Qwen3-VL-235B-A22B-Instruct-int4-mixed-AutoRound

from tqdm import tqdm
from accelerate import init_empty_weights
from transformers import Qwen3VLMoeForConditionalGeneration, AutoProcessor, AutoConfig

model_name = "Qwen/Qwen3-VL-235B-A22B-Instruct"
model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    model_name, dtype="auto", device_map="auto"
)
config = AutoConfig.from_pretrained(model_name)
orig_state_dict = model.state_dict()
model = model.to_empty(device="cpu")
keys = list(orig_state_dict.keys())
num_experts = config.text_config.num_experts
for key in tqdm(keys, desc="convert"):
    if "gate_up_proj" in key and "expert" in key:
        for i in range(num_experts):
            new_key = key.replace("gate_up_proj","gate_up_projs")
            new_key+="."+str(i)+".weight"
            value = orig_state_dict[key][i,...].transpose(0, 1).contiguous()
            orig_state_dict[new_key] = value
        orig_state_dict[key] = None
        orig_state_dict.pop(key)
    elif "down_proj" in key and "expert" in key:
        for i in range(num_experts):
            new_key = key.replace("down_proj", "down_projs")
            new_key += "." + str(i) + ".weight"
            value = orig_state_dict[key][i, ...].transpose(0, 1).contiguous()
            orig_state_dict[new_key] = value
        orig_state_dict[key] = None
        orig_state_dict.pop(key)


with init_empty_weights():
    from modeling_qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
    modify_model = Qwen3VLMoeForConditionalGeneration._from_config(config)

modelfied_state_dict = modify_model.state_dict()
modify_model.load_state_dict(orig_state_dict,strict=True,assign=True)
modify_model.to("cpu")
output_dir="Qwen3-VL-235B-A22B-Instruct-convert-linear"
modify_model.save_pretrained(output_dir)