Function extract_layer_idx

tensorrt_llm/quantization/quantize.py:267–272

(name)

        set(exclude_modules + ['*ln_f', '*ln_embed', '*lm_head']))

    def extract_layer_idx(name):
        ss = name.split('.')
        for s in ss:
            if s.isdigit():
                return int(s)
        return None

    # Meta's LLaMA 3.1 recipe:
    # (1) Skip quantization for the first and last Transformer layers