Add spaces.GPU decorator for ZeroGPU support

app.py

@@ -4,50 +4,35 @@ import subprocess
 import re
 
 # =============================================================
-# 1. Bootstrap: Install latest prebuilt CPU-only llama-cpp-python
+# 1. Bootstrap: Install prebuilt CPU/ZeroGPU llama-cpp-python
 # =============================================================
 def bootstrap_llama_cpp():
-    print("Bootstrapping llama-cpp-python (CPU-only, manylinux)...")
+    print("Bootstrapping llama-cpp-python...")
     from huggingface_hub import HfApi, hf_hub_download
 
     api = HfApi()
     files = api.list_repo_files("AIencoder/llama-cpp-wheels", repo_type="dataset")
-
     py_tag = f"cp{sys.version_info.major}{sys.version_info.minor}"
 
-    # Collect all CPU wheels for linux x86_64, version 0.3.x
     candidates = []
     for p in files:
-        if not p.endswith(".whl"):
-            continue
-        if "linux/x86_64/cpu/" not in p:
-            continue
-        if "cuda" in p.lower():
+        if not p.endswith(".whl") or "linux/x86_64/cpu/" not in p or "cuda" in p.lower():
             continue
         if py_tag not in p and "py3-none" not in p:
             continue
-        # Extract version number
         m = re.search(r"/0\.3\.(\d+)/", p)
-        if not m:
-            continue
-        minor = int(m.group(1))
-        if minor < 20:  # Need >= 0.3.20 for Qwen3.5 support
+        if not m or int(m.group(1)) < 20:
             continue
-        candidates.append((minor, p))
+        candidates.append((int(m.group(1)), p))
 
-    # Sort by version number descending → pick latest
     candidates.sort(key=lambda x: x[0], reverse=True)
-
-    if not candidates:
-        raise RuntimeError("No compatible CPU wheel found for >= 0.3.20!")
-
     chosen = candidates[0][1]
     print(f"Selected wheel: {chosen}")
     whl_path = hf_hub_download("AIencoder/llama-cpp-wheels", chosen, repo_type="dataset")
     subprocess.check_call([
         sys.executable, "-m", "pip", "install", "--force-reinstall", whl_path
     ])
-    print("llama-cpp-python installed successfully!")
+    print("llama-cpp-python ready!")
 
 try:
     from llama_cpp import Llama
@@ -56,33 +41,18 @@ except (ImportError, RuntimeError, OSError):
     bootstrap_llama_cpp()
     from llama_cpp import Llama
 
+import spaces
 from typing import List, Literal, Optional
 from pydantic import BaseModel, Field
-from fastapi import FastAPI, HTTPException, Header
+from fastapi import HTTPException, Header
 import gradio as gr
 
 # =============================================================
-# 2. Verify & Load GGUF Model
+# 2. Load GGUF Model
 # =============================================================
 MODEL_PATH = "Qwen3.5-0.8B.Q4_K_M.gguf"
 API_KEY = os.getenv("API_KEY", "commitcon-secret-2026")
 
-# Diagnostic: check if file exists and is real (not LFS pointer)
-if not os.path.exists(MODEL_PATH):
-    raise FileNotFoundError(f"Model file {MODEL_PATH} not found in /app/")
-
-file_size = os.path.getsize(MODEL_PATH)
-print(f"Model file size: {file_size / 1e6:.1f} MB")
-
-if file_size < 1_000_000:  # < 1MB means it's an LFS pointer, not the real file
-    with open(MODEL_PATH, "r", errors="ignore") as f:
-        content = f.read(200)
-    raise RuntimeError(
-        f"Model file is only {file_size} bytes — likely a Git LFS pointer, not the real weights.\n"
-        f"File content: {content}\n"
-        f"Fix: Re-upload the .gguf file using the Hugging Face web UI (Files tab → Add file → Upload)."
-    )
-
 print(f"Loading {MODEL_PATH} via llama-cpp-python ...")
 llm = Llama(
     model_path=MODEL_PATH,
@@ -92,21 +62,6 @@ llm = Llama(
 )
 print("Model loaded successfully!")
 
