All pull requests

Pull requests

• [2/2] [graph_trainer] Classify forward mutations by phase

  ciflow/fake-pgCLA Signed
  #4990·IvanKobzarev opened Oct 1, 2026Contributor·

  ·

• Use BlockShard for routed-expert FSDP when FSDP degree > local experts

  CLA Signed
  #4989·weifengpy opened Oct 1, 2026Contributor·

  ·

• [checkpoint] Optional parallel DCP reader for faster checkpoint loads

  ciflow/rlCLA Signed
  #4988·felipemello1 opened Oct 1, 2026Contributor·

  ·

• [rl] Bug fix: Do not give vLLM a fake eos_token_id

  ciflow/fake-pgciflow/rlCLA Signed
  #4987·wwwjn opened Oct 1, 2026Contributor·

  ·
  3

• [Perf][Compile] Add local compilation for cos/sin RoPE

  ciflow/b200ciflow/fake-pgciflow/h100.8ciflow/real-pgciflow/rlCLA Signed
  #4985·acisseJZhong opened Oct 1, 2026Contributor·5/5·

  ·

• [profiler] add opt-in Cuspy profiling backend

  ciflow/rlCLA Signed
  #4984·pianpwk opened Oct 1, 2026Contributor·

  ·

• [graph_trainer] Pass one config through GraphRuntime construction

  ciflow/fake-pgCLA Signed
  #4983·SherlockNoMad opened Sep 30, 2026Contributor·

  ·
  1

• [graph_trainer] Resolve one GraphExecutionPlan for PP=1 and PP>1

  ciflow/fake-pgCLA Signed
  #4980·SherlockNoMad opened Sep 30, 2026Contributor·

  ·
  1

• [graph_trainer] Always enable graph pass instrumentation

  ciflow/fake-pgCLA Signed
  #4977·SherlockNoMad opened Sep 30, 2026Contributor·

  ·
  2

• [MoE] Annotate routed-expert weights on the dense mesh when EP is off

  ciflow/fake-pgciflow/rlCLA Signed
  #4973·pianpwk opened Sep 30, 2026Contributor·

  ·

• [CI] Update Qwen3 MoE Real PG numerics golden

  ciflow/fake-pgciflow/real-pgCLA Signed
  #4972·wconstab opened Sep 30, 2026Contributor·

  ·
  2

• Split WANDB_RUN_TAGS into a list before passing to wandb

  ciflow/fake-pgCLA Signed
  #4965·u7k4rs6 opened Sep 30, 2026·

  ·
  8

• [not for land] use mxfp8 kernels from PyTorch core

  CLA Signed
  #4964·vkuzo opened Sep 30, 2026Contributor·

  ·
  1

• [DO NOT review, stack on #4656][Kimi K3] PP: store each attention-residual block once per rank and free it at its last reader

  CLA Signed
  #4963·QIU023 opened Sep 30, 2026Contributor·

  ·
  1

• [DistMuon] Unify matrix compute with BlockShard

  CLA Signed
  #4962·shuhuayu opened Sep 30, 2026Contributor·

  ·

• [Float8] Remove Float8 training support

  ciflow/fake-pgciflow/rlCLA Signed
  #4961·acisseJZhong opened Sep 30, 2026Contributor·

  ·
  7

• Declare remat regions around MoE token reordering and combine

  ciflow/rlCLA Signed
  #4958·anijain2305 opened Sep 30, 2026Contributor·

  ·

• Give every Linear its own remat region

  ciflow/rlCLA Signed
  #4956·anijain2305 opened Sep 30, 2026Contributor·

  ·

• [Kimi K3] Make vision QKV checkpoint placement explicit

  CLA Signed
  #4947·elfiegg opened Sep 29, 2026Contributor·

  ·

• [rl] Stage generator weights in CPU memory

  ciflow/fake-pgciflow/rlCLA Signed
  #4945·tushar00jain opened Sep 29, 2026Contributor·

  ·
  19

• [RL] Add Terminal-Bench 2.1 experiment via Verifiers (Qwen3.5 9B, 35B-A3B)

  ciflow/fake-pgciflow/rlCLA Signed
  #4942·wwwjn opened Sep 29, 2026Contributor·

  ·
  3

• [rl] Publish a tied weight once per trainer-to-generator weight sync

  ciflow/rlCLA Signed
  #4940·lokic233 opened Sep 29, 2026·

  ·
  1

• [rl] Size the trainer LR schedule from async_loop.num_training_steps

  ciflow/rlCLA Signed
  #4939·lokic233 opened Sep 29, 2026·

  ·
  1

• [rl] Fix default validation console keys so the reward is printed

  ciflow/rlCLA Signed
  #4938·lokic233 opened Sep 29, 2026·

  ·
  1

• [graph_trainer] Prefer live SimpleFSDP unshard chains

  ciflow/fake-pgCLA Signed
  #4934·sanketpurandare opened Sep 29, 2026Contributor·

  ·
  4