Update app.py

app.py

@@ -3,10 +3,11 @@ import gradio as gr
 import subprocess, os, sys, time
 import urllib.request, urllib.error
 from fastapi import Request
-from fastapi.responses import Response, HTMLResponse
+from fastapi.responses import Response, PlainTextResponse, HTMLResponse
 
 PORT = 9001
 FM_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "file-manager")
+MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"
 
 # --- Launch the file manager as a background child process ---
 proc = {"p": None}
@@ -22,13 +23,52 @@ def start_fm():
 
 start_fm()
 
-# --- Required ZeroGPU gate (unused, but mandatory to boot) ---
+# --- GPU info: runs nvidia-smi INSIDE the gate (GPU only attached here) ---
 @spaces.GPU
-def gpu_function():
-    return "ok"
+def gpu_report():
+    import torch
+    lines = [f"torch.cuda.is_available(): {torch.cuda.is_available()}"]
+    if torch.cuda.is_available():
+        props = torch.cuda.get_device_properties(0)
+        lines.append(f"Device name: {torch.cuda.get_device_name(0)}")
+        lines.append(f"Device count: {torch.cuda.device_count()}")
+        lines.append(f"Total memory: {props.total_memory / 1024**3:.1f} GB")
+        lines.append(f"Compute capability: {props.major}.{props.minor}")
+    try:
+        smi = subprocess.run(["nvidia-smi"], capture_output=True, text=True, timeout=30)
+        lines.append("\n--- nvidia-smi ---\n" + smi.stdout + smi.stderr)
+    except Exception as e:
+        lines.append(f"nvidia-smi error: {e}")
+    return "\n".join(lines)
+
+# --- LLM inference: model loads + generates INSIDE the gate ---
+_state = {"model": None, "tok": None}
+
+@spaces.GPU
+def run_llm(prompt: str):
+    import torch
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    if _state["model"] is None:
+        _state["tok"] = AutoTokenizer.from_pretrained(MODEL_ID)
+        _state["model"] = AutoModelForCausalLM.from_pretrained(
+            MODEL_ID, torch_dtype=torch.float16, device_map="cuda")
+    tok, model = _state["tok"], _state["model"]
+    msgs = [{"role": "user", "content": prompt}]
+    text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+    inputs = tok(text, return_tensors="pt").to("cuda")
+    out = model.generate(**inputs, max_new_tokens=256)
+    reply = tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+    return f"[GPU: {torch.cuda.get_device_name(0)}]\n\n{reply}"
 
+# --- Gradio owns the ONLY public port 7860 ---
 with gr.Blocks() as demo:
-    gr.Markdown("Self-editing computer. File manager at /fm, hot-reload at /reload")
+    gr.Markdown("Self-editing computer. File manager + terminal at /fm")
+    with gr.Row():
+        prompt = gr.Textbox(label="Ask the LLM")
+        out = gr.Textbox(label="Reply")
+    gr.Button("Run LLM").click(run_llm, inputs=prompt, outputs=out)
+    gpu_out = gr.Textbox(label="GPU")
+    gr.Button("GPU info").click(gpu_report, outputs=gpu_out)
 
 demo.queue()
 demo.launch(server_name="0.0.0.0", server_port=7860,
@@ -43,6 +83,16 @@ def reload_fm():
     time.sleep(1)
     return HTMLResponse('Reloaded file-manager process. <a href="/fm/">back to /fm</a>')
 
+# --- GPU info route ---
+@app.get("/gpuinfo")
+def gpuinfo():
+    return PlainTextResponse(gpu_report())
+
+# --- LLM route ---
+@app.get("/llm")
+def llm(q: str = "Say hello and tell me which GPU you are running on."):
+    return PlainTextResponse(run_llm(q))
+
 # --- Proxy: /fm/... -> 127.0.0.1:9001/fm/... (methods + body forwarded) ---
 @app.api_route("/fm", methods=["GET", "POST"])
 @app.api_route("/fm/{path:path}", methods=["GET", "POST"])
@@ -64,5 +114,6 @@ async def proxy(request: Request, path: str = ""):
            if k.lower() in ("set-cookie", "location", "content-type")}
     return Response(content=data, status_code=status, headers=out)
 
+# --- Keep the process alive (launch returned control) ---
 while True:
     time.sleep(3600)