rust-ml-huggingface-porting

Weight loading fails during HuggingFace porting to Candle due to tensor naming mismatches between checkpoints. The skill identifies these mismatches using pattern detection and prefix handling to ensure robust safetensors loading.

Fix tensor naming mismatches in safetensors by applying runtime pattern detection and adjusting model prefix paths. This skill catalogs common tensor name patterns across HuggingFace families to resolve weight loading errors.

Matmul failures occur when tensors lack contiguity before operations in Candle. Enforce tensor contiguity to avoid runtime errors, and verify tensor naming patterns to ensure correct weight loading from safetensors.

Detect the model prefix by checking for the presence or absence of the 'model.' string in tensor names. The skill handles prefix detection automatically, adjusting tensor paths accordingly to align HuggingFace checkpoints with Candle.

The skill catalogs common tensor name patterns across various HuggingFace families to identify naming mismatches. It applies runtime pattern detection to resolve weight loading and matmul errors across different safetensors checkpoints.