[Bug] _load_bin_or_path_file hides the real torch.load error behind UnboundLocalError

[David-Wu1119]

System Info

Code inspection plus a reproduction of the control flow; observed on main (134fa24).

Who can help?

Code owners for tensorrt_llm/_torch/models (@NVIDIA/trt-llm-models-devs per CODEOWNERS).

Information

• The official example scripts
• My own modified scripts

Reproduction

HfWeightLoader._load_bin_or_path_file (tensorrt_llm/_torch/models/checkpoints/hf/weight_loader.py) returns from a finally block:

@staticmethod
def _load_bin_or_path_file(file):
    try:
        part_weights = torch.load(file, weights_only=True, map_location='cpu', mmap=True)
    except Exception:
        logger.warning(f"Failed to load {file} with mmap=True, fallback to mmap=False")
        part_weights = torch.load(file, weights_only=True, map_location='cpu', mmap=False)
    finally:
        return part_weights

When the mmap=False retry also raises, part_weights was never assigned, so the return in finally raises UnboundLocalError and replaces the real error. This path is taken whenever a checkpoint has no *.safetensors and the loader falls back to *.bin/*.pth.

For example, a .bin-only directory that contains HF Trainer's training_args.bin fails torch.load(..., weights_only=True) with UnpicklingError: Weights only load failed ... Unsupported global: ... TrainingArguments, but the user sees:

[ERROR] Error executing _load_bin_or_path_file with args ('.../training_args.bin',): cannot access local variable 'part_weights' where it is not associated with a value

The load fails either way; the problem is that the error reported points at the loader rather than at the file and the real cause. Python 3.14 also emits SyntaxWarning: 'return' in a 'finally' block for this function (PEP 765).

Expected behavior

The original torch.load exception propagates.

Actual behavior

UnboundLocalError: cannot access local variable 'part_weights' ...

Additional notes

The fix is to drop the finally: and return after the try/except, which keeps the successful paths unchanged. This was also pointed out by an automated review on #7171. I have a PR ready and will link it here.

Written with AI assistance and reviewed line by line.

[David-Wu1119]

Fix proposed in #19565.