Add disable_mmap kwarg to from_pretrained with hf-mount auto-detection

[rtrompier]

What

Adds a new disable_mmap kwarg to PreTrainedModel.from_pretrained (and plumbs it through load_state_dict / LoadStateDictConfig / _load_pretrained_model). When enabled, safetensors checkpoints are read fully into memory and parsed via safetensors.torch.load(bytes) instead of being memory-mapped with safe_open.

Why

• Loading a safetensors model from a FUSE filesystem can deadlock the Python process when mmap + kernel readahead + the parallel state-dict loader trigger many concurrent page-faults. We observed this in production on HF Spaces/Endpoints, which expose Hub repos via a FUSE mount named hf-mount: captured kernel stacks showed Python hanging on page-faults with a saturated FUSE queue.
• Bypassing mmap (reading the file straight into RAM) side-steps the deadlock with a modest RAM cost.
• diffusers already ships the exact same kwarg (disable_mmap in diffusers/models/model_loading_utils.py); adding it to transformers makes the two libraries consistent and unblocks users who currently hit TypeError: ...__init__() got an unexpected keyword argument 'disable_mmap' when they try to pass it through.

How

• New _is_on_hf_mount(path) helper parses /proc/mounts on Linux and returns True if path resolves under a mountpoint whose device string is hf-mount (returns False on non-Linux).
• load_state_dict(...) gains a disable_mmap: bool | None = None kwarg. When None, it auto-detects via _is_on_hf_mount. When truthy (and map_location != "meta"), the safetensors file is loaded with safetensors.torch.load(open(path, "rb").read()).
• LoadStateDictConfig gets a matching disable_mmap field; from_pretrained forwards its new kwarg into it; the multi-shard safetensors branch in _load_pretrained_model mirrors the same no-mmap path per file.
• Auto-detection is a no-op on every platform except Linux with an hf-mount FUSE mount, so there's no behavior change for existing users.

Tests

Adds DisableMmapLoadingTest in tests/utils/test_modeling_utils.py covering:

• _is_on_hf_mount matches when /proc/mounts lists hf-mount for the path.
• _is_on_hf_mount does not match for a regular filesystem.
• _is_on_hf_mount short-circuits to False on non-Linux platforms.
• load_state_dict(..., disable_mmap=True) returns tensors equal to the mmap path.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

sounds great ty for adding

[ArthurZucker]

@bot /style

[github-actions]

Style fix bot fixed some files and pushed the changes.