Module weights 

HuggingFace safetensors weight loading and key resolution.

Functions

detect_weight_prefix

Detect the common prefix for transformer weight keys.

load_hf_weights

Load and merge safetensors shards from disk.

resolve_bias

Resolve an optional bias tensor by logical suffix (e.g. layers.0.self_attn.q_proj). Returns None when the model has no bias for that layer (e.g. Llama/Mistral).

resolve_lm_head

Resolve lm_head — may live outside the model prefix.

resolve_weight

Resolve a weight tensor by logical suffix (e.g. layers.0.self_attn.q_proj).

tie_word_embeddings_from_config