): de-duplicate triton MoE runner path into shared helpers (#23019)

2026-04-17 17:05:13 -07:00

qwen2_rm.py

[Score API] Add return_pooled_hidden_states to Scoring API for SequenceClassification / RewardModel (#22427)

2026-04-15 14:58:56 -07:00

qwen2_vl.py

VLM: add Conv2dLayer/Conv3dLayer to fix PyTorch 2.9.1 CuDNN Conv3d (#20282)

2026-03-15 19:17:44 +08:00

qwen2.py

fix qwen2_5_math_rm_72b (#21295)

2026-04-07 14:36:57 +08:00

qwen3_5_mtp.py

Fix qwen3.5 mtp eplb related issues (#19767)

2026-03-09 16:05:32 +08:00

qwen3_5.py

[AMD] Enable share expert fusion with router experts for Qwen3.5 BF16 & FP8 (#20736)

2026-04-14 18:52:36 -07:00

qwen3_asr.py

model: support qwen3-asr (#22073)

2026-04-07 13:27:05 +08:00

qwen3_classification.py

[Score API] Add SequenceClassification Model support (#22118)

2026-04-08 01:30:58 -07:00

qwen3_moe.py

Enable DFLASH support for additional model backends (#22358)

2026-04-09 14:36:12 -07:00

qwen3_next_mtp.py

Fix qwen3.5 mtp eplb related issues (#19767)

2026-03-09 16:05:32 +08:00

qwen3_next.py

Enable DFLASH support for additional model backends (#22358)

2026-04-09 14:36:12 -07:00

qwen3_omni_moe.py

[Omni] Optimize AudioEncoder for Qwen3_Omni_Thinker (#18185)

2026-03-14 23:00:17 +08:00

qwen3_rm.py

[Model] Add Qwen3ForRewardModel and fix Qwen3ForSequenceClassification (#17992)

2026-02-16 19:44:41 +08:00

qwen3_vl_moe.py

[PP] qwen3 vl skip layer id for pp (#19135)

2026-04-03 10:51:53 +08:00

qwen3_vl.py

Upgrade transformers to 5.5.3 and refactor hf_transformers_utils into subpackage (#21569)

2026-04-15 20:03:44 -07:00

qwen3.py

Restore Qwen3 rope config fallback (#22739)

2026-04-13 21:47:37 -07:00

qwen.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

radio.py

Support nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 (and nvidia/C-RADIOv2-H) (#12277)

2025-11-26 16:28:52 -07:00

registry.py

tiny unify environ usage (#15335)

2025-12-17 23:31:43 +08:00

roberta.py

…

sarashina2_vision.py

…

sarvam_moe.py

[Sarvam] Add inference support for Sarvam MoE LLMs (#18938)

2026-03-04 15:28:00 -08:00

sdar_moe.py

Feature/sdar support (#19044)

2026-02-19 21:58:15 -08:00

sdar.py

Feature/sdar support (#19044)

2026-02-19 21:58:15 -08:00

siglip2.py

model: support LFM2-VL (Liquid Foundation Model 2 Vision-Language) (#21230)

2026-04-04 16:36:04 +08:00

siglip.py

[CPU] Add frontend support for Gemma (#12590)

2026-03-18 23:02:26 -07:00

solar.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

stablelm.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

starcoder2.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

step3_vl_10b.py

[VLM] Chunk-aware ViT encoding with per-image cache and lazy device transfer (#22038)

2026-04-04 16:55:17 +08:00

step3_vl.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

step3p5_mtp.py

[misc] refine outdated comments for chain-style multi-layer MTP (#22996)

2026-04-16 14:49:43 -07:00

step3p5.py

[Step3p5] Optimize allreduce in MoE layers (#22773)

2026-04-16 09:33:12 +08:00

teleflm.py

model: support teleflm (#10573)

2025-11-15 03:14:49 +08:00

torch_native_llama.py

Upgrade transformers==5.3.0 (#17784)

2026-03-18 13:50:43 -07:00

transformers.py

[Feature] Stronger transformers modeling backend with TP, PP, MoE, VLMs, and torch compile (#19163)

2026-04-02 16:02:33 -07:00

utils.py

Revert "perf: optimize PCG inductor path for FP8 models (#21734)" (#23159)

2026-04-19 23:32:50 -07:00

voxtral.py

[model] support voxtral (speech-to-text) (#21635)

2026-04-05 15:46:30 +08:00

whisper.py

[Whisper] Batch encoder forward for concurrent prefill requests (#22361)

2026-04-12 14:15:14 +08:00

xverse_moe.py

refactor(moe): de-duplicate triton MoE runner path into shared helpers (#23019)

2026-04-17 17:05:13 -07:00

xverse.py

fix: use get_rope_config() to support models without rope_parameters (#21135)

2026-03-26 11:22:12 -07:00

yivl.py

…