All issues

Search results

HybridEngineRollout: synchronized early stopping across ZeRO-3 data-parallel ranks

Status: Open.

#8321 In deepspeedai/DeepSpeed;

· delock opened on Aug 26, 2026

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

Status: Open.

#8315 In deepspeedai/DeepSpeed;

· delock opened on Aug 25, 2026

[Ulysses SP] process-wide _ulysses_num_kv_heads global breaks a second model with a different head count in the same process

Status: Open.

#8291 In deepspeedai/DeepSpeed;

· delock opened on Aug 22, 2026

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

Status: Open.

#8290 In deepspeedai/DeepSpeed;

· delock opened on Aug 22, 2026

[BUG] DeepSeek-V3 MLA rank changes trigger 16-byte data_ptr alignment failure in grouped_mm under DeepSpeed

Status: Open.

#8276 In deepspeedai/DeepSpeed;

· fwerkor opened on Aug 19, 2026

Hybrid Engine registers partial inference policies for unsupported Qwen architectures

Status: Open.

#8263 In deepspeedai/DeepSpeed;

· LiRunGuo opened on Aug 17, 2026

ZeRO-3 OPSD rollout deadlocks when data-parallel ranks finish generation at different lengths

Status: Open.

#8262 In deepspeedai/DeepSpeed;

· LiRunGuo opened on Aug 17, 2026

Define affine IR for checkpointing

enhancementNew feature or requestNew feature or request

Status: Open.

#8252 In deepspeedai/DeepSpeed;

· delock opened on Aug 14, 2026

IndexError in is_nfs_path() when df wraps a long device name (breaks import deepspeed)

Status: Open.

#8249 In deepspeedai/DeepSpeed;

· avnerv opened on Aug 13, 2026

[BUG]DeepSpeed latest version fails during NCCL process-group initialization with CUDA error 600, while 0.17.2 works

bugSomething isn't workingSomething isn't working

training

Status: Open.

#8248 In deepspeedai/DeepSpeed;

· janelu9 opened on Aug 13, 2026

Cleanup: remove tp_shard process-wide scalar globals (num_kv_heads / tp_grain_size / ...), thread explicitly

Status: Open.

#8231 In deepspeedai/DeepSpeed;

· delock opened on Aug 7, 2026

[RFC] Universal checkpoint: non-semantic per-parameter geometric shard map

Status: Open.

#8230 In deepspeedai/DeepSpeed;

· delock opened on Aug 7, 2026