Update

app.py

@@ -13,25 +13,16 @@ from llmcompressor.transformers import oneshot, wrap_hf_model_class
 class CommitInfo:
     repo_url: str
 
-HF_TOKEN = os.environ.get("HF_TOKEN")
-
-def get_model_class(class_name: str):
-    """Dynamically import and return the specified model class from transformers"""
-    try:
-        # Default to AutoModelForCausalLM if not specified
-        if not class_name:
-            from transformers import AutoModelForCausalLM
-            return AutoModelForCausalLM
-            
-        exec(f"from transformers import {class_name}")
-        return eval(class_name)
-    except Exception as e:
-        raise ValueError(f"Failed to import model class {class_name}: {str(e)}")
+# def get_model_class(class_name: str):
+#     """Dynamically import and return the specified model class from transformers"""
+#     try:
+#         exec(f"from transformers import {class_name}")
+#         return eval(class_name)
+#     except Exception as e:
+#         raise ValueError(f"Failed to import model class {class_name}: {str(e)}")
 
 def parse_ignore_list(ignore_str: str) -> List[str]:
     """Parse comma-separated ignore list string into list"""
-    if not ignore_str:
-        return ["lm_head"]  # Default ignore list
     return [item.strip() for item in ignore_str.split(',') if item.strip()]
 
 def create_quantized_model(
@@ -46,7 +37,8 @@ def create_quantized_model(
     errors = []
     try:
         # Get the appropriate model class
-        model_class = get_model_class(model_class_name)
+        exec(f"from transformers import {class_name}")
+        model_class = eval(class_name)
         wrapped_model_class = wrap_hf_model_class(model_class)
         
         # Load model with ZeroGPU
@@ -257,17 +249,16 @@ The steps are:
 2. Enter the model ID you want to quantize
 3. (Optional) Customize ignored layers and model class
 4. Click "Submit"
-5. You'll get a link to your new quantized model repository! 🚀
+5. You'll get a link to your new quantized model repository on your profile! 🚀
 
 ## Advanced Options:
 - **Ignore List**: Comma-separated list of layer patterns to ignore during quantization. Examples:
   - Llama: `lm_head`
   - Phi3v: `re:.*lm_head,re:model.vision_embed_tokens.*`
-  - Pixtral: `re:.*lm_head,re:multi_modal_projector.*`
   - Llama Vision: `re:.*lm_head,re:multi_modal_projector.*,re:vision_model.*`
 - **Model Class**: Specific model class from transformers (default: AutoModelForCausalLM). Examples:
+  - `AutoModelForCausalLM`
   - `MllamaForConditionalGeneration`
-  - `Qwen2VLForConditionalGeneration`
   - `LlavaForConditionalGeneration`
 
 Note: 
@@ -299,8 +290,8 @@ with gr.Blocks(title=title) as demo:
             ignore_str = gr.Text(
                 max_lines=1,
                 label="ignore_list (comma-separated)",
-                placeholder="lm_head,re:vision_model.*",
-                value="lm_head"
+                placeholder="re:.*lm_head,re:vision_model.*",
+                value="re:.*lm_head"
             )
             model_class_name = gr.Text(
                 max_lines=1,