deepspeedai

Issues somebody may already be working on are listed last. List everything by date

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8899 · Opened Oct 24, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Support RMS-matched Muon updates via adjust_lr_fn configuration

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 54/100

  deepspeedai/DeepSpeed#8865 · 2 comments · Opened Oct 20, 2026

  Maintainers usually reply within 3 days

• Clarify safe cross-rank aggregation for MoE forward timing breakdown

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8863 · Opened Oct 19, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Add an OPSD training batch builder for teacher-prefix assembly and response-logit alignment

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  deepspeedai/DeepSpeed#8850 · 1 comment · 1 reaction · Opened Oct 17, 2026

  Maintainers usually reply within 3 days

• nightly timeout: cpu-torch-multi-latest multi-rank suite exceeded its time budget

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  nightly-triage

  Difficulty 3/5 1-2 days Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8834 · 4 comments · Opened Oct 16, 2026

  Maintainers usually reply within 3 days

• [RFC] Skip collectives on absent or single-rank groups with deepspeed.comm.pg_utils.is_parallel

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8814 · 2 comments · Opened Oct 13, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Make ZeRO-1 custom-optimizer tensor flattening explicit or expose tensor metadata

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8812 · Opened Oct 13, 2026

  Maintainers usually reply within 3 days

• [RFC] Vocabulary-parallel per-token ops for RL and on-policy distillation

  May be free again @jinyouzhi claimed this 15 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 15/100

  deepspeedai/DeepSpeed#8790 · 4 comments · Opened Oct 9, 2026

  Maintainers usually reply within 3 days

• [Triton] `is_triton_supported()` semantics for non-CUDA Triton backends

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8789 · 8 comments · Opened Oct 9, 2026

  Maintainers usually reply within 3 days

• [RFC] Backend-pluggable low-precision training for Dense Linear and AutoEP grouped GEMM (Ascend HiFloat8)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 6/100

  deepspeedai/DeepSpeed#8777 · 3 comments · Opened Oct 8, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Legacy MoE gate: opt-in route_norm, with historical combine weights left as the default

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8768 · 1 comment · Opened Oct 6, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Inference V2: add one model family on the existing kernels (dense Qwen3 proposed)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8762 · Opened Oct 6, 2026

  Maintainers usually reply within 3 days

• {{ env.GITHUB_WORKFLOW }} CI test failure

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  ci-failure

  Difficulty 4/5 3-5 days Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8741 · Opened Oct 4, 2026

  Maintainers usually reply within 3 days

• [REQUEST] DeepCompile ZeRO-1: reduce-to-owner instead of AllReduce

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8700 · 2 comments · Opened Sep 29, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Add pipeline section in `config.json` for pp

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 5/5 Over a week Newbie friendliness 30/100

  deepspeedai/DeepSpeed#8696 · Opened Sep 28, 2026

  Maintainers usually reply within 3 days

• DeepSpeed - Q4 2026 Roadmap

  May be free again @delock claimed this 29 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  roadmap

  deepspeedai/DeepSpeed#8672 · 14 comments · 3 reactions · 5 assignees · Opened Sep 26, 2026

  Maintainers usually reply within 3 days

• init_inference rejects dtype=torch.int8 on every accelerator

  May be free again A pull request for this issue was closed without being merged. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 78/100

  deepspeedai/DeepSpeed#8577 · Opened Sep 18, 2026

  Maintainers usually reply within 3 days

• Deprecate unused DeepSpeed features

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 20/100

  deepspeedai/DeepSpeed#8489 · 7 comments · Opened Sep 11, 2026

  Maintainers usually reply within 3 days

• [RFC] Enable Correct and High-Performance Muon Optimizer under ZeRO CPU Offload (Stage 1/2/3)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8463 · 1 reaction · Opened Sep 8, 2026

  Maintainers usually reply within 3 days

• Default gradient_clipping divides every Muon update by its own norm, shrinking the step by a model-sized factor

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8439 · 2 comments · Opened Sep 6, 2026

  Maintainers usually reply within 3 days

• Muon under AutoTP orthogonalizes each shard, so the update depends on the TP degree

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8437 · 1 comment · Opened Sep 6, 2026

  Maintainers usually reply within 3 days

• [Feature] Per-Head Muon: per-head Newton-Schulz orthogonalization for attention projections

  May be free again @0z5a claimed this 37 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement training

  Difficulty 5/5 Over a week Newbie friendliness 45/100

  deepspeedai/DeepSpeed#8367 · 5 comments · 4 reactions · Opened Aug 30, 2026

  Maintainers usually reply within 3 days

• [BUG]Unfreed parameters and unnecessary peak memory in autoEP

  May be free again @tohtana claimed this 56 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  deepspeedai/DeepSpeed#8353 · 1 comment · 1 assignee · Opened Aug 29, 2026

  Maintainers usually reply within 3 days

• [REQUEST] XPU: register native pinned host memory with the device runtime (follow-up to #8283)

  May be free again @delock claimed this 60 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  deepspeedai/DeepSpeed#8315 · 1 comment · 2 reactions · 1 assignee · Opened Aug 25, 2026

  Maintainers usually reply within 3 days

• HF transformers main injects 'embedding_rowwise' into tp_plan for tied-embedding models; AutoTP rejects the whole plan

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8290 · 3 comments · 1 reaction · Opened Aug 22, 2026

  Maintainers usually reply within 3 days

• Add logging support for better observability

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8140 · 1 comment · Opened Jul 14, 2026

  Maintainers usually reply within 3 days

• DeepSpeed - Q3 Roadmap

  May be free again @delock claimed this 105 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  roadmap

  deepspeedai/DeepSpeed#8104 · 6 comments · 9 reactions · 5 assignees · Opened Jun 30, 2026

