All pull requests

Pull requests

• [tokenizer] Preserve non-special added tokens

  ciflow/fake-pgCLA Signed
  #5470·mottopanikeiku opened Oct 17, 2026·

  ·
  1

• [DO NOT review, pending finalizing][Kimi K3] Support multimodal video training

  ciflow/rlCLA Signed
  #5469·QIU023 opened Oct 17, 2026Contributor·

  ·
  1

• [DistMuon] Balance Owned and sharded compute placement

  CLA Signed
  #5468·JavaZeroo opened Oct 17, 2026Contributor·

  ·

• [models] Let the attn_gym facade defer missing-backend import errors

  ciflow/fake-pgciflow/rlCLA Signed
  #5467·mkhilari opened Oct 17, 2026·

  ·
  5

• [overrides] Run fused SwiGLU and Helion RoPE overrides on any accelerator

  ciflow/fake-pgCLA Signed
  #5466·githubsgi opened Oct 16, 2026Contributor·

  ·
  3

• [dist_moe] Use fast fused SwiGLU mega kernel for inference

  ciflow/b200ciflow/fake-pgciflow/rlCLA Signed
  #5465·andrewor14 opened Oct 16, 2026Contributor·

  ·
  2

• Make TorchTitan compile backend configurable

  ciflow/fake-pgciflow/rlCLA Signed
  #5463·ryan-li-ryanllli-aws opened Oct 16, 2026·

  ·
  2

• .gitignore: ignore .monarch/ and structured_logs/

  CLA Signed
  #5462·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [rl][Observability] Rollout workers write structured logs

  ciflow/rlCLA Signed
  #5461·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [compile][perf] SwiGLU region: write the packed gate/up gradient with one store per half

  ciflow/rlCLA Signed
  #5460·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [muse_spark][perf] Compile RoPE with the q/k norms; drop the Helion RoPE override

  CLA Signed
  #5459·felipemello1 opened Oct 16, 2026Contributor·5/5·

  ·

• [muse_spark][perf] Compile the residual-stream RMSNorms (muse_spark_rmsnorm region)

  CLA Signed
  #5458·felipemello1 opened Oct 16, 2026Contributor·4/5·

  ·

• [rl] Verifiers 0.4.0

  ciflow/rlCLA Signed
  #5456·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [rl] Trainer MFU over real documents: no padding, attention at each document's length

  ciflow/rlCLA Signed
  #5455·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [rl] Verifiers generation server: faster request path, required sampling params

  ciflow/rlCLA Signed
  #5454·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [rl] Verifiers generation server: no per-request access log

  ciflow/rlCLA Signed
  #5453·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [rl] Rubric.truncation_penalty: keep partial credit on truncated rollouts, add a penalty

  ciflow/rlCLA Signed
  #5451·felipemello1 opened Oct 16, 2026Contributor·

  ·

• Use torch.accelerator for device, stream and memory APIs

  ciflow/fake-pgciflow/rlCLA Signed
  #5450·jemitche1 opened Oct 16, 2026·

  ·
  7

• [rl] Set the trainer's CPU thread count after NUMA binding

  ciflow/fake-pgciflow/rlCLA Signed
  #5447·felipemello1 opened Oct 16, 2026Contributor·2 tasks·

  ·
  1

• [attention] Flex attention: recompile limit 16, and fail instead of running eager

  ciflow/rlCLA Signed
  #5442·felipemello1 opened Oct 16, 2026Contributor·

  ·

• [moe][perf] Rank small route sets pairwise instead of a one-hot cumsum (9/9)

  ciflow/rlCLA Signed
  #5437·felipemello1 opened Oct 15, 2026Contributor·9/9·

  ·

• [moe][perf] Argsort expert ids on the narrowest integer type (8/9)

  ciflow/rlCLA Signed
  #5436·felipemello1 opened Oct 15, 2026Contributor·8/9·

  ·

• [muse_spark][perf] Compile the attention's per-head norms and output gate

  CLA Signed
  #5434·felipemello1 opened Oct 15, 2026Contributor·3/5·

  ·

• [muse_spark][perf] Enable the Muse Spark regions in every recipe

  CLA Signed
  #5433·felipemello1 opened Oct 15, 2026Contributor·2/5·

  ·

• [muse_spark][perf] Run the fp32-stream RMSNorms on the fused kernel

  CLA Signed
  #5432·felipemello1 opened Oct 15, 2026Contributor·1/5·

  ·