All pull requests

Pull requests

• release: bump miles version to 0.1.3

  #4275·HJSang opened Oct 22, 2026Collaborator·

  ·
  4

• Add the miles-ext-kernels submodule behind --use-miles-ext-kernels

  #4274·yueming-yuan opened Oct 22, 2026Collaborator·

  ·

• [2/4] MLLM Refactor Core: build training inputs in token space

  #4272·Rockdu opened Oct 22, 2026Collaborator·2/2·3 of 6 tasks·

  ·

• feat(async): bound staleness when the producer submits a group

  #4270·Arist12 opened Oct 21, 2026Contributor·

  ·
  1

• [2/10] Add docker/Dockerfile.tpu: miles on TPU v7x from sglang-torchtpu's release image

  #4268·HJSang opened Oct 21, 2026Collaborator·1/9·

  ·
  12

• [6/6] Add --use-output-store

  #4267·guapisolo opened Oct 21, 2026Collaborator·6/6·

  ·
  2

• [5/10] TPU trainer: --titan-fsdp-gather-once gathers FSDP params once per pass

  #4265·HJSang opened Oct 21, 2026Collaborator·4/9·

  ·
  4

• [5/6] Delete output store objects once nothing can read them

  #4264·guapisolo opened Oct 21, 2026Collaborator·5/6·

  ·

• build(docker): bump torch_memory_saver to 118b09b

  run-ci-image
  #4247·guapisolo opened Oct 21, 2026Collaborator·

  ·
  1

• Add Bellman Policy Optimization (BPO) by Apodex

  #4246·wl45361 opened Oct 21, 2026·

  ·
  1

• [4/6] Materialize the replay columns of output store plans on the trainer

  #4244·guapisolo opened Oct 21, 2026Collaborator·4/6·

  ·

• [3/6] Build output store plans from rollout responses

  run-ci
  #4243·guapisolo opened Oct 21, 2026Collaborator·3/6·

  ·
  11

• [2/6] Carry the output store plan on samples, the session reply and the train dict

  run-ci
  #4242·guapisolo opened Oct 21, 2026Collaborator·2/6·

  ·
  6

• perf(step5): reuse frozen indexer selections during recompute

  #4241·ShuZihan opened Oct 21, 2026·

  ·
  1

• [1/4] MLLM Refactor Core: send media refs to SGLang

  #4240·Rockdu opened Oct 21, 2026Collaborator·1/2·2 of 6 tasks·

  ·

• fix(fsdp): synchronize model D2H before shard repadding

  #4239·Palind-Rome opened Oct 21, 2026·

  ·
  1

  2

• docs: note that --async-save requires --use-persistent-ckpt-worker

  #4237·Arist12 opened Oct 21, 2026Contributor·

  ·
  1

• [1/6] Add the output store merge plan: Seg, Segments, StorePlan

  #4236·guapisolo opened Oct 21, 2026Collaborator·1/6·

  ·
  13

• Reject speculative decoding under --true-on-policy-mode

  #4234·HJSang opened Oct 20, 2026Collaborator·

  ·
  8

• [AMD] fix(mfu): add Instinct MI300X/MI325X/MI350X/MI355X to the peak BF16 table

  #4233·Arist12 opened Oct 20, 2026Contributor·

  ·

• fix(rollout): measure fully-async rollout throughput between batches, not over the drain

  #4232·Arist12 opened Oct 20, 2026Contributor·

  ·

• [AMD] fix(trainer): enable ROCm TransformerEngine MXFP8 for --fp8-recipe mxfp8

  #4231·Arist12 opened Oct 20, 2026Contributor·

  ·

• fix(train_async): save on --save-trigger-sentinel while waiting for a rollout

  #4230·Arist12 opened Oct 20, 2026Contributor·

  ·

• fix(megatron): do not step the LR scheduler twice on resume

  #4229·Arist12 opened Oct 20, 2026Contributor·

  ·

• fix(loss): keep empty CP ranks in the graph without anchors

  #4228·HJSang opened Oct 20, 2026Collaborator·

  ·
  4