  Maintainers usually reply within 3 days

• Exception type too broad in error handling

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 42/100

  deepspeedai/DeepSpeed#8093 · 1 comment · Opened Jun 26, 2026

  Maintainers usually reply within 3 days

• zero.Init partitioning of large fused MoE-expert tensors spikes a single GPU (transient full materialization) -> OOM during load even when the sharded model fits

  May be free again A pull request for this issue was closed without being merged. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 45/100

  deepspeedai/DeepSpeed#8085 · 4 comments · Opened Jun 22, 2026

  Maintainers usually reply within 3 days

• Add type hints to public API functions

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#8074 · 1 comment · Opened Jun 17, 2026

  Maintainers usually reply within 3 days

• EvoformerAttention should auto-detect CUTLASS instead of requiring CUTLASS_PATH

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  deepspeedai/DeepSpeed#8001 · 1 comment · Opened May 8, 2026

  Maintainers usually reply within 3 days

• [BUG]Loss connection to ranks

  May be free again @Shreyas-S-809 claimed this 171 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  Difficulty 4/5 3-5 days Newbie friendliness 45/100

  deepspeedai/DeepSpeed#7991 · 4 comments · Opened May 4, 2026

  Maintainers usually reply within 3 days

• [BUG] crashed training gpt model using 3d parallelization

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  deepspeedai/DeepSpeed#7922 · 2 comments · Opened Mar 25, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Put NVTX instrumentation in its own domain (and potentially categories)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  deepspeedai/DeepSpeed#7912 · 3 comments · Opened Mar 18, 2026

  Maintainers usually reply within 3 days

• Deepspeed zero 3 seems to save empty adapter_model.safetensors

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 28/100

  deepspeedai/DeepSpeed#7900 · 1 comment · Opened Mar 13, 2026

  Maintainers usually reply within 3 days

• [RFC] SDMA-Based Collective Communication Backend for ZeRO-3 on AMD GPUs

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  deepspeedai/DeepSpeed#7884 · 10 comments · 4 reactions · Opened Mar 4, 2026

  Maintainers usually reply within 3 days

• [Roadmap] DeepSpeed Roadmap Q2 2026

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#7861 · 4 comments · 5 reactions · Opened Feb 19, 2026

  Maintainers usually reply within 3 days

• [Bug] Missing input validation for fp16 loss_scale accepts float('inf')

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  Difficulty 2/5 1-3 hours Newbie friendliness 55/100

  deepspeedai/DeepSpeed#7852 · Opened Feb 13, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Ulysses support for VLMs

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#7834 · 1 comment · Opened Feb 6, 2026

  Maintainers usually reply within 3 days

• [BUG] DeepSpeed ZeRO Stage-3 + CPU offloaded optimizer (CPUAdam) inconsistency metadata between subgroup

  May be free again @tohtana claimed this 249 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  deepspeedai/DeepSpeed#7819 · 3 comments · 1 assignee · Opened Jan 27, 2026

  Maintainers usually reply within 3 days

• [BUG] ZenFlow Stage 3 with `full_warm_up_rounds=0` fails due to missing `complete_column_offset` attribute

  May be free again @Antlera claimed this 278 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  deepspeedai/DeepSpeed#7796 · 1 assignee · Opened Jan 20, 2026

  Maintainers usually reply within 3 days

• [REQUEST] Python types?

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  enhancement

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#7772 · Opened Jan 11, 2026

  Maintainers usually reply within 3 days

• [BUG] ZenFlow: loss values becomes NaN after 'update_interval' number of steps.

  May be free again @Antlera claimed this 293 days ago, and no pull request is open. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  deepspeedai/DeepSpeed#7759 · 3 comments · 1 assignee · Opened Jan 4, 2026

  Maintainers usually reply within 3 days

• [BUG] Grad norm value differs significantly across DeepSpeed version

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug training

  Difficulty 4/5 3-5 days Newbie friendliness 42/100

  deepspeedai/DeepSpeed#7749 · 4 comments · Opened Dec 29, 2025

  Maintainers usually reply within 3 days

• Please publish DeeepSpeed Python 3.13 Windows wheels already

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  ci-failure

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  deepspeedai/DeepSpeed#7744 · 2 comments · Opened Dec 23, 2025

  Maintainers usually reply within 3 days

• Follow-ups for restoring fp32 masters without optimizer state

  Possibly taken @tjruwase claimed this 2 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  deepspeedai/DeepSpeed#8890 · 1 assignee · Opened Oct 22, 2026

  Maintainers usually reply within 3 days

• [BUG] Universal checkpoint resume puts CPU `step` into fused/capturable optimizer state

  Possibly taken @ArjunPakhan claimed this 2 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 74/100

  deepspeedai/DeepSpeed#8884 · 1 comment · Opened Oct 22, 2026

  Maintainers usually reply within 3 days

• [BUG] cpu-multi-triage "report multi-rank test failures" cannot parse ANSI-colored FAILED lines

  Possibly taken @delock claimed this 3 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8881 · Opened Oct 22, 2026

  Maintainers usually reply within 3 days

• [BUG] ZeRO-3 resume can fail when a resident adapter becomes an external dependency after tracing

  Possibly taken @DimensionSTP claimed this 4 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8874 · Opened Oct 21, 2026

  Maintainers usually reply within 3 days

• Incorrect bus bandwidth calculation for subgroup collectives

  Possibly taken @marcusacosta claimed this 5 days ago. Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 1-2 days Newbie friendliness 25/100

  deepspeedai/DeepSpeed#8861 · Opened Oct 19, 2026

  Maintainers usually reply within 3 days