Weights

Safetensors loading and the WeightStore that resolves unquantized vs. quantized (4-bit) tensors using each checkpoint's quantization metadata.

anllm.weights.WeightStore

anllm.weights.load_safetensors(path)