All pull requests

Pull requests

• [megatron, model] refactor: modernize Megatron integration - use native output callbacks [2/5]

  #8439·ISEEKYAN opened Oct 29, 2026Collaborator·4 of 5 tasks·

  ·
  1

• [megatron] refactor: modernize Megatron integration - centralize compatibility [1/5]

  #8438·ISEEKYAN opened Oct 29, 2026Collaborator·5 of 6 tasks·

  ·
  1

• [tests,hardware] pytest changes for device-agnostic case

  #8437·kahlun opened Oct 29, 2026Contributor·5 of 7 tasks·

  ·
  4

• [rollout] fix: honor repetition penalty in agent loops

  #8436·taking-lying-flat opened Oct 29, 2026·7 of 8 tasks·

  ·
  3

• [rollout, sglang] fix: reduce tensor lifetime overlap during bucketed weight sync

  #8435·Levius-Fubuki opened Oct 29, 2026·5 of 8 tasks·

  ·
  1

• [trainer] fix: seed SFT dataloader generators so resume keeps the restored CPU RNG

  #8434·JonnesLin opened Oct 28, 2026·5 of 8 tasks·

  ·
  1

  2

• [hardware, doc] feat: update the ROCm Dockerfile for Primus v26.8

  #8432·tmm77 opened Oct 28, 2026Contributor·5 of 8 tasks·

  ·
  5

• [megatron, doc] feat: MXFP4 MoE recipe and Qwen3-30B-A3B example

  #8426·0z5a opened Oct 28, 2026Contributor·5 of 6 tasks·

  ·
  1

• [megatron, perf] fix: fake-quantize all MXFP4 expert weights of a grouped linear in one call

  #8425·0z5a opened Oct 28, 2026Contributor·5 of 6 tasks·

  ·
  1

• [trainer, algo] fix: give GDPO its reward components on the V1 trainer

  #8424·flyingdoubleG opened Oct 28, 2026·4 of 7 tasks·

  ·
  6

• [algo, doc] fix: use alpha = B/(B-M) in GPG advantage estimator

  #8422·Beiciccc opened Oct 28, 2026·6 of 8 tasks·

  ·
  3

• [algo] fix: drop IcePop-rejected tokens from geo_mean loss

  #8421·mirzasnowy opened Oct 28, 2026·4 of 7 tasks·

  ·
  1

  2

• [data] fix: accept null media columns in MultiTurnSFTDataset

  #8420·kjgggggg opened Oct 28, 2026·5 of 6 tasks·

  ·
  2

• [fsdp, veomni] feat: opt in to FP32 logits before temperature scaling

  #8419·Li-Jicheng opened Oct 28, 2026·5 of 7 tasks·

  ·
  2

• [veomni, lora, vllm] feat: support M3.1-Flash expert LoRA rollout export

  #8418·Li-Jicheng opened Oct 28, 2026·3 of 8 tasks·

  ·
  2

• [veomni, vllm, model] feat: support M3.1-Flash multimodal full-weight RL

  #8417·Li-Jicheng opened Oct 28, 2026·3 of 7 tasks·

  ·
  2

• [vllm, rollout] fix: fold unquantized MoE refits through backend aliases

  #8416·Li-Jicheng opened Oct 28, 2026·3 tasks done·

  ·
  2

• [vllm, rollout] fix: reserve wildcard engine ports and close reservations on failure

  #8415·Li-Jicheng opened Oct 28, 2026·3 tasks done·

  ·
  2

• [model] perf: build Qwen3-VL/Qwen3.5 pos-embed interpolation tensors without Python lists

  #8414·Turoad opened Oct 28, 2026·7 tasks done·

  ·
  2

• [trainer] fix: preserve the token axis when packing position IDs

  #8413·yue-zeng-yue opened Oct 28, 2026Contributor·

  ·
  1

• [single_controller] feat: track unavailable placement group slots

  #8411·KunWuLuan opened Oct 28, 2026Contributor·5 of 6 tasks·

  ·
  1

  6

• [rollout, vllm, ray] feat: isolate standalone rollout Python runtime

  #8407·Levius-Fubuki opened Oct 28, 2026·6 of 7 tasks·

  ·
  6

• examples: add single-GPU 24GB GRPO + LoRA recipe (Qwen2.5-1.5B)

  #8405·Rich-Young98 opened Oct 27, 2026·

  ·
  8

• [hardware, megatron, doc] feat: upgrade Ascend 950 to Megatron-Core 0.18 and add MXFP8 GRPO example

  #8404·Rain6ow opened Oct 27, 2026·8 tasks·

  ·
  9

• [recipe] feat: DeepSeek-V4.1 LoRA with bitwise train/inference log-probs and 262K context on 8 GPUs

  #8403·ISEEKYAN opened Oct 27, 2026Collaborator·

  ·
  6