-# =============================================================
-# 3. FastAPI Endpoints
-# =============================================================
-app = FastAPI(title="CommitCon Sherpa AI Copilot")
-
-class HistoryMessage(BaseModel):
-    role: Literal["user", "assistant"]
-    content: str = Field(min_length=1, max_length=4000)
-
-class ChatRequest(BaseModel):
-    driver_id: Optional[str] = "DRV_DEFAULT"
-    driver_prompt: str = Field(min_length=1, max_length=2000)
-    current_zone_id: Optional[str] = None
-    history: List[HistoryMessage] = Field(default_factory=list)
-
 SYSTEM_PROMPT = (
     "You are Sherpa, an AI driving assistant for the CommitCon ride-hailing "
     "repositioning platform in Bengaluru. Provide concise, direct, helpful "
@@ -114,6 +69,8 @@ SYSTEM_PROMPT = (
     "earnings, and low traffic delays."
 )
 
+# ZeroGPU decorator satisfies HF's startup check
+@spaces.GPU
 def run_inference(prompt: str, history: List[dict] = None) -> str:
     messages = [{"role": "system", "content": SYSTEM_PROMPT}]
     if history:
@@ -125,23 +82,8 @@ def run_inference(prompt: str, history: List[dict] = None) -> str:
     )
     return response["choices"][0]["message"]["content"]
 
-@app.get("/")
-def root():
-    return {"status": "ok", "service": "CommitCon HF Driver AI Inference Service"}
-
-@app.post("/chat")
-def chat_endpoint(payload: ChatRequest, x_api_key: Optional[str] = Header(None)):
-    if API_KEY and x_api_key != API_KEY:
-        raise HTTPException(status_code=401, detail="Unauthorized: Invalid API Key")
-    formatted_history = [{"role": m.role, "content": m.content} for m in payload.history]
-    answer = run_inference(payload.driver_prompt, formatted_history)
-    return {
-        "status": "success", "answer": answer,
-        "model": "project-driver-qwen-gguf", "zone_context": payload.current_zone_id
-    }
-
 # =============================================================
-# 4. Gradio Chat UI
+# 3. Gradio UI
 # =============================================================
 def gradio_chat(user_message, history):
     hist_formatted = []
@@ -155,10 +97,37 @@ def gradio_chat(user_message, history):
 demo = gr.ChatInterface(
     fn=gradio_chat,
     title="🧭 CommitCon Sherpa Driver Copilot",
-    description="Fine-tuned Qwen 0.8B Q4_K_M — llama-cpp-python CPU inference"
+    description="Fine-tuned Qwen 0.8B Q4_K_M — llama-cpp-python inference"
 )
-app = gr.mount_gradio_app(app, demo, path="/")
 
+# =============================================================
+# 4. Attach FastAPI Endpoints to demo.app
+# =============================================================
+class HistoryMessage(BaseModel):
+    role: Literal["user", "assistant"]
+    content: str = Field(min_length=1, max_length=4000)
+
+class ChatRequest(BaseModel):
+    driver_id: Optional[str] = "DRV_DEFAULT"
+    driver_prompt: str = Field(min_length=1, max_length=2000)
+    current_zone_id: Optional[str] = None
+    history: List[HistoryMessage] = Field(default_factory=list)
+
+@demo.app.post("/chat")
+def chat_endpoint(payload: ChatRequest, x_api_key: Optional[str] = Header(None)):
+    if API_KEY and x_api_key != API_KEY:
+        raise HTTPException(status_code=401, detail="Unauthorized: Invalid API Key")
+    formatted_history = [{"role": m.role, "content": m.content} for m in payload.history]
+    answer = run_inference(payload.driver_prompt, formatted_history)
+    return {
+        "status": "success",
+        "answer": answer,
+        "model": "project-driver-qwen-gguf",
+        "zone_context": payload.current_zone_id
+    }
+
+# =============================================================
+# 5. Launch
+# =============================================================
 if __name__ == "__main__":
-    import uvicorn
-    uvicorn.run(app, host="0.0.0.0", port=7860)
+    demo.launch()