All pull requests

Pull requests

• Check that converting through an affine map writes every element once

  #8975·Achyuthan-S opened Nov 1, 2026Contributor·

  ·
  2

• Keep fp32 CPU optimizer states in fp32 when loading a state_dict

  #8974·xylian86 opened Nov 1, 2026Collaborator·

  ·
  3

• ZeRO-3 partial offload: fix LR schedules and checkpoints of the device sub-groups

  #8973·xylian86 opened Nov 1, 2026Collaborator·

  ·
  4

• SuperOffload: checkpoint the CPU Adam state, fix unused parameters and rollbacks, speculate with clipping

  #8972·xylian86 opened Nov 1, 2026Collaborator·

  ·
  5

• Log Train/Samples/train_loss to monitors again

  #8971·arun-elr opened Nov 1, 2026·

  ·
  5

• ZeRO: apply the clip coefficient at full precision to bf16/fp16 gradients

  #8970·sfc-gh-abkulkarni opened Nov 1, 2026Contributor·3 tasks done·

  ·
  3

• Skip blank and comment lines in .deepspeed_env

  #8969·vineethsaivs opened Oct 31, 2026Contributor·

  ·

• Skip redundant Stage-0 gradient reduction at DP=1

  #8968·jinyouzhi opened Oct 31, 2026Contributor·

  ·

• [AffineIR] Analyze affine transfer batches before execution (Plan analysis & calibration Part-1)

  #8967·jinyouzhi opened Oct 31, 2026Contributor·

  ·

• [AffineIR] Calibrate affine transfer links per path (Plan analysis & calibration Part-2)

  #8966·jinyouzhi opened Oct 31, 2026Contributor·

  ·
  3

• ZeRO-1/2: keep every request's gradients until the optimizer step

  #8964·sfc-gh-truwase opened Oct 31, 2026Collaborator·

  ·
  4

• [ZeRO] Keep the offload overflow flag on the device

  #8963·0z5a opened Oct 31, 2026Contributor·

  ·

• Fix AutoTP multi-query GPTBigCode training at TP > 1

  #8962·VenusTZZ opened Oct 31, 2026Contributor·

  ·
  1

• Divide SequenceTiledCompute's incoming gradient out of place

  #8960·vineethsaivs opened Oct 30, 2026Contributor·

  ·

• Honor explicit dtype in ZeRO tensor constructors

  #8959·jinyouzhi opened Oct 30, 2026Contributor·

  ·
  1

  6

• [VocabTP] Add vocabulary-parallel generalized JSD with Triton kernels

  #8958·jinyouzhi opened Oct 30, 2026Contributor·

  ·
  1

• Fix ZeRO-1/2 lazy optimizer-state linking for all parameter groups

  #8956·YidaWeng opened Oct 30, 2026·

  ·
  1

  8

• Retire stale checkpoint artifacts after replacing a ZeRO export

  #8954·ai-yang opened Oct 30, 2026·

  ·
  1

  1

• Initialize OneCycle momentum at the upper bound

  #8953·ai-yang opened Oct 30, 2026·

  ·
  1

  1

• Fix torch-latest workflow unit failures

  #8950·sfc-gh-truwase opened Oct 30, 2026Collaborator·2 of 6 tasks·

  ·
  6

• ZeRO-1/2: preserve gradients across request boundaries until step

  #8949·stas00 opened Oct 29, 2026Collaborator·

  ·
  8

• Sum the OPSD JSD loss in float32 for bf16 and fp16 logits

  #8947·vineethsaivs opened Oct 29, 2026Contributor·

  ·

• Merge final_linear shards when loading a Megatron checkpoint on fewer TP ranks

  #8946·udsy19 opened Oct 29, 2026·

  ·
  1

• [VocabTP] vocab_parallel_ce support fused LM-head projection + loss for the HF forward

  will-review
  #8945·jinyouzhi opened Oct 29, 2026Contributor·

  ·
  15

• Run the v1 multi-rank CPU suite on pull requests

  #8944·delock opened Oct 29, 2026Collaborator·

  ·
  6