All pull requests

Pull requests

• bf16 wgrad: dual-source varlen-K GEMM and merge of deferred wgrads

  #95·adam-xiaoyao opened Oct 25, 2026·

  ·

• Use paddle cpp extension loader for cold JIT

  #69·Manfredss opened Jul 16, 2026·

  ·

• Optimize DeepEP metadata-to-GEMM launch bridge

  #68·A-nnonymous opened Jul 13, 2026Collaborator·

  ·

• Tune BF16 wgrad and fuse DeepEP metadata scale packing

  #64·A-nnonymous opened Jul 9, 2026Collaborator·

  ·

• Add bit-exact tests for SonicMoE weight layout kernels

  #62·A-nnonymous opened Jul 7, 2026Collaborator·

  ·

• feat(gemm): unified gemm_num_sms SM-count cap for all frontier GEMMs

  #58·A-nnonymous opened Jul 2, 2026Collaborator·

  ·

• feat(overlap): global gemm_num_sms SM cap for frontier GEMMs

  #57·A-nnonymous opened Jul 2, 2026Collaborator·

  ·

• Fix SonicMoE JIT lock warm-cache path

  #47·A-nnonymous opened Jun 16, 2026Collaborator·

  ·

• [Perf] Improve FP8 dispatch storage release

  #43·A-nnonymous opened Jun 10, 2026Collaborator·

  ·
  1