All pull requests

Pull requests

• feat(rollout): add OPSD response token loss primitives

  #8677·nathon-lee opened Sep 26, 2026Contributor·

  ·

• fix(rollout): forward top-k sampling configuration

  #8676·nathon-lee opened Sep 26, 2026Contributor·

  ·

• Check AutoSP sequence length and head count divisibility

  #8675·jinyouzhi opened Sep 26, 2026Contributor·

  ·

• Validate the AutoSP SP/DP mesh when registering process groups

  #8674·jinyouzhi opened Sep 26, 2026Contributor·

  ·

• Fuse mixed-precision gradients and parameter writeback into FusedAdam

  #8673·CharlesXu-HQ opened Sep 26, 2026·

  ·

• AutoEP: reorder expert rows with gathers on the collective path

  #8671·yh0903 opened Sep 26, 2026Contributor·

  ·

• AutoEP + AutoTP folding under ZeRO-3

  #8670·pengdurice opened Sep 25, 2026Contributor·

  ·

• Apply weight decay to the CPU Adagrad update

  #8669·vineethsaivs opened Sep 25, 2026Contributor·

  ·

• Apply the ZeRO-3 predivide postscale when the factor equals world size

  #8668·vineethsaivs opened Sep 25, 2026Contributor·

  ·

• Skip ZeRO DeepCompile backward hooks for AutoSP

  #8667·jinyouzhi opened Sep 25, 2026Contributor·

  ·

• fix: attention_unfused alpha missing norm_factor squaring

  will-review
  #8666·delock opened Sep 25, 2026Collaborator·

  ·
  3

• Fix FPDT K/V gradient buffers under grouped-query attention

  will-review
  #8665·jinyouzhi opened Sep 25, 2026Contributor·

  ·

• Fused RMSNorm: multi-row programs and a one-pass backward for narrow norms (stacked on #8645)

  #8664·yh0903 opened Sep 25, 2026Contributor·

  ·

• Add an opt-in fused Triton RoPE for Hugging Face split-half models

  #8663·yh0903 opened Sep 25, 2026Contributor·

  ·

• Fix FP16 gradient unscaling on loss scale growth steps

  #8662·taking-lying-flat opened Sep 25, 2026Contributor·

  ·

• Record global grad norm when the engine clips fp32 gradients

  #8661·nniiovoo opened Sep 25, 2026·

  ·
  2

• Add deepspeed.models.glm_moe_dsa: TileLang sparse attention for GLM-5.2

  #8659·pengdurice opened Sep 24, 2026Contributor·

  ·

• Add deepspeed.models.minimax_m3_vl: block-sparse attention for MiniMax M3

  #8658·pengdurice opened Sep 24, 2026Contributor·

  ·

• Fix autotuning tuner error log that dropped the exception type

  #8657·David-Wu1119 opened Sep 24, 2026·

  ·
  1

• Reject Adam settings SuperOffload silently replaces with AdamW

  #8656·vineethsaivs opened Sep 24, 2026Contributor·

  ·

• Keep Muon's update at full size under fp16 loss scaling

  will-review
  #8655·alanhuangyoo opened Sep 24, 2026Contributor·

  ·

• Release each BF16 gradient once immediate_grad_update has accumulated it

  will-review
  #8654·yh0903 opened Sep 24, 2026Contributor·

  ·

• Add an opt-in chunked causal-LM loss that keeps no full-vocabulary FP32 tensor

  will-review
  #8653·yh0903 opened Sep 24, 2026Contributor·

  ·

• AutoEP: fused Triton kernel for the clamped expert activations

  #8652·pengdurice opened Sep 24, 2026Contributor·

  ·

• AutoEP: MiniMax-M3 preset

  #8651·pengdurice opened Sep 24, 2026Contributor·

  ·