app.py

import gradio as gr
from transformers import AutoModel, AutoConfig
import torch
import json
from collections import defaultdict, OrderedDict

def analyze_model_parameters(model_path, hf_token=None, show_layer_details=False):
    try:
        # Prepare token parameter
        token_kwargs = {}
        if hf_token and hf_token.strip():
            token_kwargs['token'] = hf_token.strip()
        
        # Load model configuration first
        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True, **token_kwargs)
        
        # Load model on CPU - FIXED: Use low_cpu_mem_usage=False to avoid meta tensors
        model = AutoModel.from_pretrained(
            model_path, 
            torch_dtype=torch.float32,  # Explicitly set dtype
            low_cpu_mem_usage=False,    # Disable meta tensors
            trust_remote_code=True, 
            **token_kwargs
        )
        
        # Ensure model is on CPU
        model = model.cpu()
        
        # Initialize counters
        total_params = 0
        trainable_params = 0
        embedding_params = 0
        non_embedding_params = 0
        
        # Track unique parameters to handle weight tying
        unique_params = {}
        param_details = []
        layer_breakdown = defaultdict(lambda: {'total': 0, 'trainable': 0, 'params': []})
        
        # Embedding layer patterns (common names for embedding layers)
        embedding_patterns = [
            'embeddings', 'embed', 'wte', 'wpe', 'word_embedding', 
            'position_embedding', 'token_embedding', 'embed_tokens',
            'embed_positions', 'embed_layer_norm'
        ]
        
        def is_embedding_param(name):
            name_lower = name.lower()
            return any(pattern in name_lower for pattern in embedding_patterns)
        
        def get_layer_name(param_name):
            """Extract layer information from parameter name"""
            parts = param_name.split('.')
            if len(parts) >= 2:
                # Handle common transformer architectures
                if 'layer' in parts or 'layers' in parts:
                    for i, part in enumerate(parts):
                        if part in ['layer', 'layers'] and i + 1 < len(parts):
                            try:
                                layer_num = int(parts[i + 1])
                                return f"Layer {layer_num}"
                            except ValueError:
                                pass
                # Handle other patterns
                if 'encoder' in parts:
                    return "Encoder"
                elif 'decoder' in parts:
                    return "Decoder"
                elif any(emb in param_name.lower() for emb in embedding_patterns):
                    return "Embeddings"
                elif 'classifier' in param_name.lower() or 'head' in param_name.lower():
                    return "Classification Head"
                elif 'pooler' in param_name.lower():
                    return "Pooler"
                elif 'ln' in param_name.lower() or 'norm' in param_name.lower():
                    return "Layer Norm"
            return "Other"
        
        # Analyze all parameters
        for name, param in model.named_parameters():
            param_size = param.numel()
            is_trainable = param.requires_grad
            is_embedding = is_embedding_param(name)
            layer_name = get_layer_name(name)
            
            # Handle weight tying by using data pointer
            ptr = param.data_ptr()
            if ptr not in unique_params:
                unique_params[ptr] = {
                    'name': name,
                    'size': param_size,
                    'trainable': is_trainable,
                    'embedding': is_embedding,
                    'layer': layer_name,
                    'shape': list(param.shape)
                }
                
                # Add to totals
                total_params += param_size
                if is_trainable:
                    trainable_params += param_size
                if is_embedding:
                    embedding_params += param_size
                else:
                    non_embedding_params += param_size
                    
                # Add to layer breakdown
                layer_breakdown[layer_name]['total'] += param_size
                if is_trainable:
                    layer_breakdown[layer_name]['trainable'] += param_size
                
            # Add parameter details
            param_details.append({
                'name': name,
                'shape': list(param.shape),
                'size': param_size,
                'trainable': is_trainable,
                'embedding': is_embedding,
                'layer': layer_name,
                'shared': ptr in [p['ptr'] for p in param_details if 'ptr' in p],
                'ptr': ptr
            })
            
            # Add to layer breakdown details
            layer_breakdown[layer_name]['params'].append({
                'name': name,
                'shape': list(param.shape),
                'size': param_size,
                'trainable': is_trainable
            })
        
        # Format the summary
        summary = f"""
🔍 **MODEL ANALYSIS: {model_path}**

📊 **PARAMETER SUMMARY**
├── Total Parameters: {total_params:,}
├── Trainable Parameters: {trainable_params:,}
├── Non-trainable Parameters: {total_params - trainable_params:,}
└── Trainable Percentage: {(trainable_params/total_params*100):.1f}%

🧠 **PARAMETER BREAKDOWN**
├── Embedding Parameters: {embedding_params:,} ({embedding_params/total_params*100:.1f}%)
└── Non-embedding Parameters: {non_embedding_params:,} ({non_embedding_params/total_params*100:.1f}%)

📋 **MODEL INFO**
├── Model Type: {config.model_type if hasattr(config, 'model_type') else 'Unknown'}
├── Architecture: {config.architectures[0] if hasattr(config, 'architectures') and config.architectures else 'Unknown'}
└── Hidden Size: {getattr(config, 'hidden_size', 'Unknown')}
"""

