All pull requests

Pull requests

• feat: support speculative decoding in task pipeline.

  #2391·oleanderrr opened Sep 28, 2026Contributor·

  ·

• feat: align anthropic messages api with vllm.

  #2389·XuZhang99 opened Sep 27, 2026Collaborator·

  ·

• feat: align existing OpenAI APIs with vLLM.

  #2388·XuZhang99 opened Sep 27, 2026Collaborator·

  ·

• refactor: group pending overlap step state.

  #2378·liutongxuan opened Sep 26, 2026Collaborator·

  ·

• perf: route GLM Lightning Indexer through torch_npu.

  #2355·PixelFlat opened Sep 20, 2026Contributor·

  ·

• refactor: unify qwen image sequence parallel handling.

  #2322·moonseeker1 opened Sep 9, 2026·

  ·

• feat: add pybind multimodal package to setup.py.

  #2296·yingxudeng opened Sep 1, 2026Collaborator·

  ·

• feat: restore linear-state slots in base step_for_schedule_overlap.

  #2295·yingxudeng opened Sep 1, 2026Collaborator·

  ·

• feat: add x_attention_v2 operator use pypto as compile backend.

  #2286·Yangxu-China opened Aug 29, 2026·

  ·
  1

• feat: support EPLB for DeepSeek-V3.2 Python model executor.

  #2283·yingxudeng opened Aug 24, 2026Collaborator·

  ·

• feat: add lag confidence to decouple adaptive speculative pruning from critical path.

  #2280·weizhehuang0827 opened Aug 20, 2026Collaborator·

  ·

• feat: add Qwen3.8 DFlash2 decoding support.

  #2279·pjgao opened Aug 20, 2026Collaborator·

  ·
  1

• feat: support Qwen3-ASR on NPU device.

  #2278·shan-chen-feng opened Aug 20, 2026Collaborator·

  ·

• perf: token-level multimodal mm token expansion.

  #2268·kongweiqian opened Aug 19, 2026·

  ·

• feat: support Qwen3.5 eager-mode on NPU with Python executor.

  #2264·Wang-1F opened Aug 18, 2026Collaborator·

  ·
  3

• perf: keep raw dit image tensors on cpu until vae preprocess.

  #2261·chuanhaoyu opened Aug 18, 2026·

  ·

• feat: support separate DiT/VAE service instances for Flux.

  #2256·z-jun03 opened Aug 17, 2026Collaborator·

  ·

• feat: add RegionE support for Qwen-Image-Edit.

  #2250·BurnWan opened Aug 17, 2026Collaborator·

  ·

• perf: route low-token qwen3 decode through optimized nz quantmatmul.

  #2248·maojunx99 opened Aug 17, 2026Collaborator·

  ·

• feat: support client-supplied uuid for multimodal processor cache.

  #2243·wly-115 opened Aug 14, 2026Collaborator·

  ·

• bugfix: fix tilelang g_tilingKey multiple definition across kernel families.

  #2242·fengz72 opened Aug 14, 2026Contributor·

  ·

• feat: add DeepSeek V4 NPU kernel Python bindings.

  #2241·0moyi0-2024 opened Aug 14, 2026·

  ·
  23

• feat: add opt-in FIA decode for Qwen3.5.

  #2240·Enguikong opened Aug 14, 2026Contributor·

  ·
  13

• feat: support DeepSeek-V4 in the Python model executor

  #2238·0moyi0-2024 opened Aug 14, 2026·

  ·

• refactor: adopt SequenceParallelMixin for Wan2.2 sequence parallelism and communication-coputation overlap.

  #2237·zhangjieshao98 opened Aug 14, 2026Contributor·

  ·