All pull requests

Pull requests

• [WIP][DistMuon] Compile local optimizer updates in three stages

  CLA Signed
  #6029·jiesihu opened Nov 11, 2026Contributor·

  ·

• [rl] SGLang generator: share KV across generators through Mooncake Store

  ciflow/rlCLA Signed
  #6027·acisseJZhong opened Nov 11, 2026Contributor·

  ·

• [rl] Muse Spark image prompts and a SLAKE LoRA GRPO recipe for Muse Spark 816B

  ciflow/fake-pgciflow/rlCLA Signed
  #6025·wwwjn opened Nov 11, 2026Contributor·2/3·

  ·
  14

• [rl] SGLang engine: take the request's Prompt and the vision_processor argument

  ciflow/fake-pgciflow/rlCLA Signed
  #6024·wwwjn opened Nov 11, 2026Contributor·

  ·
  1

• Schedule-driven activation offload for pipeline stages

  CLA Signed
  #6022·anijain2305 opened Nov 11, 2026Contributor·2/2·

  ·

• [rl] SGLang generator: use SGLang's public attention-layer and FA kernel hooks

  ciflow/rlCLA Signed
  #6021·acisseJZhong opened Nov 11, 2026Contributor·

  ·

• Activation offloading on torch_remat saved-tensor hooks

  ciflow/fake-pgCLA Signed
  #6020·anijain2305 opened Nov 10, 2026Contributor·1/2·

  ·

• [rl] generator: receive SGLang scheduler outputs through output_sink

  ciflow/rlCLA Signed
  #6018·acisseJZhong opened Nov 10, 2026Contributor·

  ·

• [CI deps] Bump attn-gym pin to v0.0.19 (0d0197fb)

  ciflow/dockerciflow/fake-pgciflow/rlCLA Signed
  #6017·shuhuayu opened Nov 10, 2026Contributor·

  ·

• [rl][docs] README: bring the RL model support table up to date

  ciflow/rlCLA Signed
  #6016·wwwjn opened Nov 10, 2026Contributor·

  ·

• Support tensor and context parallelism with parallelism.fsdp_backend=FlexShardConfig()

  ciflow/rlCLA Signed
  #6014·anshul-si opened Nov 10, 2026Contributor·

  ·

• Support Hugging Face checkpoint loads with parallelism.fsdp_backend=FlexShardConfig()

  ciflow/rlCLA Signed
  #6013·anshul-si opened Nov 10, 2026Contributor·

  ·

• Support checkpointing with parallelism.fsdp_backend=FlexShardConfig()

  ciflow/rlCLA Signed
  #6012·anshul-si opened Nov 10, 2026Contributor·

  ·

• Support ChunkedLossWrapper with the FlexShard backend

  CLA Signed
  #6011·anshul-si opened Nov 10, 2026Contributor·

  ·
  11

• Support expert parallelism with parallelism.fsdp_backend=FlexShardConfig()

  ciflow/fake-pgciflow/rlCLA Signed
  #6010·anshul-si opened Nov 10, 2026Contributor·

  ·
  2

• [rl] Admit new rollout groups by KV room, before their envs start

  ciflow/rlCLA Signed
  #6008·felipemello1 opened Nov 10, 2026Contributor·3 tasks·

  ·

• [kimi_k3_1] Add Kimi K3.1 Pro with Gated DSA

  ciflow/rlCLA Signed
  #6007·JavaZeroo opened Nov 10, 2026Contributor·

  ·
  1

• [rl] Install Monarch nightly in new RL environments

  ciflow/dockerciflow/fake-pgciflow/rlCLA Signed
  #6006·pzhan9 opened Nov 10, 2026Contributor·

  ·
  11

• [recipes] Muse Spark SLAKE SFT: full vision tower + LoRA decoder

  ciflow/fake-pgciflow/rlCLA Signed
  #6004·fegin opened Nov 10, 2026Contributor·

  ·
  3

• [lora] Keep selected modules fully trainable in LoRATransform

  ciflow/fake-pgciflow/rlCLA Signed
  #6003·fegin opened Nov 10, 2026Contributor·

  ·
  8

• [rl] Sync the model state dict with ModelStateDictTransport behind a config flag

  ciflow/rlCLA Signed
  #5997·d4l3k opened Nov 10, 2026Member·

  ·
  3

• [rl] Add ModelStateDictTransport over torch.distributed._transport

  ciflow/rlCLA Signed
  #5996·d4l3k opened Nov 10, 2026Member·

  ·
  1

• [config] Reject unsupported TP/EP degrees for FLUX

  ciflow/fake-pgCLA Signed
  #5995·bantianyue opened Nov 10, 2026Contributor·

  ·
  1

  3

• [data] Add vision_encoder_cost for heaviest-first multimodal ordering

  ciflow/rlCLA Signed
  #5993·pianpwk opened Nov 9, 2026Contributor·

  ·

• [data] Order each window of microbatches heaviest first across DP ranks

  ciflow/rlCLA Signed
  #5992·pianpwk opened Nov 9, 2026Contributor·

  ·