Issues

• CVE-2026-69112 / GHSA-4j2p-28q2-5m79: advisory says <= 1.14.0, but 1.15.0 has the same code. Please clarify the project's position.

  huggingface/accelerate#4525 · Filter by author mohamedasni opened on Nov 11, 2026

• merge_fsdp_weights with remove_checkpoint_dir=True deletes the merged model when output_path is inside checkpoint_dir

  huggingface/accelerate#4514 · Filter by author Dhruv-Sharma29 opened on Nov 9, 2026

• FSDP2: prepare() keeps the whole unsharded model alive during sharding when cpu_ram_efficient_loading=False

  huggingface/accelerate#4511 · Filter by author Jayce-Ping opened on Nov 8, 2026

• FSDP2 QLoRA: with bnb_4bit_quant_storage=bf16 (the storage the #3987 warning recommends), the all-gather alters packed 4-bit weights at world size ≥ 2

  huggingface/accelerate#4510 · Filter by author CyberNativeAI opened on Nov 8, 2026
  1

• BatchSamplerShard padding mutates retained sampler batches and changes original loader samples

  huggingface/accelerate#4505 · Filter by author davrens101 opened on Nov 6, 2026
  1

• Checkpoints saved during gradient accumulation lose partial gradients on resume

  huggingface/accelerate#4503 · Filter by author ely2ba opened on Nov 6, 2026

• enable_model_cpu_offload(): first post-load generation thrashes the Windows pagefile (~6x step time), silently recovers

  huggingface/accelerate#4488 · Filter by author AIwork4me opened on Nov 1, 2026

• Stateful DataLoader resume at an epoch boundary replays the completed epoch (shard iteration counter is not checkpointed)

  huggingface/accelerate#4480 · Filter by author anishpentyala opened on Oct 31, 2026
  1

• Sequence parallelism: allow an opt-out of the linear-attention refusal, and make disable_in_eval configurable

  huggingface/accelerate#4476 · Filter by author eilamc14 opened on Oct 30, 2026

• FSDP2 + cpu_ram_efficient_loading: ignored_modules are sharded anyway, and models with persistent buffers crash

  huggingface/accelerate#4474 · Filter by author eilamc14 opened on Oct 30, 2026

• accelerator.reduce(tensor, "mean") fails on integer tensors when more than one process is used

  huggingface/accelerate#4469 · Filter by author MaiAditya opened on Oct 29, 2026

• get_max_memory() budgets CPU offload from host RAM, not the container memory limit

  huggingface/accelerate#4465 · Filter by author prabhkodes opened on Oct 27, 2026

• save_state(total_limit=...) deletes the newest checkpoint and load_state() resumes from an old one on Windows

  huggingface/accelerate#4454 · Filter by author Sonica-B opened on Oct 25, 2026

• Stateful split-batch dispatcher replays consumed samples after restoring

  huggingface/accelerate#4452 · Filter by author AryanRG opened on Oct 25, 2026

• Optimizer exceptions leave a stale AMP step marker and advance the scheduler after overflow

  huggingface/accelerate#4445 · Filter by author salma2vec opened on Oct 23, 2026

• Failed dataloader iteration leaves stale gradient state

  huggingface/accelerate#4443 · Filter by author ska2704 opened on Oct 23, 2026

• save_state deletes retained checkpoints before a replacement save succeeds

  huggingface/accelerate#4436 · Filter by author salma2vec opened on Oct 21, 2026

• load_checkpoint_and_dispatch ignores dtype with default device_map=None

  huggingface/accelerate#4434 · Filter by author akshitkhokhani opened on Oct 20, 2026
  3

• preserve_rng_stage=False in fsdp2_apply_ac causes bug when using dropout with activation checkpointing

  huggingface/accelerate#4430 · Filter by author YisuZhou opened on Oct 20, 2026
  2

• FSDP2 forward does not run under mixed precision autocast, unlike DDP and FSDP1

  huggingface/accelerate#4419 · Filter by author 12kimih opened on Oct 18, 2026

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  huggingface/accelerate#4411 · Filter by author michellearoth opened on Oct 16, 2026

• gather_for_metrics drops or duplicates examples with SubsetRandomSampler

  huggingface/accelerate#4410 · Filter by author gorkadabo1 opened on Oct 15, 2026
  3

• test2

  huggingface/accelerate#4409 · Filter by author Asthenia0412 opened on Oct 15, 2026

• pad_across_processes turns bool tensors into int64, which breaks gather across ranks

  huggingface/accelerate#4399 · Filter by author arthurdjn opened on Oct 13, 2026

• send_to_device ignores skip_keys for Mappings that have .to() method

  huggingface/accelerate#4397 · Filter by author arthurdjn opened on Oct 13, 2026