llmfan46

/

Qwen3.6-35B-A3B-uncensored-heretic

like 90

The vision model weights in this repository are nested under an incorrect prefix than the original model from Qwen repository.

The original model uses model.visual.* as a prefix, but this repo uses model.language_model.visual.*.

Because the architecture Qwen3_5MoeForConditionalGeneration expects the vision tower to be a sibling of the language model rather than a child of it, loader scripts for conversion tools fail with a ValueError reporting hundreds of "extra" parameters.

The affected files are model.safetensors.index.json and model-00002-of-00002.safetensors. To fix this at the source, the keys inside the safetensors and the index map must be renamed. Below is a Python script that performs this correction:

import json
import os
from safetensors import safe_open
from safetensors.torch import save_file

model_path = "./llmfan46/Qwen3.6-35B-A3B-uncensored-heretic"
index_file = os.path.join(model_path, "model.safetensors.index.json")

# 1. Update Index Map
with open(index_file, "r") as f:
    index = json.load(f)

weight_map = index["weight_map"]
new_weight_map = {}
files_to_update = set()

for key, val in weight_map.items():
    if key.startswith("model.language_model.visual."):
        new_key = key.replace("model.language_model.visual.", "model.visual.")
        new_weight_map[new_key] = val
        files_to_update.add(val)
    else:
        new_weight_map[key] = val

index["weight_map"] = new_weight_map
with open(index_file, "w") as f:
    json.dump(index, f, indent=2)

# 2. Update Safetensors Files
for filename in files_to_update:
    file_path = os.path.join(model_path, filename)
    tensors = {}
    with safe_open(file_path, framework="pt", device="cpu") as f:
        for key in f.keys():
            tensor = f.get_tensor(key)
            if key.startswith("model.language_model.visual."):
                new_key = key.replace("model.language_model.visual.", "model.visual.")
                tensors[new_key] = tensor
            else:
                tensors[key] = tensor

    save_file(tensors, file_path)