All pull requests

Pull requests

• local_spmd regions with spmd.local_map

  ciflow/fake-pgciflow/rlCLA Signed
  #4960·pianpwk opened Sep 30, 2026Contributor·

  ·
  2

• Stop pinning the routed-expert w2 output

  ciflow/rlCLA Signed
  #4959·anijain2305 opened Sep 30, 2026Contributor·

  ·

• Declare remat regions around MoE token reordering and combine

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

  ·

• Keep the EP token-count sync inside its remat region

  ciflow/rlCLA Signed
  #4957·anijain2305 opened Sep 30, 2026Contributor·

  ·

• Declare remat regions inside ColumnParallelLinear

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

  ·

• [Fix] RoutedExperts output_postprocess parameter under FSDP

  ciflow/fake-pgciflow/rlCLA Signed
  #4954·jinsooihm opened Sep 29, 2026Contributor·

  ·

• [graph_trainer] Link public Slack channel in README

  ciflow/fake-pgCLA Signed
  #4953·SherlockNoMad opened Sep 29, 2026Contributor·

  ·

• [graph_trainer] Remove graph-based EP chunking

  ciflow/fake-pgCLA Signed
  #4952·SherlockNoMad opened Sep 29, 2026Contributor·

  ·

• [graph_trainer] Keep Shard(0) for routed experts when EP covers the experts

  ciflow/fake-pgCLA Signed
  #4949·wconstab opened Sep 29, 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/rlCLA Signed
  #4945·tushar00jain opened Sep 29, 2026Contributor·2/2·

  ·

• [graph_trainer] Avoid PP=1 metadata dry run for decoder models

  ciflow/fake-pgCLA Signed
  #4944·IvanKobzarev opened Sep 29, 2026Contributor·

  ·
  1

  1

• [DSV4] Support packed documents

  CLA Signed
  #4943·drisspg opened Sep 29, 2026Contributor·6/6·

  ·

• [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·

  ·

• Remove torchcomms dependency

  ciflow/fake-pgciflow/rlCLA Signed
  #4941·tushar00jain opened Sep 29, 2026Contributor·1/2·

  ·
  1

  1

• [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·

  ·
  2

• [DSV4] Place compressed KV at its group's first token when positions is None

  ciflow/fake-pgCLA Signed
  #4932·drisspg opened Sep 28, 2026Contributor·5/6·

  ·

• [Perf][Compile] Add local compilation for SwiGLU

  ciflow/fake-pgciflow/rlCLA Signed
  #4931·acisseJZhong opened Sep 28, 2026Contributor·4/4·

  ·
  2

• Fix per-depth MTP loss normalization

  ciflow/fake-pgciflow/rlCLA Signed
  #4929·marcusacosta opened Sep 28, 2026·

  ·
  4

• [Perf][Compile] Add local compilation for OffsetRMSNorm

  ciflow/fake-pgciflow/rlCLA Signed
  #4928·acisseJZhong opened Sep 28, 2026Contributor·3/4·

  ·

• [2/2] [graph_trainer] PP=1 GA schedule with FSDP overlap

  ciflow/dockerciflow/fake-pgciflow/rlCLA Signed
  #4927·IvanKobzarev opened Sep 28, 2026Contributor·

  ·
  17

• Add Fp32OutputLinear for the LM head and router gate; delete CastLinear and RouterGateLinear

  ciflow/fake-pgciflow/rlCLA Signed
  #4923·felipemello1 opened Sep 28, 2026Contributor·

  ·
  11