All pull requests

Pull requests

• feat(fuse_moe): add small-batch FP8 warp_decode kernels for SM90

  #102·Religious-J opened Oct 30, 2026Contributor·

  ·

• feat: add SM120 fused MoE (MXFP4 x MXFP8, 128-block FP8)

  #101·rynwang opened Oct 22, 2026·

  ·
  3

• feat: add SM120 CuTe DSL fused MoE (128-block FP8 and MXFP4 x MXFP8)

  #100·rynwang opened Oct 12, 2026·

  ·

• [Fused MoE] Extend W13 with peer-indexed EP input

  #89·taoyuanyuan opened Aug 25, 2026·

  ·

• chore: unify remaining copyright headers to Tencent 2026

  #88·xueyangcs opened Aug 25, 2026Contributor·

  ·

• fix: move pipeline barriers into dynamic SharedStorage for 128-byte alignment (#42)

  #87·XFDG opened Aug 24, 2026·

  ·
  2

• # docs: add layout algebra comment to make_tiled_copy_V_interleave_trans (#44)

  #86·XFDG opened Aug 24, 2026·

  ·

• fix: fix kTaskLoopPolicy type in grouped gemm host code

  #85·icyhearts opened Aug 23, 2026·

  ·
  2

• fix(fuse_moe): correct dowm_tmas typo

  #83·cherishhh opened Aug 19, 2026·

  ·

• Improve source build robustness

  #78·liuyao0322 opened Jul 31, 2026·

  ·

• Fix AllReduce benchmark README script name

  #77·liuyao0322 opened Jul 31, 2026·

  ·

• [ROCm] Add ROCm support and norm op.

  #69·WhatGhost opened Jul 20, 2026·

  ·
  1

• feat: add qwen3 tts rope operator

  #67·l-wave opened Jul 17, 2026·

  ·

• fix: change kHalfFltEps from 1e-20 to 5e-8 in sampler rng, aligned to torch

  #65·LINZHI97 opened Jul 13, 2026·

  ·

• feat: add qwen3 tts topk gumbel sampler

  #64·l-wave opened Jul 10, 2026·

  ·

• feat: relax heads_per_group constraint to support 1~8 for BF16 and FP8 decode attention

  #58·Religious-J opened Jun 25, 2026Contributor·

  ·

• feat: bf16 fused moe kernel

  #41·ZelinMa557 opened Apr 9, 2026·

  ·
  5

• feat: add bf16 group gemm kernel

  #34·ZelinMa557 opened Mar 23, 2026·

  ·