All pull requests

Pull requests

• frost(sdpa): sm107 d512 prefill -- stream the O TMA store per published chunk, pipeline the epilogue readout (bf16 / fp8 / mxfp8), release BMM2 early in the bf16 correction

  area:frostcat-enhancementsmod-cutedslmod-frostorig-nv-eng
  #1288·RomanAnders90 opened Sep 29, 2026Contributor·

  ·

  · Frontend 1.31.0
  6

  

• Add SM80 native-GQA bitwise, LPT_L2 short-last-block and head-ratio tests

  #1287·harshithkantamneni opened Sep 29, 2026·

  ·
  1

• [FROST][KDA] Disable gate-decay splitting by default

  #1286·icavan opened Sep 29, 2026Contributor·

  ·
  7

• Update Benchmarking Artifacts - backend 9.27.0.21, frontend 0cb8e7d

  #1285·brandonfzhang opened Sep 29, 2026Collaborator·

  ·
  1

• [FROST][KDA] Right-size chunk-16 bprop grid for MBS1 pretraining

  #1284·icavan opened Sep 29, 2026Contributor·

  ·
  6

• Cross Proxy fence to avoid reading stale descriptors in Wgrad kernel

  #1283·vthumbe1503 opened Sep 29, 2026Contributor·

  ·
  8

• Fix grouped SwiGLU SM110 reduction and per-call AMAX correctness

  #1280·0z5a opened Sep 29, 2026Contributor·

  ·
  1

• Frost fp8 sm103 softmax

  area:frostcat-perf-bugorig-nv-eng
  #1278·yanzhuo607 opened Sep 28, 2026Collaborator·

  ·

  · Frontend 1.31.0
  9

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

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

  ·
  1

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

  #1259·0z5a opened Sep 28, 2026Contributor·

  ·
  1

• Prefer FROST for bounded B200 paged D256 prefill

  area:frostarea:global_attentioncat-perf-bugorig-nv-eng
  #1241·YangXu1990uiuc opened Sep 27, 2026Collaborator·

  ·

  · Frontend 1.31.0
  1

• Prepare adaptive THD prefill launches and packed MLA split-KV

  area:frostarea:global_attentioncat-perf-bugorig-nv-eng
  #1240·YangXu1990uiuc opened Sep 27, 2026Collaborator·

  ·

  · Frontend 1.31.0
  1

• Reduce grouped GLU/quant host overhead and route grouped SwiGLU through unified GLU

  #1233·hwanseoc opened Sep 26, 2026Member·

  ·

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

  #1221·Butterfingrz opened Sep 24, 2026Contributor·

  ·
  1

• BSA: optimize SM90 blk64 forward performance

  #1213·Butterfingrz opened Sep 24, 2026Contributor·

  ·
  11

• frost(sdpa/bwd): sm107 (Rubin) d=256 backward -- bf16/fp16 and per-tensor FP8 engines, kernels, tests, perf

  area:frostcat-featuremod-cutedslmod-frostorig-nv-eng
  #1212·RomanAnders90 opened Sep 24, 2026Contributor·

  ·

  · Frontend 1.31.0
  7

  

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

  #1210·scottyokim opened Sep 24, 2026·

  ·
  1

• Add bounded causal GQA backward for SM107

  #1207·layalir opened Sep 23, 2026·

  ·
  11

• [DSv4.1] Reduce Engram backward register pressure on SM100

  area:frostcat-enhancementsorig-nv-eng
  #1204·YangXu1990uiuc opened Sep 23, 2026Collaborator·

  ·

  · Frontend 1.31.0
  4

• [DSv4.1] Support native N-major BF16 expert input gradients

  area:frostcat-featureorig-nv-eng
  #1199·YangXu1990uiuc opened Sep 23, 2026Collaborator·

  ·

  · Frontend 1.31.0
  7

• Add aligned HCA backward for GB300 and Rubin

  area:sparse_attention
  #1198·layalir opened Sep 22, 2026·

  ·

  · Frontend 1.31.0
  25

  

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

  area:global_attentioncat-cleanuporig-nv-eng
  #1186·YangXu1990uiuc opened Sep 22, 2026Collaborator·

  ·

  · Frontend 1.31.0
  70

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

  area:global_attentioncat-cleanuporig-nv-eng
  #1185·YangXu1990uiuc opened Sep 22, 2026Collaborator·

  ·

  · Frontend 1.31.0
  88

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

  area:global_attentioncat-cleanuporig-nv-eng
  #1184·YangXu1990uiuc opened Sep 22, 2026Collaborator·

  ·

  · Frontend 1.31.0
  78

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

  area:sparse_attentioncat-cleanuporig-nv-eng
  #1183·YangXu1990uiuc opened Sep 22, 2026Collaborator·

  ·

  · Frontend 1.31.0
  93