Fix safetensors detection in try_collect_weight_map (missing dot in glob)

[Osamaali313]

Problem

try_collect_weight_map (optimum/fx/parallelization/utils.py) builds its weight-file glob patterns with the safetensors dot missing:

use_safetensors, weight_patterns = False, ["*safetensors", "*.bin"]
for pattern in weight_patterns:
    if len(glob.glob(os.path.join(folder_path, pattern))) > 0:
        use_safetensors = pattern == "*.safetensors"   # compares against the DOTTED form
        break

"*safetensors" still matches real files (glob("*safetensors") matches model.safetensors), so the loop breaks on it — but use_safetensors = ("*safetensors" == "*.safetensors") is always False, even for a safetensors-only model.

Impact

With use_safetensors stuck False, for a safetensors model:

• index_path uses WEIGHTS_INDEX_NAME (pytorch_model.bin.index.json) instead of SAFE_WEIGHTS_INDEX_NAME → the real model.safetensors.index.json is never read.
• weight_files globs *.bin → empty for a safetensors model.
• convert_bin_to_safetensors runs on every call (acquires a file lock; no-op with no bin files).

try_collect_weight_map is called from optimum/fx/parallelization/api.py in the FX tensor-parallelism parallelize_model flow.

Evidence (siblings prove intent)

The three other references in the same function all use the dotted "*.safetensors":

• use_safetensors = pattern == "*.safetensors"
• glob.glob(os.path.join(folder_path, "*.safetensors" if use_safetensors else "*.bin"))
• glob.glob(os.path.join(folder_path, "*.safetensors"))

The sibling "*.bin" entry in the same list has its dot; only the safetensors entry is missing it.

Reproduction

Synthetic sharded-safetensors dir (model-0000{1,2}-of-00002.safetensors + model.safetensors.index.json):

	use_safetensors	index used	weight_files
before (*safetensors)	False ❌	pytorch_model.bin.index.json	[]
after (*.safetensors)	True ✅	model.safetensors.index.json	both shards

Fix

use_safetensors, weight_patterns = False, ["*.safetensors", "*.bin"]

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.