Repository Issue Activity (beta)

#8290 HF transformers main injects 'embedding_rowwise' into tp_plan for tied-embedding models; AutoTP rejects the whole plan

#8260 [BUG] reduce_bucket_size=0 passes validation but crashes ZeRO backward with ZeroDivisionError

#7413 [BUG] FlopsProfiler accumulates metrics when called multiple times

#8156 compute_elastic_config raises bare ZeroDivisionError for return_microbatch=True on non-0.2 elasticity versions without explicit world_size

#8297 [BUG] ZeRO-1/2 fail mapping zero-sized parameter inside a partition

#8197 [REQUEST] OPSD Profile and improve HybridEngine rollout performance

#8279 [BUG] ZeRO-1/2 fail on trailing zero-sized trainable parameters

#8291 [Ulysses SP] process-wide _ulysses_num_kv_heads global breaks a second model with a different head count in the same process

#8285 [AutoTP] update_mp_params shards vision-tower attributes (patch_embed.embed_dim) without sharding weights, breaking multimodal forward

#8193 [REQUEST] Dynamic/Adaptive Prefetching Window for ZeRO-3

#8173 [RFC] Support sharding LM heads and adopting Online Softmax

#102 max_grad_norm is ignored in FP16 training

#7775 [BUG] ZeRO-3 with `zero_quantized_weights=true` incorrectly casts bf16 inputs to fp16, causing BERT training failure

#8276 [BUG] DeepSeek-V3 MLA rank changes trigger 16-byte data_ptr alignment failure in grouped_mm under DeepSpeed

#8263 Hybrid Engine registers partial inference policies for unsupported Qwen architectures

#8262 ZeRO-3 OPSD rollout deadlocks when data-parallel ranks finish generation at different lengths

#8076 [BUG]ds_z3_config error

#8249 IndexError in is_nfs_path() when df wraps a long device name (breaks `import deepspeed`)

#8252 Define affine IR for checkpointing

#8230 [RFC] Universal checkpoint: non-semantic per-parameter geometric shard map

#8248 [BUG]DeepSpeed latest version fails during NCCL process-group initialization with CUDA error 600, while 0.17.2 works

#7384 [BUG] Ulysses DistributedAttention silently produces incorrect output when #GPUs does not divide global sequence length

#8231 Cleanup: remove tp_shard process-wide scalar globals (num_kv_heads / tp_grain_size / ...), thread explicitly

#7433 [BUG] TypeError: 'staticmethod' object is not callable, in deepcompile (patch_compiled_func.py)

#8224 ZeRO-2 + bf16 silently computes incorrect gradients at multi-rank when a submodule is used more than once per step, a regression from #7665, still present in 0.19.3