        # Add layer breakdown summary
        if layer_breakdown:
            summary += "\n🏗️ **LAYER BREAKDOWN SUMMARY**\n"
            sorted_layers = sorted(layer_breakdown.items(), key=lambda x: (
                0 if x[0] == "Embeddings" else
                1 if x[0].startswith("Layer") else
                2 if x[0] == "Layer Norm" else
                3 if x[0] == "Pooler" else
                4 if x[0] == "Classification Head" else 5
            ))
            
            for layer_name, info in sorted_layers:
                percentage = info['total'] / total_params * 100
                summary += f"├── {layer_name}: {info['total']:,} params ({percentage:.1f}%)\n"
        
        # Detailed layer breakdown if requested
        layer_details = ""
        if show_layer_details:
            layer_details = "\n" + "="*60 + "\n"
            layer_details += "🔍 **DETAILED LAYER-BY-LAYER BREAKDOWN**\n"
            layer_details += "="*60 + "\n"
            
            for layer_name, info in sorted_layers:
                layer_details += f"\n📁 **{layer_name.upper()}**\n"
                layer_details += f"   Total: {info['total']:,} | Trainable: {info['trainable']:,}\n"
                layer_details += f"   Parameters:\n"
                
                for param_info in info['params']:
                    trainable_mark = "✓" if param_info['trainable'] else "✗"
                    layer_details += f"   {trainable_mark} {param_info['name']}: {param_info['shape']} → {param_info['size']:,}\n"
        
        return summary + layer_details
        
    except Exception as e:
        error_msg = str(e)
        if "401" in error_msg or "authentication" in error_msg.lower():
            return f"🔒 **Authentication Error:** This model requires a valid HuggingFace token.\n\nPlease provide your HuggingFace token in the token field above.\n\nOriginal error: {error_msg}"
        elif "404" in error_msg or "not found" in error_msg.lower():
            return f"🔍 **Model Not Found:** The model '{model_path}' was not found.\n\nPlease check:\n- Model path is correct\n- Model exists on HuggingFace Hub\n- You have access to the model (use token if private)\n\nOriginal error: {error_msg}"
        else:
            return f"❌ **Error loading model:** {error_msg}\n\nPlease check that the model path is correct and accessible."

def count_parameters_basic(model_path, hf_token=None):
    """Basic parameter counting without layer details"""
    return analyze_model_parameters(model_path, hf_token, show_layer_details=False)

def count_parameters_detailed(model_path, hf_token=None):
    """Detailed parameter counting with layer-by-layer breakdown"""
    return analyze_model_parameters(model_path, hf_token, show_layer_details=True)

# Create Gradio interface with multiple outputs
with gr.Blocks(title="🤗 Advanced HuggingFace Model Parameter Analyzer", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🤗 Advanced HuggingFace Model Parameter Analyzer
    
    Enter any HuggingFace model path to get detailed parameter analysis including:
    - **Total & trainable parameter counts**
    - **Embedding vs non-embedding breakdown** 
    - **Layer-by-layer analysis**
    - **Weight sharing detection**
    - **Private model access** with HuggingFace token
    """)
    
    with gr.Row():
        with gr.Column(scale=2):
            model_input = gr.Textbox(
                label="🔍 HuggingFace Model Path",
                placeholder="e.g., bert-base-uncased, gpt2, microsoft/DialoGPT-medium",
                value="bert-base-uncased"
            )
            
        with gr.Column(scale=1):
            hf_token_input = gr.Textbox(
                label="🔑 HuggingFace Token (Optional)",
                placeholder="hf_...",
                type="password",
                info="Required for private models or gated models"
            )
            
    with gr.Row():
        analyze_btn = gr.Button("📊 Analyze Model", variant="primary")
        detailed_btn = gr.Button("🔍 Detailed Analysis", variant="secondary")
    
    output_text = gr.Textbox(
        label="📋 Analysis Results",
        lines=20,
        max_lines=50,
        show_copy_button=True
    )
    
    # Event handlers
    analyze_btn.click(
        fn=count_parameters_basic,
        inputs=[model_input, hf_token_input],
        outputs=output_text
    )
    
    detailed_btn.click(
        fn=count_parameters_detailed,
        inputs=[model_input, hf_token_input],
        outputs=output_text
    )
    
    # Example models
    gr.Examples(
        examples=[
            ["bert-base-uncased"],
            ["gpt2"],
            ["roberta-base"],
            ["distilbert-base-uncased"],
            ["microsoft/DialoGPT-medium"],
            ["facebook/bart-base"],
            ["t5-small"],
            ["google/flan-t5-small"]
        ],
        inputs=model_input,
        label="🎯 Example Models"
    )
    
    gr.Markdown("""
    ### 📝 Notes:
    - **Weight tying detection**: Automatically handles shared parameters (e.g., input/output embeddings)
    - **Layer categorization**: Groups parameters by transformer layers, embeddings, etc.
    - **Detailed analysis**: Click "Detailed Analysis" for parameter-by-parameter breakdown
    - **Private models**: Use your HuggingFace token to access private or gated models
    - **Token security**: Token is only used for this session and not stored
    - **Model compatibility**: Works with most HuggingFace transformer models
    """)

if __name__ == "__main__":
    demo.launch()