hugging-face-space-deployer

# Hugging Face Space Deployer

A skill for AI engineers to create, configure, and deploy interactive ML demos on Hugging Face Spaces.

## CRITICAL: Pre-Deployment Checklist

**Before writing ANY code, gather this information about the model:**

### 1. Check Model Type (LoRA Adapter vs Full Model)

**Use the HF MCP tool to inspect the model files:**
```
hf-skills - Hub Repo Details (repo_ids: ["username/model"], repo_type: "model")
```

**Look for these indicators:**

| Files Present | Model Type | Action Required |
|---------------|------------|-----------------|
| `model.safetensors` or `pytorch_model.bin` | Full model | Load directly with `AutoModelForCausalLM` |
| `adapter_model.safetensors` + `adapter_config.json` | LoRA/PEFT adapter | Must load base model first, then apply adapter with `peft` |
| Only config files, no weights | Broken/incomplete | Ask user to verify |

**If adapter_config.json exists, check for `base_model_name_or_path` to identify the base model.**

### 2. Check Inference API Availability

Visit the model page on HF Hub and look for "Inference Providers" widget on the right side.

**Indicators that model HAS Inference API:**
- Inference widget visible on model page
- Model from known provider: `meta-llama`, `mistralai`, `HuggingFaceH4`, `google`, `stabilityai`, `Qwen`
- High download count (>10,000) with standard architecture

**Indicators that model DOES NOT have Inference API:**
- Personal namespace (e.g., `GhostScientist/my-model`)
- LoRA/PEFT adapter (adapters never have direct Inference API)
- Missing `pipeline_tag` in model metadata
- No inference widget on model page

### 3. Check Model Metadata

- Ensure `pipeline_tag` is set (e.g., `text-generation`)
- Add `conversational` tag for chat models

### 4. Determine Hardware Needs

| Model Size | Recommended Hardware |
|------------|---------------------|
| < 3B parameters | ZeroGPU (free) or CPU |
| 3B - 7B parameters | ZeroGPU or T4 |
| > 7B parameters | A10G or A100 |

### 5. Ask User If