Search results

FSDP2 + SHARDED_STATE_DICT: optimizer checkpoint load fails with missing step key. (RuntimeError: Missing key in checkpoint state_dict: optimizer.state.0.step.)

Status: Open.

#3971 In huggingface/accelerate;

· iavinas opened on Mar 11, 2026

loading weight is so slow for fsdp2

Status: Open.

#3968 In huggingface/accelerate;

· jiaqiw09 opened on Mar 10, 2026

[Bug] prepare() silently double-wraps models and double-serializes checkpoints when called twice on the same object

Status: Open.

#3967 In huggingface/accelerate;

· iavinas opened on Mar 9, 2026

Security: Unsafe torch.load in FSDP2 scaler path (CWE-502)

Status: Open.

#3964 In huggingface/accelerate;

· spartan8806 opened on Mar 9, 2026

[Bug] TrackioTracker.log() ignores step parameter

Status: Open.

#3963 In huggingface/accelerate;

· iavinas opened on Mar 7, 2026

[Bug] RNG states from multiple backends (e.g. CUDA + HPU) are saved but only one is restored on load_state

Status: Open.

#3960 In huggingface/accelerate;

· iavinas opened on Mar 6, 2026

infer_auto_device_map does not place submodule buffers on device_map when submodule is split

Status: Open.

#3956 In huggingface/accelerate;

· dg845 opened on Mar 5, 2026

AssertionError for FP8 Benchmarks

Status: Open.

#3954 In huggingface/accelerate;

· Hukongtao opened on Mar 4, 2026

Segfault in device_map='auto' weight dispatch on RTX 5090 D v2 (Blackwell, sm_120, GB202-240)

Status: Open.

#3933 In huggingface/accelerate;

· XHXIAIEIN opened on Feb 24, 2026

why does (Naive MP) show better memory efficiency than FSDP1/2 ??

Status: Open.

#3932 In huggingface/accelerate;

· Lim-Sung-Jun opened on Feb 24, 2026

_prepare_fsdp2 does not work with nn.Module on meta device

Status: Open.

#3930 In huggingface/accelerate;

· jiayuanmark opened on Feb 21, 2026

notebook_launcher raises "Cannot re-initialize CUDA in forked subprocess"

Status: Open.

#3925 In huggingface/accelerate;

· minhtriet opened on Feb 10, 2026