# 2. 額外載入模型的 Config
        config = AutoConfig.from_pretrained(VLM_MODEL, trust_remote_code=True)

        # 3. 建立一個「空的」模型骨架 (不載入權重)
        #    這一步非常快，且不佔用 CPU RAM
        with init_empty_weights():
            empty_model = Qwen3VLForConditionalGeneration(config)
            # (可選，但推薦) 確保權重一致性
            empty_model.tie_weights() 


        max_memory_map = {
            0: "9GiB", 
            1: "10.5GiB", 
            2: "10.5GiB"
        }

        # 4. 根據「空的」模型骨架和記憶體限制，計算 device_map
        #    這就是您 ??? 的答案：
        device_map = infer_auto_device_map(
            empty_model,  # <-- 填入空模型
            max_memory=max_memory_map,
            # 這是 Qwen-VL 的特殊要求，告訴 accelerate 哪些模組不可分割
            no_split_module_classes=empty_model._no_split_modules 
        )

        # (可選) 印出 device_map 檢查
        print("[INFO] Inferred device map:")
        print(device_map)

        # 5. 載入模型，這次它會使用 device_map 來分散載入
        model = Qwen3VLForConditionalGeneration.from_pretrained(
            VLM_MODEL,
            trust_remote_code=True,
            device_map=device_map, # 傳入計算好的 map
            torch_dtype=torch.float16,
            attn_implementation="flash_attention_2",
            low_cpu_mem_usage=True # 保持開啟
        )```     

  import pynvml  # type: ignore[import]
Traceback (most recent call last):
  File "/media/pc14700k/ssd1tb/vlm/soiling/soiling_test_for_img_tag-hugging-face.py", line 764, in process_images
    model = Qwen3VLForConditionalGeneration.from_pretrained(
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pc14700k/anaconda3/envs/vlm/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/pc14700k/anaconda3/envs/vlm/lib/python3.12/site-packages/transformers/modeling_utils.py", line 5001, in from_pretrained
    hf_quantizer.preprocess_model(
  File "/home/pc14700k/anaconda3/envs/vlm/lib/python3.12/site-packages/transformers/quantizers/base.py", line 225, in preprocess_model
    return self._process_model_before_weight_loading(model, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pc14700k/anaconda3/envs/vlm/lib/python3.12/site-packages/transformers/quantizers/quantizer_awq.py", line 119, in _process_model_before_weight_loading
    model, has_been_replaced = replace_with_awq_linear(
                               ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/pc14700k/anaconda3/envs/vlm/lib/python3.12/site-packages/transformers/integrations/awq.py", line 134, in replace_with_awq_linear
    from awq.modules.linear.gemm import WQLinear_GEMM
ModuleNotFoundError: No module named 'awq.modules'; 'awq' is not a package
[ERROR] Failed to load model QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ: No module named 'awq.modules'; 'awq' is not a package


### Who can help?

_No response_

### Information

- [ ] The official example scripts
- [x] My own modified scripts

### Tasks

- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] My own task or dataset (give details below)

### Reproduction

use QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ
use Qwen3-VL-30B-A3B-Instruct is OK
but use Qwen3-VL-30B-A3B-Instruct-AWQ got the problem log

### Expected behavior

support qwen3-vl awq