All pull requests

Pull requests

• [Fix] Honor deterministic algorithms in SM90 bmk,bnk->mn einsum

  #457·LiRunGuo opened Sep 22, 2026·

  ·
  4

• [Fix] Reject FP32 bmk,bnk->mn einsum without C instead of crashing

  #456·LiRunGuo opened Sep 22, 2026·

  ·
  2

• [Fix] SM90 heuristic offers block M 16/32 to the FP8 1D1D kernel (JIT failure for M <= 32)

  #455·LiRunGuo opened Sep 22, 2026·

  ·
  6

• Add SM90 NVFP4+FP8 MegaMoE support

  #454·lengrongfu opened Sep 21, 2026·

  ·
  24

• Fix SM120 BF16 GEMM shared-memory reuse ordering

  #453·Sunt-ing opened Sep 21, 2026·

  ·
  16

• perf: Use full-rate FADD relu in SM90 FP8 MQA logits epilogue

  #452·Chtholly-Boss opened Sep 18, 2026·

  ·
  4

• [Bugfix] Reset MegaMoE combine readiness for CUDA graph replay

  #451·foreverrookie opened Sep 17, 2026·

  ·
  3

• feat(gemm): opt-in labels-contract checker for m-grouped contiguous GEMM

  #450·lucifer1004 opened Sep 16, 2026·

  ·
  18

• [BugFix][sm90] tf32 gemm race condition

  #448·leevan opened Sep 15, 2026·

  ·
  6

• Integrate main APIs and SM120 kernels while preserving nv_dev features

  #447·lucifer1004 opened Sep 15, 2026·

  ·
  31

• Fix test helper discovery in pytest and sanitizer runner

  #446·lucifer1004 opened Sep 14, 2026·

  ·
  8

• fix(sm90): support parameterized SwiGLU in FP8 MegaMoE

  #445·shiyang814-cpu opened Sep 14, 2026·

  ·
  11

• Add torch dispatcher op for tf32_hc_prenorm_gemm

  #442·leevan opened Sep 14, 2026·

  ·
  10

• Add a fused SM90 FP8 MegaMoE: one kernel, a lag-ordered schedule and a ring activation pool

  #440·zkyue opened Sep 13, 2026·

  ·
  15

• [Bugfix][Layout] Add cuda_fp8.h include to fix __nv_fp8_e4m3 undefined

  #439·XFDG opened Sep 11, 2026·

  ·
  1

  3

• Fix unknown CMake command "cuda_add_library" in CMakeLists.txt

  #437·PerryLink opened Sep 10, 2026·

  ·
  4

• Improve SM100 MX MQA logits clean-output performance

  #436·OutWhite opened Sep 10, 2026·

  ·
  13

• fix: derive the SM120 fp8 mqa_logits swizzle mode from head_dim

  #434·aganhui opened Sep 10, 2026·

  ·
  5

• fix: port the #343 tensormap drain to the SM120 k_grouped NT contiguous kernel

  #431·aganhui opened Sep 9, 2026·

  ·
  1

• [BUG] Fix sm90 grouped multicast deadlock

  #429·baptiste-bt opened Sep 8, 2026·

  ·
  7

• fix: reject non-positive block size multiples

  #428·nisbenz opened Sep 7, 2026·

  ·
  2

• Improve error when cuobjdump is missing or fails

  #427·monperrus opened Sep 6, 2026·

  ·
  4

• [Bug] Fix bench_kineto returning zero on the first call in a process

  #426·DiegoCao opened Sep 2, 2026·

  ·
  4

• docs: note Mega MoE requires PyTorch >= 2.9 in Requirements

  #424·pengfaming opened Aug 28, 2026·

  ·
  2

• docs: fix stale test file link in README

  #423·pengfaming opened Aug 28, 2026·

  ·
  2