All pull requests

Pull requests

• Keep the tensors a model names for fp32 in fp32 in init_inference

  #8864·udsy19 opened Oct 20, 2026·

  ·
  1

  4

• Fix subgroup communication bandwidth calculation

  #8862·marcusacosta opened Oct 19, 2026·

  ·
  1

  3

• AutoTP: sequence sharding for training

  #8860·pengdurice opened Oct 19, 2026Contributor·

  ·

• AutoTP: shard an untied input embedding along the vocabulary (`vocab_…

  #8859·pengdurice opened Oct 19, 2026Contributor·

  ·

• [Muon] CPU offload: pin the momentum and overlap the host writeback

  #8858·jinyouzhi opened Oct 19, 2026Contributor·

  ·

• Keep Yuan's o_proj bias whole on every rank and describe it as replicated

  #8857·Achyuthan-S opened Oct 18, 2026Contributor·

  ·
  5

• Fix BF16 full-gradient updates to use the accumulation buffer

  #8856·mottopanikeiku opened Oct 18, 2026·

  ·
  4

• Expect AutoTP to keep a row-parallel output head tied to the embedding

  #8855·jinyouzhi opened Oct 18, 2026Contributor·

  ·
  2

• Scale ZeRO-3 __avg_scatter_grads by sequence_parallel_size

  #8854·deep1blue opened Oct 18, 2026·

  ·
  1

  1

• Measure first-call and post-update costs of continuous rollout CUDA Graphs

  #8853·nathon-lee opened Oct 18, 2026Contributor·

  ·

• Support Transformers 5.20 cache query offsets in DeepSpeedStaticCache

  #8852·nathon-lee opened Oct 18, 2026Contributor·

  ·

• [AutoTP] Support frozen AutoTP models and manage AutoTP mode per model

  #8851·jinyouzhi opened Oct 17, 2026Contributor·

  ·

• AutoEP + AutoTP folding: Universal Checkpointing under ZeRO-3

  #8848·pengdurice opened Oct 16, 2026Contributor·

  ·

• AutoEP: support combine_impl="fused_weighted_sum" with AutoEP+AutoTP folding

  #8847·pengdurice opened Oct 16, 2026Contributor·

  ·

• AutoTP: send CPU copies in the first-batch consistency check

  #8845·pengdurice opened Oct 16, 2026Contributor·

  ·

• Universal Checkpointing: read AutoEP ZeRO-3 optimizer files one at a time

  #8844·pengdurice opened Oct 16, 2026Contributor·

  ·

• Keep the parameters a model names for fp32 in fp32 under ZeRO-3

  #8843·pengdurice opened Oct 16, 2026Contributor·

  ·

• ZeRO-3: reuse the weight partition as the master copy when it already has the master dtype

  #8842·pengdurice opened Oct 16, 2026Contributor·

  ·

• Peng add cpu adam fused step v1

  #8841·pengdurice opened Oct 16, 2026Contributor·

  ·
  1

• ZeRO-3: copy offloaded master weights back to the device through pinned buffer

  #8840·pengdurice opened Oct 16, 2026Contributor·

  ·

• [OPSD] Request last-position logits only in HybridEngine rollout prefill

  #8839·jinyouzhi opened Oct 16, 2026Contributor·

  ·
  3

• Add Triton kernels for vocab-parallel logprobs, entropy and KL

  #8838·jinyouzhi opened Oct 16, 2026Contributor·

  ·
  3

• Fix lost TP all-reduce results for non-contiguous tensors

  #8837·jinyouzhi opened Oct 16, 2026Contributor·

  ·
  5

• Skip the accelerator-only CPU optimizer tests on the cpu accelerator

  #8836·delock opened Oct 16, 2026Collaborator·

  ·
  1

• Fix out-of-bounds row access in inference layer_norm/rms_norm kernels for hidden sizes <= 128

  #8830·jayden1711 opened Oct 15, 2026·

  ·
  1

  6