Function gguf_clip_loader

loader.py:470–506

(path)

    return torch.ByteTensor(list(spm.SerializeToString()))

def gguf_clip_loader(path):
    sd, extra = gguf_sd_loader(path, is_text_model=True)
    arch = extra.get("arch_str", None)
    if arch in {"t5", "t5encoder"}:
        temb_key = "token_embd.weight"
        if temb_key in sd and sd[temb_key].shape == (256384, 4096):
            # non-standard Comfy-Org tokenizer
            sd["spiece_model"] = gguf_tokenizer_loader(path, sd[temb_key].shape)
            # TODO: dequantizing token embed here is janky but otherwise we OOM due to tensor being massive.
            logging.warning(f"Dequantizing {temb_key} to prevent runtime OOM.")
            sd[temb_key] = dequantize_tensor(sd[temb_key], dtype=torch.float16)
        sd = sd_map_replace(sd, T5_SD_MAP)
    elif arch in {"llama", "qwen2vl", "qwen3", "qwen3vl", "gemma3"}:
        # TODO: pass model_options["vocab_size"] to loader somehow
        temb_key = "token_embd.weight"
        if temb_key in sd and sd[temb_key].shape[0] >= (64 * 1024):
            if arch == "llama" and sd[temb_key].shape == (131072, 5120):
                # non-standard Comfy-Org tokenizer
                sd["tekken_model"] = gguf_tekken_tokenizer_loader(path, sd[temb_key].shape)
            elif arch == "gemma3":
                sd["spiece_model"] = gguf_gemma3_tokenizer_loader(path)
            # See note above for T5.
            logging.warning(f"Dequantizing {temb_key} to prevent runtime OOM.")
            sd[temb_key] = dequantize_tensor(sd[temb_key], dtype=torch.float16)
        if arch == "gemma3":
            sd = sd_map_replace(sd, GEMMA3_SD_MAP)
            sd = gemma3_norm_corrections(sd)
        else:
            sd = sd_map_replace(sd, LLAMA_SD_MAP)
        if arch == "llama":
            sd = llama_permute(sd, 32, 8) # L3 / Mistral
        if arch == "qwen2vl":
            vsd = gguf_mmproj_loader(path)
            sd.update(vsd)
    else:
        pass
    return sd