All pull requests

Pull requests

• fix: validate SM100 masked strides and add scheduler coverage

  #492·anencore94 opened Nov 2, 2026·

  ·

• Fix 32-bit scale indices in transpose_fp32 and pack_fp32_into_ue8m0

  #491·Theodore-Liu opened Nov 1, 2026·

  ·
  1

• docs: document DG_USE_LOCAL_VERSION in the env var list

  #489·stwilliam503388-creator opened Oct 30, 2026·

  ·

• fix: restore the env var when an @ignore_env test raises

  #488·stwilliam503388-creator opened Oct 30, 2026·

  ·

• [Fix] relax SM90 num_heads assert to allow 16

  #486·PaddyXj opened Oct 29, 2026·

  ·

• [SM90] Reduce H32 MQA logits heads within each thread

  #485·unrestrainedpointer opened Oct 27, 2026·

  ·

• fix(utils): enhance tensor view contiguity, squeeze safety, and scale indexing

  #483·reginaldalfret opened Oct 25, 2026·

  ·

• Merge math.hpp, replace Kpack4 with torch::kFloat4_e2m1fn_x2

  #482·john-kernel opened Oct 24, 2026·

  ·
  2

• Use the allocated TMEM base in the SM100 HC prenorm GEMM

  #480·XFDG opened Oct 21, 2026·

  ·
  1

• [Fix][Einsum] Reject c in SM90 fp8_einsum("bhr,hdr->bhd") to prevent silent data corruption in sm90_fp8_bmm

  #477·arxgy opened Oct 17, 2026·

  ·
  1

• Fix SM120 split-K alignment after partition caps

  #475·Sunt-ing opened Oct 15, 2026·

  ·
  1

• Add optimized fused FP8 x MXFP4 MegaMoE for SM90

  #473·AIExplorer7 opened Oct 14, 2026·

  ·

• docs: add Simplified Chinese translation (README.zh.md)

  #472·JasonYeYuhe opened Oct 14, 2026·

  ·

• Make pack_ue8m0_to_int CUDA graph capture safe

  #466·zhoubiao188 opened Oct 6, 2026·

  ·
  1

  1

• Update copyright year in LICENSE file

  #465·bishalramdam opened Oct 6, 2026·

  ·

• Update vendored CUTLASS to v4.8.0

  #464·fhgffy opened Oct 4, 2026·

  ·
  2

• Add SM90 BF16 grouped GEMM with gathered activation rows

  #463·yuweih205 opened Sep 30, 2026·

  ·

• [Fix] Bound cuBLASLt descriptor lifetime on errors

  #460·ZenAlexa opened Sep 28, 2026·

  ·

• [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·

  ·
  7

• 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