4 people 4bac767695 [Perf][Nemotron] Skip redundant latent-MoE all-reduce at TP>1 (~13% decode win) (#52301) ...

Signed-off-by: Junhao Shen <junshen@nvidia.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

2026-09-15 07:33:51 -07:00

..

attention

[Bugfix][Models] Fix OpenPangu sleep mode with static sinks (#53696)

2026-09-14 21:05:47 +08:00

fused_moe

[Perf][Nemotron] Skip redundant latent-MoE all-reduce at TP>1 (~13% decode win) (#52301)

2026-09-15 07:33:51 -07:00

fusion

[Perf][Kernel][Quantization] Fuse ReLU2 with static FP8 activation quantization (#53793)

2026-09-14 10:26:15 -07:00

hpc

[Bugfix] Fix stale HPC QK-norm weights after weight refit (#56706)

2026-09-14 21:34:41 -07:00

mamba

[6/N][warmup][DSv4] Migrate sampling, and DFlash JIT kernels (#56323)

2026-09-12 17:24:24 -07:00

minimax_rms_norm

[ROCm][MiniMax-M2] Dispatch fused QK-norm + AllReduce via AITER (#44849)

2026-07-14 03:10:16 +00:00

pooler

[Bugfix] Honor STEP token pooling in DispatchPooler.for_seq_cls (#55307)

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

quantization

[Refactor] Normalize DeepSeek V4.1 model package naming (#56741)

2026-09-13 23:17:50 -07:00

rotary_embedding

[Bugfix] Align vLLM YaRN with Transformers and stop re-scaling max_model_len (#56446)

2026-09-11 13:01:38 +00:00

__init__.py

…

activation.py

[Performance] Add Triton kernel for Gemma3n sparse GELU (#48498)

2026-09-14 07:42:15 +00:00

attention_layer_base.py

[3/N][KV-Cache Layout Refactor] Standardize Mamba cache; drop get_transfer_cache_regions (#44456)

2026-07-21 09:16:15 +00:00

conv.py

[Refactor] Remove multiple dead codes (#50285)

2026-08-03 10:59:08 -04:00

fused_allreduce_gemma_rms_norm.py

[ROCm][Perf][M3] Fused allreduce+GemmaRMSNorm fast path (#54787)

2026-09-08 07:21:49 +00:00

fused_embed_norm.py

Replicated embedding and norm fusion for DSV3 flat model (#48484)

2026-08-18 14:53:21 +08:00

fused_qk_norm_rope.py

[Kernel][Perf] Enable fused QK-norm + partial MRoPE + gate for Qwen3.6 (#52676)

2026-08-24 18:28:49 -07:00

indexer_topk.py

[Perf][Kernel] Integrate DeepSelect TopK for the DSA sparse indexer (#56464)

2026-09-13 00:44:37 -07:00

layernorm.py

[XPU] Add fused GemmaRMSNorm path for eager execution (#53678)

2026-09-03 06:11:20 +08:00

lightning_attn.py

[Kernel][Test] Extend lightning_attn and awq_triton kernel tests to XPU (#43307)

2026-06-04 14:25:59 -04:00

linear.py

[Perf][Kernel][Quantization] Fuse ReLU2 with static FP8 activation quantization (#53793)

2026-09-14 10:26:15 -07:00

logits_processor.py

[Hy4] support Hy4-preview model (#54160)

2026-08-29 12:07:28 +08:00

mhc.py

[ROCm][DSV4.1][Perf] Fold the mHC post step into the delayed pre projection (#56513)

2026-09-14 16:44:56 -05:00

mla.py

[3/N] HiSparse: host-resident sparse-MLA decode hot-buffering (#53781)

2026-09-12 00:36:47 -04:00

resampler.py

…

sparse_attn_indexer_kpool.py

[PCP][DCP] Enable PCP+DCP on sparse-MLA models (#56157)

2026-09-12 23:24:44 +00:00

sparse_attn_indexer.py

[Perf][Kernel] Integrate DeepSelect TopK for the DSA sparse indexer (#56464)

2026-09-13 00:44:37 -07:00

utils.py

[ROCm][Bugfix][Perf] Tune multi-stream shared experts use; wvSplitKrc fixes (#56098)

2026-09-10 15:14:22 -07:00

vocab_parallel_embedding.py

[Qwen4Exp] Support UVA PLE-offload and Engram tensor parallelism (#54371)

2026-09-09 07:32:34 -07:00