zen-family/tools/deployment/push_to_hf.py

#!/usr/bin/env python3
"""
Push fine-tuned Zen-Omni model to HuggingFace
"""

import os
from pathlib import Path
from huggingface_hub import HfApi, login, create_repo, upload_folder
from transformers import AutoTokenizer

print("""
╔═══════════════════════════════════════════════════════╗
║     PUSHING ZEN-OMNI TO HUGGINGFACE                  ║
╚═══════════════════════════════════════════════════════╝
""")

# Model details
model_path = "./zen-omni-m1-finetuned"
repo_name = "zenlm/zen-omni-1.5b-lora"  # Will be under your account

print("🔐 Logging in to HuggingFace...")
# This will use your token from HF_TOKEN env or prompt for login
try:
    api = HfApi()
    # Check if already logged in
    user_info = api.whoami()
    print(f"✅ Logged in as: {user_info['name']}")
    username = user_info['name']
    repo_id = f"{username}/zen-omni-1.5b-lora"
except:
    print("Please login with your HuggingFace token:")
    login()
    api = HfApi()
    user_info = api.whoami()
    username = user_info['name']
    repo_id = f"{username}/zen-omni-1.5b-lora"

print(f"\n📦 Preparing to push to: {repo_id}")

# Create repository
print("🚀 Creating repository...")
try:
    create_repo(
        repo_id=repo_id,
        repo_type="model",
        exist_ok=True,
        private=False  # Set to True for private repo
    )
    print(f"✅ Repository ready: https://huggingface.co/{repo_id}")
except Exception as e:
    print(f"Repository exists or error: {e}")

# Create model card
print("\n📝 Creating model card...")
model_card = f"""---
language:
- en
- zh
license: apache-2.0
base_model: Qwen/zen-1.5B-Instruct
tags:
- multimodal
- zen-omni
- hanzo-mcp
- lora
- qwen
- text-generation
library_name: peft
pipeline_tag: text-generation
---

# Zen-Omni 1.5B LoRA

Fine-tuned version of zen-1.5B with specialized knowledge about:
- **Zen-Omni**: Multimodal AI architecture based on Qwen3-Omni
- **Hanzo MCP**: Model Context Protocol tools
- **Thinker-Talker**: MoE architecture for low-latency streaming

## Model Details

- **Base Model**: Qwen/zen-1.5B-Instruct
- **Fine-tuning**: QLoRA with rank 4
- **Training Device**: Apple M1 Max
- **Parameters**: 544,768 trainable (0.035% of model)
- **Training Steps**: 50 (quick demo)

## Key Knowledge Areas

### 1. Zen-Omni Architecture
- Based on Qwen3-Omni-30B-A3B architecture
- Supports 119 text languages, 19 speech input, 10 speech output
- 234ms first-packet latency
- Thinker-Talker MoE design

### 2. Hanzo MCP Integration
- Python: `pip install hanzo-mcp`
- Node.js: `npm install @hanzo/mcp`
- Unified multimodal search across text, image, audio, video

### 3. Technical Components
- AuT encoder: 650M params, 12.5Hz token rate
- Vision: SigLIP2-So400M (540M params)
- Thinker: 30B-A3B MoE
- Talker: 3B-A0.3B MoE
- Code2wav: 200M ConvNet

## Usage

### With PEFT

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/zen-1.5B-Instruct",
    torch_dtype=torch.float16,
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "{repo_id}")

# Use tokenizer
tokenizer = AutoTokenizer.from_pretrained("{repo_id}")

# Generate
prompt = "User: What is Zen-Omni?\\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Training Data

Trained on specialized examples about:
- Zen-Omni multimodal capabilities
- Hanzo MCP tools usage (Python/Node.js)
- Thinker-Talker architecture
- Low-latency streaming techniques
- Cross-modal reasoning

## Performance

- Training Loss: 2.9 → 1.6 (50 steps)
- Inference Speed: ~1.25 steps/sec on M1 Max
- Response Quality: Learning concepts, needs more training for production

## Limitations

- Quick demo training (only 50 steps)
- Small base model (1.5B)
- Limited to text modality in this version
- Needs more epochs for full knowledge retention

## Citation

```bibtex
@misc{{zen-omni-2024,
  title={{Zen-Omni: Fine-tuned Multimodal Assistant}},
  author={{Zen Team}},
  year={{2024}},
  url={{https://huggingface.co/{repo_id}}}
}}
```

## License

Apache 2.0 (inherited from zen)
"""

# Save model card
model_card_path = Path(model_path) / "README.md"
model_card_path.write_text(model_card)
print("✅ Model card created")

# Upload to HuggingFace
print("\n📤 Uploading model to HuggingFace...")
try:
    api.upload_folder(
        folder_path=model_path,
        repo_id=repo_id,
        repo_type="model",
        commit_message="Upload Zen-Omni 1.5B LoRA fine-tuned on M1 Max"
    )
    print(f"\n✅ Model successfully uploaded!")
    print(f"🔗 View your model at: https://huggingface.co/{repo_id}")
    print(f"\n📦 Install with:")
    print(f"   from peft import PeftModel")
    print(f"   model = PeftModel.from_pretrained(base_model, '{repo_id}')")
    
except Exception as e:
    print(f"❌ Upload failed: {e}")
    print("\nTry manual upload:")
    print(f"1. huggingface-cli login")
    print(f"2. huggingface-cli upload {repo_id} {model_path}")

print("\n" + "="*60)
print("Next steps:")
print(f"1. View model: https://huggingface.co/{repo_id}")
print("2. Test inference from HF Hub")
print("3. Share with community!")