Search results

DataLoader shuffle sequence replays from epoch 0 after resuming from a checkpoint

Status: Open.

#3996 In huggingface/accelerate;

· XiangY-Q opened on Apr 5, 2026

device_map="auto": silent corruption of tensors captured in register_forward_hook (3+ GPUs, inference_mode, stale bare references)

Status: Open.

#3995 In huggingface/accelerate;

· keenanpepper opened on Apr 3, 2026

Isssue when using torch.compile

Status: Open.

#3991 In huggingface/accelerate;

· tritam593 opened on Mar 28, 2026

[Feature] Save model-only feature in save_state

Status: Open.

#3990 In huggingface/accelerate;

· 3manifold opened on Mar 27, 2026

[FSDP2] auto-exclude incompatible Params4bit from fully_shard to prevent silent QLoRA corruption

Status: Open.

#3983 In huggingface/accelerate;

· roycho96 opened on Mar 19, 2026

[Bug] FSDP2 mixed-precision upcast to fp32 is a silent no-op since v1.13.0

Status: Open.

#3979 In huggingface/accelerate;

· iavinas opened on Mar 18, 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] 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