switch to docker space for huggingface deployment

Dockerfile

@@ -14,8 +14,14 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY . .
 
 # Create necessary directories
-RUN mkdir -p checkpoints export data/shards tokenizer finetune/checkpoints
+RUN mkdir -p checkpoints export data/shards tokenizer \
+             finetune/checkpoints frontend
 
-EXPOSE 3001
+# HuggingFace requires port 7860
+EXPOSE 7860
+
+# HuggingFace runs as non-root user 1000
+RUN useradd -m -u 1000 user
+USER user
 
 CMD ["python", "serve.py"]

README.md

@@ -1,56 +1,30 @@
+cat > README.md << 'EOF'
 ---
 title: My LLM
 emoji: 🧠
 colorFrom: green
 colorTo: black
-sdk: static
+sdk: docker
+app_port: 7860
 pinned: true
 ---
 
-## Quickstart
-
-### 1. Install dependencies
-```bash
-python -m venv env
-source env/bin/activate
-pip install -r requirements.txt
-```
-
-### 2. Train tokenizer
-```bash
-python tokenizer/train_tokenizer.py
-```
-
-### 3. Prepare data
-```bash
-python data/prepare.py
-```
-
-### 4. Train model
-```bash
-# CPU
-python train.py
+# My LLM — Built from Scratch
 
-# GPU (recommended)
-python train.py --device cuda
-```
+A complete, production-grade LLM built entirely from scratch.
 
-### 5. Export + quantize
-```bash
-python export/export_gguf.py --checkpoint checkpoints/ckpt_0000500.pt
-python export/quantize.py --weights export/model_weights.pt
-```
+## Architecture
+- Decoder-only transformer (GPT-style)
+- 35M parameters
+- Grouped Query Attention (GQA)
+- SwiGLU FFN + RoPE + RMSNorm
+- Int8 quantization for CPU inference
 
-### 6. Run server
-```bash
-python serve.py
-# Open http://localhost:3001
-```
-
-### 7. Docker deployment
-```bash
+## Quickstart
+\`\`\`bash
 docker compose up --build
-```
+# Open http://localhost:3001
+\`\`\`
 
 ## Training Results
 | Steps | Loss | Perplexity |
@@ -58,9 +32,5 @@ docker compose up --build
 | 500   | 5.81 | 335        |
 
 ## Tech Stack
-- PyTorch
-- SentencePiece
-- FastAPI + Uvicorn
-- HuggingFace Datasets
-- Docker
+PyTorch · SentencePiece · FastAPI · Docker · HuggingFace
 EOF

serve.py

@@ -129,7 +129,7 @@ def root():
 
 @app.get("/health")
 def health():
-    return {"status": "ok", "model": "my-llm", "params": model.num_params()}
+    return {"status": "ok" if model is not None else "no model loaded", "model": "my-llm", "params": model.num_params() if model is not None else 0}
 
 
 @app.post("/generate", response_model=GenerateResponse)
@@ -174,4 +174,5 @@ def chat(req: GenerateRequest):
     return generate(req)
 
 if __name__ == "__main__":
-    uvicorn.run("serve:app", host="0.0.0.0", port=3001, reload=False)
+    port = int(os.environ.get("PORT", 7860))
+    uvicorn.run("serve:app", host="0.0.0.0", port=port, reload=False)