app_simple_backup.py

"""
Zen VL Training - Simplified & Working
Just trains zen-vl-4b with our datasets
"""

import gradio as gr
import torch
from transformers import AutoModel, AutoProcessor, TrainingArguments, Trainer
from datasets import load_dataset

def train_zen_vl():
    """Simple one-button training for zen-vl-4b"""
    
    logs = []
    
    def log(msg):
        print(msg)
        logs.append(msg)
        yield "\n".join(logs)
    
    try:
        yield from log("🧘 Starting Zen VL 4B Training")
        yield from log("=" * 80)
        
        # GPU check
        has_gpu = torch.cuda.is_available()
        yield from log(f"🎮 GPU: {has_gpu}")
        if has_gpu:
            yield from log(f"   {torch.cuda.get_device_name(0)}")
            yield from log(f"   {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f}GB")
        
        # Load model
        yield from log("\n📦 Loading zen-vl-4b-instruct...")
        model = AutoModel.from_pretrained(
            "zenlm/zen-vl-4b-instruct",
            torch_dtype=torch.bfloat16,
            device_map="auto",
            trust_remote_code=True
        )
        processor = AutoProcessor.from_pretrained("zenlm/zen-vl-4b-instruct")
        yield from log("✅ Model loaded")
        
        # Load datasets
        yield from log("\n📚 Loading datasets...")
        all_data = []
        
        datasets_to_load = [
            ("ADP Synatra", "neulab/agent-data-collection", "synatra", 7500),
            ("ADP Code Feedback", "neulab/agent-data-collection", "code_feedback", 7500),
            ("ADP Go Browse", "neulab/agent-data-collection", "go-browse-wa", 7500),
            ("xLAM Function Calling", "Salesforce/xlam-function-calling-60k", None, 7500)
        ]
        
        for name, hf_id, config, max_samples in datasets_to_load:
            yield from log(f"  Loading {name}...")
            try:
                if config:
                    ds = load_dataset(hf_id, config, split="train", streaming=True)
                else:
                    ds = load_dataset(hf_id, split="train", streaming=True)
                
                samples = []
                for i, example in enumerate(ds):
                    if i >= max_samples:
                        break
                    samples.append(example)
                
                all_data.extend(samples)
                yield from log(f"    ✅ {len(samples)} samples")
            except Exception as e:
                yield from log(f"    ⚠️  Error: {e}")
        
        yield from log(f"\n✅ Total: {len(all_data)} samples")
        
        # Training
        yield from log("\n⚙️  Training Configuration:")
        yield from log("  Epochs: 3")
        yield from log("  Batch Size: 1")
        yield from log("  Learning Rate: 2e-5")
        yield from log("  Output: zenlm/zen-vl-4b-agent")
        
        training_args = TrainingArguments(
            output_dir="./zen-vl-output",
            num_train_epochs=3,
            per_device_train_batch_size=1,
            learning_rate=2e-5,
            logging_steps=10,
            save_steps=500,
            bf16=True,
            push_to_hub=True,
            hub_model_id="zenlm/zen-vl-4b-agent",
            report_to="none",
        )
        
        trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=all_data if len(all_data) > 0 else None,
        )
        
        yield from log("\n🔥 TRAINING STARTED")
        yield from log("=" * 80)
        
        result = trainer.train()
        
        yield from log("\n✅ TRAINING COMPLETED!")
        yield from log(f"📊 Final Loss: {result.training_loss:.4f}")
        yield from log("☁️  Uploading to zenlm/zen-vl-4b-agent...")
        
        trainer.push_to_hub()
        
        yield from log("\n🎉 SUCCESS! Model live at zenlm/zen-vl-4b-agent")
        
    except Exception as e:
        yield from log(f"\n❌ ERROR: {str(e)}")
        import traceback
        yield from log(f"\n{traceback.format_exc()}")

# Simple interface
with gr.Blocks(title="Zen VL Training") as demo:
    gr.Markdown("""
    # 🧘 Zen VL 4B Training
    
    Trains zen-vl-4b-instruct → zen-vl-4b-agent
    
    **Datasets**: ADP (Synatra, Code Feedback, Go Browse) + xLAM (60k)
    **Total**: ~30k samples
    **Time**: ~6-8 hours on A10G
    **Output**: zenlm/zen-vl-4b-agent
    """)
    
    start_btn = gr.Button("🚀 Start Training", variant="primary", size="lg")
    output = gr.Textbox(label="Training Logs", lines=30)
    
    start_btn.click(train_zen_vl, outputs=output)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)