Faster per-token fp8 group quant packed kernel for blackwell (#41326)

Yongye Zhu committed 6 months ago

[DeepSeek] Use torch.mm for bf16xbf16->fp32 gemm (#41300)

Woosuk Kwon committed 6 months ago

[Bugfix] Fix RoutedExpertsCapturer for Gemma 4 MoE (top_k_experts) (#41401)

Tran Le committed 6 months ago

[DSV4] Add BF16 and MXFP8 A2A support for flashinfer a2a one sided (#40960)

Yongye Zhu committed 6 months ago

[Model Runner v2] Fix v2 compile counter `num_gpu_runner_capture_triggers` and `num_cudagraph_captured` (#41285)

Wentao Ye committed 6 months ago

Fix typo in log message for indexer cache (#41419)

Michael Goin committed 6 months ago

[NVFP4][Hopper/AMD Instinct] Add Triton kernels for NVFP4 dequantization and QDQ emulation (#40033)

fxmarty-amd committed 6 months ago

[CI/Build] Skip Prithvi/Terratorch model-registry tests when terratorch is missing (#41389)

Stefano Castagnetta committed 6 months ago

[CI] Add MTP coverage: Qwen3.5 correctness + no-sync spec decode (#40472)

Stefano Castagnetta committed 6 months ago

[DSV4] Avoid redundant dtype conversion. (#41374)

Jee Jee Li committed 6 months ago

xpu docker: pin oneAPI to 2025.3 and avoid unintended 2026 upgrade (#41380)

wenjun liu committed 6 months ago

[CI/Build] Skip terratorch + torchgeo while PyPI has lightning quarantined (#41377)

Stefano Castagnetta committed 6 months ago

[ROCm] ROCm DeepEP API updated to latest (#39721)

tej committed 6 months ago

[UX][Bugfix] Fix OOM by setting PyTorch `max_split_size_mb` during model loading (#41268)

Matthew Bonanni committed 6 months ago

[EPLB] Optimize memory overhead in Nixl communicator (#40013)

Ilya Markov committed 6 months ago

Stop mergify labelling from skipping pre-commit (#41362)

Harry Mellor committed 6 months ago

[Doc] Fix RTD build: pytorch.org/docs/stable/objects.inv returns 404 (#41353)

Stefano Castagnetta committed 6 months ago

[MoE] Make MoERunnerInterface a PluggableLayer for OOT support (#35178)

Xiaoshuang Wang committed 6 months ago

[P/D] Prefill compute optimizations with bi-directional KV cache transfers between P and D nodes (#32553)

snadampal committed 6 months ago

[KVConnector] MultiConnector SupportsHMA (#39571)

Nicolò Lucchesi committed 6 months ago