Description

The HF bin/PTH loader retries without mmap when the initial load fails. Its return currently sits in a finally block, so if the fallback also fails, Python replaces the loader exception with an UnboundLocalError.

Move the return after the retry block. Successful loads and the mmap fallback are unchanged, while fallback failures now propagate their original exception.

This restores the behavior from before bin checkpoint loading was parallelized.

Test Coverage

• Added a unit test that forces both mmap attempts to fail, verifies the retry order, and checks that the fallback exception is preserved.
• pre-commit run --files tensorrt_llm/_torch/models/checkpoints/hf/weight_loader.py tests/unittest/_torch/models/checkpoints/hf/test_weight_loader.py
• python3 -Werror::SyntaxWarning -m py_compile tensorrt_llm/_torch/models/checkpoints/hf/weight_loader.py tests/unittest/_torch/models/checkpoints/hf/test_weight_loader.py

PR Checklist

• Reviewed the pull request checklist as applicable.

Dev Engineer Review

• Moved the return outside the finally block so fallback-load exceptions propagate unchanged.
• Preserves successful loads and mmap retry behavior without API or performance changes.
• No configuration or test-list changes.

QA Engineer Review

• Added test_bin_loader_propagates_fallback_error.
• Verifies mmap-first loading, non-mmap fallback, call order, arguments, and fallback exception propagation.
• No corresponding test-db/ or qa/ coverage entry was identified.
• Verdict: insufficient.