All pull requests

Pull requests

• ci: write the rules for adding a CI test in .claude/rules

  #3953·guapisolo opened Oct 7, 2026Collaborator·

  ·

• Request session top-k candidates as flat base64 arrays

  #3952·HJSang opened Oct 7, 2026Collaborator·

  ·

• p2p: ModelReplica, a model replica without parameter storage in the engine's reload state

  run-ci
  #3950·yueming-yuan opened Oct 7, 2026Collaborator·

  ·

• Allow score centering with DFlash speculative decoding

  #3946·jvmncs opened Oct 6, 2026·

  ·

• p2p: add StagingBuffers, bounded transfer memory reused across loads

  #3945·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• [DO NOT MERGE] debug: H200 CI runner IB/NVSHMEM probe for DeepEP failures

  #3944·Zhichenzzz opened Oct 6, 2026Collaborator·

  ·
  2

• p2p: keep model replicas, the transport and the registration for the process

  #3943·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• p2p: reject a layout mismatch before any byte is written

  #3942·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• p2p: a write thread per rollout engine; fail the update when any write fails or times out

  #3941·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• p2p: assign rollout engine ranks over the engines and placement handed over

  #3940·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• p2p: move the protocol's helpers into protocols/utils and protocols/transports (no behavior change)

  #3939·yueming-yuan opened Oct 6, 2026Collaborator·

  ·

• test(inkling): run Inkling-Small CI on a 6-layer slice with a global-attention layer

  #3937·HJSang opened Oct 6, 2026Collaborator·

  ·
  4

• fsdp: one ArchAdapter per architecture; HF padding-free kwargs for GatedDeltaNet packing

  run-ci-fsdp
  #3935·HJSang opened Oct 6, 2026Collaborator·

  ·

• feat: add offline text SFT for DiffusionGemma on FSDP2

  #3931·adavyas opened Oct 6, 2026·

  ·

• Add experimental TMax GRPO score-centering recipe for Qwen3.5-9B

  #3929·HJSang opened Oct 6, 2026Collaborator·

  ·

• Honor the weight-check selector for the startup snapshot and reset

  #3928·jvmncs opened Oct 6, 2026·

  ·

• fix(tools): refuse an FSDP export that would ship untrained weights

  #3926·ysjprojects opened Oct 5, 2026·

  ·

• fix(tools): read the checkpoint args the fork stores, not common.pt

  #3925·ysjprojects opened Oct 5, 2026·

  ·

• fix(megatron): set variable_seq_lengths in the shared Megatron defaults

  #3924·ysjprojects opened Oct 5, 2026·

  ·

• fix(ray): resolve the head node from GCS, not the dashboard

  #3923·ysjprojects opened Oct 5, 2026·

  ·
  1

• release: bump miles version to 0.1.2

  #3922·Zhichenzzz opened Oct 5, 2026Collaborator·

  ·

• chore(args): warn that --over-sampling-batch-size is ignored under --fully-async

  #3921·lizamd opened Oct 5, 2026Contributor·

  ·
  4

• feat(filter_hub): add mask_truncated_zero_reward sample filter

  #3920·lizamd opened Oct 5, 2026Contributor·

  ·
  5

• fix(swe-agent-harbor): resume from save_dir instead of retraining from ref_load

  #3919·lizamd opened Oct 5, 2026Contributor·

  ·
  2

• fix(swe-agent-harbor): cap engine context in run.py and the GLM-4.7 async launcher

  #3918·lizamd opened Oct 5, 2026Contributor·

  ·
  5