Issues

• should pytest --forked be supported for GPU unit test runs?

  deepspeedai/DeepSpeed#9075 · Filter by author hwchen2017 opened on Nov 13, 2026

• [BUG] save_checkpoint writes latest before other ranks finish their shards

  deepspeedai/DeepSpeed#9067 · Filter by author asonnino opened on Nov 12, 2026
  1

• [BUG] save_checkpoint after load_checkpoint(tag=<older>) rewrites the abandoned run's tag in place

  deepspeedai/DeepSpeed#9066 · Filter by author asonnino opened on Nov 12, 2026

• [REQUEST] Index group offsets once when materializing lazy ZeRO-3 checkpoint tensors

  deepspeedai/DeepSpeed#9062 · Filter by author itsrayeedscodespace opened on Nov 11, 2026

• Triton inference softmax(input, mask) reads the wrong mask row for a per-batch mask (wrong batch / out of bounds)

  deepspeedai/DeepSpeed#9060 · Filter by author cervoliu opened on Nov 11, 2026

• [BUG] ZeRO-1 with fp32 grad accumulation rounds the window sum to the communication dtype

  deepspeedai/DeepSpeed#9047 · Filter by author sfc-gh-truwase opened on Nov 10, 2026

• nightly failure without a failure class: cpu-torch-multi-latest died unclassified

  nightly-triage
  deepspeedai/DeepSpeed#9038 · Filter by author github-actions opened on Nov 9, 2026

• [RFC] Two-stage relay for cross-node weight sync, priced by a calibrated cost model

  deepspeedai/DeepSpeed#9026 · Filter by author Achyuthan-S opened on Nov 6, 2026
  16

• [BUG] AutoTP counts replicated GPTBigCode key/value rows once per tensor-parallel rank in the global gradient norm

  bugtraining
  deepspeedai/DeepSpeed#9024 · Filter by author VenusTZZ opened on Nov 6, 2026
  1

• AutoTP rollout engine: generate with a tensor-parallel copy of a ZeRO training model

  enhancement
  deepspeedai/DeepSpeed#9022 · Filter by author OkoroAchraf opened on Nov 5, 2026

• CPU: test_finite_step_after_overflow fp16 cases fail with 'Type fp16 is not supported on your device'

  deepspeedai/DeepSpeed#9019 · Filter by author delock opened on Nov 5, 2026
  2

• segKI graph-capture decode produces wrong tokens for row 0 at b=8 and all rows at b=2 (eager path is correct)

  bug
  deepspeedai/DeepSpeed#9017 · Filter by author delock opened on Nov 5, 2026
  1

• [BUG] zero_quantized_nontrainable_weights corrupts frozen weights and crashes ZeRO-3 fetches

  bugtraining
  deepspeedai/DeepSpeed#8990 · Filter by author VenusTZZ opened on Nov 3, 2026

• fp16 hysteresis: 0 (or negative) makes DynamicLossScaler never lower the loss scale, so every overflowing step is skipped forever

  deepspeedai/DeepSpeed#8982 · Filter by author damienen opened on Nov 2, 2026

• [BUG] AutoTP multi-query GPTBigCode training fails at TP > 1

  bugtraining
  deepspeedai/DeepSpeed#8961 · Filter by author VenusTZZ opened on Oct 31, 2026

• ZeRO-1/2 gradient accumulation: remaining test coverage

  buildtraining
  deepspeedai/DeepSpeed#8957 · Filter by author sfc-gh-truwase opened on Oct 30, 2026
  

• [BUG] Re-exporting a ZeRO checkpoint can leave a stale index that loads old weights

  deepspeedai/DeepSpeed#8952 · Filter by author ai-yang opened on Oct 30, 2026

• [BUG] OneCycle applies the minimum momentum to the first Adam update

  deepspeedai/DeepSpeed#8951 · Filter by author ai-yang opened on Oct 30, 2026

• [BUG] zero.Init casts tensors created with an explicit dtype=torch.float (RoPE inv_freq computed in bf16)

  bugtraining
  deepspeedai/DeepSpeed#8948 · Filter by author eilamc14 opened on Oct 29, 2026

• [RFC] Agentic RL training on the DeepSpeed hybrid engine

  deepspeedai/DeepSpeed#8942 · Filter by author anyadontfly opened on Oct 29, 2026

• [RFC] Batch replica selection for affine transfer plans

  deepspeedai/DeepSpeed#8935 · Filter by author 0z5a opened on Oct 28, 2026
  4

• [RFC] Deterministic training and versioned Megatron bitwise compatibility for production-common configurations

  deepspeedai/DeepSpeed#8915 · Filter by author ziming-zh opened on Oct 26, 2026

• [Precision] BF16 AutoEP folding rounds FP32 TP reduction before FP32 gradient accumulation

  deepspeedai/DeepSpeed#8913 · Filter by author ziming-zh opened on Oct 26, 2026

• [BUG] get_global_norm_of_tensors crashes on ranks with an empty norm contribution

  deepspeedai/DeepSpeed#8912 · Filter by author ziming-zh opened on Oct 26, 2026

• [BUG] ZeRO stage 1/2 marks optimizer-state fragments linked before later groups initialize

  deepspeedai/DeepSpeed#8911 · Filter by author ziming-zh opened on Oct 26, 2026
  1