All pull requests

Pull requests

• fix(trainer): skip_optimizer saves bf16 weights only (ENG-6561)

  #3915·cabrinha opened Oct 7, 2026Collaborator·

  ·

• feat(sft): add deterministic source selection cycles

  #3914·nxphi47 opened Oct 7, 2026Contributor·

  ·

• fix(inference): let the LoRA adapter take the base model name on vLLM 0.31

  #3913·faresobeid opened Oct 7, 2026Contributor·

  ·

• perf(trainer): make the FSDP reduce-scatter buffer cap configurable and default it to 1

  #3912·garrett361 opened Oct 7, 2026Contributor·

  ·

• fix(trainer): run FA2 and FA3 through custom ops so fullgraph compile works

  #3911·garrett361 opened Oct 6, 2026Contributor·

  ·

• fix(ci): skip GPU tests on docs-only changes

  #3910·samsja opened Oct 6, 2026Member·

  ·

• fix(ci): uv cache on h200 runners, drop phantom 4xa6000 job, cap fan-out

  #3909·samsja opened Oct 6, 2026Member·

  ·

• fix(sft): apply per-message loss masks from a message_loss_mask column

  #3908·nxphi47 opened Oct 6, 2026Contributor·

  ·
  2

• feat: overlap trainer image preprocessing with GPU compute

  #3907·KyleWong288 opened Oct 6, 2026Member·

  ·
  1

• fix(resume): resume from the latest complete checkpoint, not the newest step dir

  #3906·faresobeid opened Oct 6, 2026Contributor·

  ·

• refactor(trainer): clean up and document ParallelDims meshes

  #3905·garrett361 opened Oct 6, 2026Contributor·

  ·

• feat: by-handle payload fields table (TRR weights, score-centering logprobs)

  #3904·faresobeid opened Oct 6, 2026Contributor·

  ·

• feat: write router-replay ids and sampling masks by handle on multi-node runs

  #3903·faresobeid opened Oct 6, 2026Contributor·

  ·

• feat: by-handle payload segments: reader, packing, CP-window reads

  #3902·faresobeid opened Oct 6, 2026Contributor·

  ·

• fix(orchestrator): remove hardcoded TARGET_LAG

  #3901·jurij-jukic opened Oct 6, 2026·

  ·

• perf(inference): pack /generate per-token payloads and tune API-server GC

  #3900·faresobeid opened Oct 6, 2026Contributor·

  ·
  4

• feat(eval): stream infinite tasksets through a standalone eval

  #3898·parkerpettit opened Oct 6, 2026Contributor·

  ·

• feat(inference): check native KV offload fits in /dev/shm at launch

  #3896·faresobeid opened Oct 6, 2026Contributor·

  ·

• feat(router): configurable vllm-router health check and circuit breaker

  #3895·faresobeid opened Oct 6, 2026Contributor·

  ·

• perf(trainer): fused SignSGD step, skip no-op gradient rescales, real debug random init

  #3894·samsja opened Oct 6, 2026Member·

  ·

• fix: keep live trace previews scoped to their branches

  #3893·hallerite opened Oct 6, 2026Member·

  ·

• perf(trainer): DeepSeek-V4.1 on Blackwell: cuDNN attention + indexer, fused SignSGD

  #3891·samsja opened Oct 5, 2026Member·

  ·

• Add ModelExpress public-client weight broadcast

  #3890·zhengluo-nv opened Oct 5, 2026·

  ·
  24

• security(ci): pin BuildKit builder to v0.33.1 and rotate registry buildcache (CVE-2026-93318)

  #3889·Apocrathia opened Oct 5, 2026·

  ·
  6

• perf(microbatch refinement) refinement swapping changes

  #3886·LazaroHurtado opened Oct 5, 2026·

  ·