All pull requests

Pull requests

• Replay one CUDA graph per accumulation group

  ciflow/b200ciflow/fake-pgciflow/rlCLA Signed
  #5088·fegin opened Oct 6, 2026Contributor·

  ·

• [checkpoint] Parallelize Hugging Face checkpoint loading

  ciflow/fake-pgciflow/rlCLA Signed
  #5087·fegin opened Oct 6, 2026Contributor·

  ·
  6

• [compile] Gated activations: unbind gate/up inside the compiled region

  ciflow/rlCLA Signed
  #5086·felipemello1 opened Oct 6, 2026Contributor·1 task·

  ·
  4

• [Do not review] [dist_moe] Use bounded token capacity API

  ciflow/rlCLA Signedmeta-exported
  #5085·sanketpurandare opened Oct 6, 2026Contributor·

  ·

• [dist_moe] Reset context after PP metadata inference

  ciflow/rlCLA Signed
  #5084·sanketpurandare opened Oct 6, 2026Contributor·

  ·
  1

• [rl] Fix RL CPU unit tests resolving a month-old torch nightly

  ciflow/fake-pgciflow/rlCLA Signed
  #5083·wwwjn opened Oct 6, 2026Contributor·

  ·
  3

• [rl] Keep only the NCCL env vars batch invariance needs

  ciflow/fake-pgciflow/rlCLA Signed
  #5081·wwwjn opened Oct 6, 2026Contributor·

  ·

• Support multiple EMA copies

  ciflow/fake-pgCLA Signed
  #5074·ruisizhang123 opened Oct 6, 2026Contributor·

  ·
  3

• Make FlexAttention compiler backend and options configurable

  ciflow/rl
  #5073·aws-dixuan opened Oct 6, 2026·

  ·
  2

• [observability] Profile Neuron devices with the PrivateUse1 activity

  #5072·kvshbg-aws opened Oct 5, 2026·

  ·
  1

• [CP] Add graph-safe varlen K/V all-gather

  ciflow/rlCLA Signed
  #5071·fegin opened Oct 5, 2026Contributor·

  ·

• Save trailing adds so replay skips them

  ciflow/fake-pgciflow/rlCLA Signed
  #5070·anijain2305 opened Oct 5, 2026Contributor·

  ·

• [3/3] [graph_trainer] Bucket routed experts by parent scope

  ciflow/fake-pgCLA Signed
  #5067·IvanKobzarev opened Oct 5, 2026Contributor·3/3·

  ·

• [TorchTitan] Make DeepSeek V3's HF save file map match what the save writes

  CLA Signed
  #5065·ivy-zhou opened Oct 5, 2026Contributor·

  ·

• [TorchTitan] Use unquantized GPT-OSS expert names unless loading MXFP4

  ciflow/fake-pgCLA Signed
  #5064·ivy-zhou opened Oct 5, 2026Contributor·

  ·
  1

• [TorchTitan] Tell to_hf and from_hf when an HF load is quantized

  ciflow/fake-pgciflow/rlCLA Signed
  #5063·ivy-zhou opened Oct 5, 2026Contributor·

  ·
  2

• Add precomputed routing for weight sync

  ciflow/fake-pgciflow/rlCLA Signed
  #5061·tushar00jain opened Oct 5, 2026Contributor·

  ·
  11

• [2/3] [graph_trainer] Fuse WGrad accumulation through views

  ciflow/fake-pgciflow/h100.8CLA Signed
  #5059·IvanKobzarev opened Oct 5, 2026Contributor·2/3·

  ·

• [1/3] [graph_trainer] Extract FSDP split-pad-cat together with reduce_grad chains

  ciflow/fake-pgciflow/h100.8CLA Signed
  #5058·IvanKobzarev opened Oct 5, 2026Contributor·1/3·

  ·

• Refine new_args to include only provided arguments

  ciflow/fake-pgCLA Signed
  #5056·MaddipatlaChetan24 opened Oct 5, 2026Contributor·

  ·
  2

• Fix per-depth MTP loss normalization

  ciflow/fake-pgciflow/rlCLA Signed
  #5055·tianyu-l opened Oct 5, 2026Contributor·

  ·
  3

• [dist_moe] Add DistMoE for the vLLM generator inference path

  ciflow/rlCLA Signed
  #5046·wwwjn opened Oct 3, 2026Contributor·

  ·
  21

• [debug] Save the device-mesh layout with debug.save_parallelism_file

  ciflow/fake-pgciflow/rlCLA Signed
  #5045·francesco-bertolotti opened Oct 3, 2026Contributor·

  ·
  13

• [graph_trainer] Pass unbacked symbol bindings to GraphPP dW graph

  CLA Signed
  #5037·wconstab opened Oct 2, 2026Contributor·

  ·

• [Perf][Compile] Add local compilation for Kimi attention residual

  ciflow/rlCLA Signed
  #5033·acisseJZhong opened Oct 2, 2026Contributor·

  ·