All pull requests

Pull requests

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

  #4228·HJSang opened Oct 20, 2026Collaborator·

  ·

• [9/9] Test the sglang-torchtpu APIs the TPU path calls, on a CPU runner

  #4226·HJSang opened Oct 20, 2026Collaborator·9/9·

  ·
  3

• [8/9] Add docker/Dockerfile.tpu: miles on TPU v7x through torch_tpu

  #4225·HJSang opened Oct 20, 2026Collaborator·8/9·

  ·

• Index rollout replay data as row segments and add --enable-delay-token-replay-assembly

  #4224·guapisolo opened Oct 20, 2026Collaborator·

  ·

• docs: add SYS_PTRACE to the NVIDIA docker run so colocated weight updates work (#364)

  #4223·wangweinan opened Oct 20, 2026·

  ·
  1

• fix(megatron): reset Mamba state at packed-sample boundaries in THD

  #4222·omri-manolo opened Oct 20, 2026·

  ·

• [DO NOT MERGE] test new main

  #4221·fzyzcjy opened Oct 20, 2026Collaborator·

  ·
  8

• [7/9] Routing replay on torchtitan/TPU; on TPU route uncaptured rows by the router

  #4218·HJSang opened Oct 20, 2026Collaborator·7/9·

  ·

• [6/9] Add --activation-in-fp32: SiLU in fp32 in the trainer and the rollout engines

  #4217·HJSang opened Oct 20, 2026Collaborator·6/9·

  ·

• [5/9] TPU trainer: splash attention, fused log-prob, bf16 FSDP gradient reduce (opt-in)

  #4215·HJSang opened Oct 20, 2026Collaborator·5/9·

  ·

• [4/9] TPU: compute log-probs and the policy loss at the padded bucket shape

  #4214·HJSang opened Oct 20, 2026Collaborator·4/9·

  ·

• [3/9] Run the torchtitan trainer on TPU (torch_tpu)

  #4213·HJSang opened Oct 20, 2026Collaborator·3/9·

  ·

• Read rollout replay data from the SGLang output store on the trainer

  #4212·guapisolo opened Oct 20, 2026Collaborator·

  ·

• feat(lora): native LoRA for Muse Spark with merged rollout weights

  #4211·guapisolo opened Oct 20, 2026Collaborator·

  ·

• [2/9] Add disk and shared-memory weight sync for engines without NCCL or CUDA IPC

  #4210·HJSang opened Oct 20, 2026Collaborator·2/9·

  ·

• [1/9] Route the training path's device calls through miles.utils.accelerator

  #4209·HJSang opened Oct 20, 2026Collaborator·1/9·

  ·

• Share the DeepGEMM JIT cache across rollout engines

  #4208·adrenaline21 opened Oct 20, 2026Collaborator·2 of 3 tasks·

  ·

• DeepSeek-V4.1-Flash LoRA through native LoRA

  #4205·yueming-yuan opened Oct 19, 2026Collaborator·

  ·

• fix(dp_schedule): drop trailing rollouts when long compact samples cannot align

  #4203·anthonyhe-bot opened Oct 19, 2026Contributor·

  ·

• feat(rollout): add checkpointed multi-dataset mixing

  #4200·saiyambharara opened Oct 19, 2026·

  ·

• Keep service and Ray worker ports out of the host's ephemeral port range

  #4199·bobboli opened Oct 19, 2026Contributor·

  ·

• feat(rollout): select partial groups at admission

  #4198·nanjiangwill opened Oct 19, 2026Contributor·

  ·

• feat(rollout): support partial groups on abort for compact outputs

  #4167·nanjiangwill opened Oct 18, 2026Contributor·

  ·
  2

• fix(rollout): zero uniform float32 rewards

  #4165·mmbajo opened Oct 18, 2026·

  ·
  1

• fix(sglang): initialize parser plugins before registry use

  #4164·mmbajo opened Oct 18, 2026·

  ·
  1