accelerate · Issues· 107 open

`simple_launcher` discards the child's error, unlike the elastic launch path

all_gather_into_tensor is deprecated, _gpu_gather warns on every gather

Skip automatic DataLoader sharding when the loader is already rank-sharded

`PrefixedDataset` violates the `Mapping` contract

compile_regions runs the uncompiled model when the root module has an accelerate hook

prepare_model unconditionally imports DTensor, breaking single-GPU training on torch builds without a distributed backend

[Bug] Fused AdamW AMP overflow leaves step_was_skipped false and advances the scheduler

ACCELERATE_BYPASS_DEVICE_MAP is the only boolean env flag read case-sensitively, so =True is silently ignored

`use_stateful_dataloader` under multi-process training restores the cursor, not the shuffle order — sampler permutation is never serialized

[Community Contributions] examples on distributed inference using Accelerate

`use_stateful_dataloader` + `num_workers>0` draws epoch-0 permutation during `prepare()` before cross-rank RNG sync — corrupts DDP data partition

Feature request - SLURM support

accelerator.prepare

`datasets` Iterable Dataset sharding support

how to cleanly exit when using `accelerate launch`