All pull requests

Pull requests

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

  #8944·delock opened Oct 29, 2026Collaborator·

  ·
  6

• [AffineIR] Storage maps of AutoEP experts under EDP and ZeRO

  #8941·jinyouzhi opened Oct 29, 2026Contributor·

  ·
  4

• Fix legacy MoE dispatch capacity agreement across expert-parallel ranks

  #8940·Levius-Fubuki opened Oct 29, 2026·

  ·
  1

• feat(rollout): non-greedy sampling on hybrid engine graph decode path

  #8939·anyadontfly opened Oct 28, 2026·

  ·
  1

  5

• [ops] Warn when DeepSpeedCPUAdam is limited to one OpenMP thread

  #8938·0z5a opened Oct 28, 2026Contributor·

  ·

• [ROCm] Fix the ROCm build of inference_core_ops

  #8937·rraminen opened Oct 28, 2026Contributor·

  ·
  1

• [checkpoint] Batch replica selection for affine transfer plans

  #8936·0z5a opened Oct 28, 2026Contributor·

  ·

• [Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators

  #8933·VenusTZZ opened Oct 28, 2026Contributor·

  ·
  1

  3

• [Fix] Reject duplicate block ids in BlockedAllocator.free to prevent KV-cache double allocation

  #8931·lkh-bug opened Oct 28, 2026·

  ·
  3

• Cap CPU accelerator memory at the cgroup limit of the process

  #8930·prabhkodes opened Oct 27, 2026·

  ·
  1

  14

• Sum GPTBigCode's shared key/value gradient across tensor-parallel ranks

  #8928·Achyuthan-S opened Oct 27, 2026Contributor·

  ·
  1

  5

• Add klxpu (KUNLUNXIN XPU) self-hosted CI for the v1 unit suite

  #8923·GuCarpenter opened Oct 27, 2026Contributor·

  ·
  8

• Check where column-split pieces write in a shard

  #8922·Achyuthan-S opened Oct 27, 2026Contributor·Queued
  8

• Preserve FP32 folded gradients through FP32 accumulation

  #8920·Levius-Fubuki opened Oct 27, 2026·

  ·
  1

• Avoid optimizer updates during low-precision initialization

  #8919·Levius-Fubuki opened Oct 27, 2026·

  ·
  1

  4

• Keep empty gradient-norm ranks in collective reductions

  #8918·Levius-Fubuki opened Oct 27, 2026·

  ·
  1

  6

• Add an opt-in deterministic execution policy

  #8917·ziming-zh opened Oct 26, 2026·

  ·

• Fix the default data parallel checkpoint writer crash on dp_rank

  #8916·vineethsaivs opened Oct 26, 2026Contributor·

  ·

• Fix ZeRO-3 NVMe module-only checkpoint weight refresh

  #8908·Levius-Fubuki opened Oct 26, 2026·

  ·
  6

• Emit intra_op_parallelism and JSON booleans from ds_nvme_tune

  #8907·Sonica-B opened Oct 26, 2026·

  ·
  1

• fix: inference norm missing threadsPerGroup==32 dispatch + ROCm 7.2.2 bf16 ABI

  #8906·umechand-amd opened Oct 25, 2026Collaborator·

  ·
  5

• [Feature] Extend affine live-state transfer to CPU-offloaded optimizer storage

  #8905·ArjunPakhan opened Oct 25, 2026·

  ·
  1

  9

• Track implicit package initializer dependencies in CI test selection

  #8903·PointerL opened Oct 25, 2026·

  ·
  1

• Recover legacy Torch checkpoints in zero_to_fp32

  #8901·Levius-Fubuki opened Oct 24, 2026·

  ·

• Complete ZeRO-3 serial checkpoint handoff for universal loads and rank boundaries

  #8898·Levius-Fubuki opened Oct 24, 2026·

  ·
  7