All pull requests

Pull requests

• fix: repair decode CUDA graph capture and replay

  #105·dandan20031031 opened Oct 22, 2026·

  ·

• fix: wrong shape of the output mask

  #99·HappyMankind opened Sep 16, 2026·

  ·

• fix(block_manager): never share live blocks — cache hits cannot be consumed and silently corrupt prefill output

  #98·MiracleLex opened Aug 31, 2026Contributor·

  ·
  1

• perf: add split-KV flash decoding for batch-1 attention

  #93·irregu opened Aug 5, 2026·

  ·

• Fixed weight loading error in multi-gpu environment as stated in issue #38

  #92·Leamonz opened Jul 23, 2026·

  ·

• Fix Prefix Cache, Scheduler Correctness, Qwen3 Tensor Parallel Loading, and Multi-GPU Execution

  #88·KevinWu8192 opened Jul 19, 2026·

  ·

• [fix] Fix the incorrect rms_norm_epsilon default for Qwen3

  #82·Fan-Yunfan opened Jun 25, 2026·

  ·

• fix(linear): separate bias loader and fix row-parallel bias handling

  #81·HaoYuan-Gao opened Jun 16, 2026·

  ·
  1

• optimize FlashAttention kernel

  #74·LeeWant opened Apr 2, 2026Collaborator·

  ·
  2

• Fix potential issues when the 'enforce_ceager' mode is set to 'False' #72

  #73·LeeWant opened Mar 31, 2026Collaborator·

  ·
  1