All pull requests

Pull requests

• Finish SDPA binder ownership and remove production Python forward fallbacks

  area:frostarea:global_attentioncat-enhancementsorig-nv-eng
  #1438·YangXu1990uiuc opened Oct 7, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  14

• sdpa fwd (cc 10.7): f16x2 exponent and pre-folded scale on every forward flavor; graph.sdpa attn_scale_prefolded; cc 10.7 test_mhas_v2 lever sweeps

  area:frostarea:global_attentioncat-enhancementsmod-cutedslmod-frostorig-nv-eng
  #1436·vedaanta opened Oct 7, 2026Collaborator·

  ·

  · Frontend 1.31.0
  3

• Unify SM80 packed THD forward host binding in native code

  area:frostarea:global_attentioncat-enhancementsorig-nv-eng
  #1434·YangXu1990uiuc opened Oct 7, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  6

• SDPA: use one native binder for half backward entry points

  area:frostarea:global_attentioncat-enhancementsorig-nv-eng
  #1433·YangXu1990uiuc opened Oct 7, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  10

• [Auto-FE] An explicit attn_scale of 0 means 0 in every FROST SDPA adapter; kernels that cannot run it refuse it

  area:frostarea:global_attentioncat-bugfixorig-nv-eng
  #1430·YangXu1990uiuc opened Oct 7, 2026Collaborator·3 of 4 tasks·

  ·

  · Frontend 1.31.0
  12

• frost(gated_attention_block, sdpa_bwd_sm107): the MXFP8 block backward -- GatedAttentionBlockBwd(quant=MxQuantSpec) over the MXFP8 record (block-scale GEMMs over transposed operands, canonical / transposed quantizer modes); the SDPA row's GQA dK fold rounded once from fp32 partials

  area:frostcat-featuremod-cutedslmod-frostorig-nv-eng
  #1429·RomanAnders90 opened Oct 7, 2026Contributor·

  ·

  · Frontend 1.31.0
  5

  

• Accept canonical layouts in grouped dGLU, dgrad quant, and wgrad

  #1422·hwanseoc opened Oct 6, 2026Member·

  ·

• sdpa(frost): re-fit SM100 placement on B200; offer the SM90, SM100 FP8 and SM100 d512-backward rows by default

  area:frostarea:global_attentioncat-enhancementsmod-frostorig-nv-eng
  #1421·YangXu1990uiuc opened Oct 6, 2026Collaborator·4 tasks done·

  ·
  14

• Read ALiBi slopes and other constant tensors from device memory

  area:global_attentioncat-bugfixorig-nv-eng
  #1419·egilliam-nv opened Oct 6, 2026Collaborator·

  ·

  · Frontend 1.31.0
  4

  

• Speed up deterministic DSA top-k boundary ties

  area:sparse_attentioncat-bugfixmod-cutedslop: DSAorig-external
  #1413·baudouinMNC opened Oct 6, 2026·3 of 4 tasks·

  ·

  · Frontend 1.31.0
  5

• grouped gemm: unified JAX GLU API on SM100 and Rubin

  cat-enhancementsmod-cutedslorig-nv-eng
  #1399·hwanseoc opened Oct 5, 2026Member·4 tasks done·

  ·

  · Frontend 1.31.0

• Resolve setmaxregister Problem and Add SwapAB for Moe Gemm

  area:frostcat-bugfixcat-featuremod-cutedslmod-frostorig-nv-eng
  #1354·yihuawei opened Oct 3, 2026Collaborator·4 tasks·

  ·

  · Frontend 1.31.0
  13

  

• [DSA] Add prepared BF16 variable-length Top-K for GB300 and Rubin

  area:sparse_attentioncat-featureop: DSAorig-nv-eng
  #1298·Anerudhan opened Sep 30, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  2

• Fix grouped SwiGLU SM110 reduction, scale output and AMAX stream ordering

  cat-bugfixmod-cutedslorig-external
  #1280·0z5a opened Sep 29, 2026Contributor·

  ·

  · Frontend 1.31.0
  10

• test(gemm): cover public INT8 large-accumulator contracts

  #1265·0z5a opened Sep 28, 2026Contributor·

  ·
  4

• test(gemm): cover block-scale public graph replay and scale edges

  #1259·0z5a opened Sep 28, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix(bsa): stabilize SM100 forward softmax for large logits

  #1221·Butterfingrz opened Sep 24, 2026Contributor·3 of 4 tasks·

  ·
  1

• BSA: optimize SM90 blk64 forward performance

  #1213·Butterfingrz opened Sep 24, 2026Contributor·3 of 4 tasks·

  ·
  11

• Add C++ and Python APIs for custom numerical weight dequantization

  #1210·scottyokim opened Sep 24, 2026·1 of 3 tasks·

  ·
  1

• Rule 8 (gated_attention_block): quant words in a workspace slot, per-device graph handle with locked re-stream

  area:global_attentioncat-cleanupmod-cutedslorig-nv-eng
  #1186·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  95

• Rule 8 (HSTU): caller-owned workspace for block-sparse metadata and the bwd fp32 accumulator, R5 declines, compile-time fakes

  area:global_attentioncat-cleanupmod-cutedslorig-nv-eng
  #1185·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  107

• Rule 8 (flex_attention): compile() from fake tensors, no device memory at compile

  area:global_attentioncat-cleanupmod-cutedslorig-nv-eng
  #1184·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  95

• Rule 8 (NSA/CSA): plan-time host ints, SWA caller-owned workspace and THD ragged offsets; wrappers unchanged

  area:sparse_attentioncat-cleanupmod-cutedslorig-nv-eng
  #1183·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  115

• Rule 8 (DSA): plan-style classes own no device memory -- required outputs, caller-owned workspaces, plan-time host ints; wrappers unchanged

  area:sparse_attentioncat-cleanupmod-cutedslop: DSAorig-nv-eng
  #1182·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  95

• Rule 8 core: caller-owned workspaces and no host blocking in the graph API, FROST MoE GEMM and SDPA, grouped GEMM and Hopper KDA; R9 detector and recipes R10/R11

  area:frostarea:global_attentionarea:linear_attentioncat-cleanupmod-cutedslmod-frostorig-nv-eng
  #1181·YangXu1990uiuc opened Sep 22, 2026Collaborator·4 tasks done·

  ·

  · Frontend 1.31.0
  139