Fix: use demo.launch() for ZeroGPU, add ninja, @spaces.GPU on gradio fn

app.py

@@ -467,18 +467,72 @@ def read_root():
         "version": "3.1.0-zerogpu"
     }
 
-# ── Gradio UI ──
-def ui_predict(user_message, max_tokens, bandwidth):
+# ── Gradio UI (must use @spaces.GPU decorated function) ──
+@spaces.GPU(duration=60)
+def gradio_inference(user_message, max_tokens, bandwidth):
+    """Gradio wrapper that triggers ZeroGPU."""
     req = ChatCompletionRequest(
         model=MODEL_ID,
         messages=[ChatMessage(role="user", content=user_message)],
         max_tokens=int(max_tokens), bandwidth=int(bandwidth)
     )
-    result = chat_completions(req)
-    return result
+    # Inline the inference here so @spaces.GPU wraps it
+    t0 = time.perf_counter()
+    prompt = tokenizer.apply_chat_template(
+        [{"role": "user", "content": user_message}],
+        tokenize=False, add_generation_prompt=True
+    )
+    inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
+    if inputs["input_ids"].shape[1] > 1000:
+        inputs["input_ids"] = inputs["input_ids"][:, -1000:]
+        if "attention_mask" in inputs:
+            inputs["attention_mask"] = inputs["attention_mask"][:, -1000:]
+
+    kalpana_cache = KalpanaHuggingFaceCache(config=model.config, bandwidth=int(bandwidth), device="cuda")
+    
+    model.to("cuda")
+    input_ids = inputs["input_ids"].to("cuda")
+    attention_mask = inputs["attention_mask"].to("cuda")
+    
+    eos_tokens = [tokenizer.eos_token_id]
+    
+    with torch.no_grad():
+        output = model.generate(
+            input_ids=input_ids, attention_mask=attention_mask,
+            max_new_tokens=int(max_tokens), temperature=0.7, do_sample=True,
+            top_p=0.9, repetition_penalty=1.1,
+            past_key_values=kalpana_cache, use_cache=True,
+            pad_token_id=tokenizer.eos_token_id, eos_token_id=eos_tokens
+        )
+    
+    generated_text = tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True)
+    t1 = time.perf_counter()
+    
+    model.to("cpu")
+    torch.cuda.empty_cache()
+    
+    prompt_tokens = int(input_ids.shape[1])
+    completion_tokens = len(tokenizer(generated_text).input_ids)
+    total_tokens = prompt_tokens + completion_tokens
+    
+    openai_cost = (prompt_tokens / 1e6) * OPENAI_INPUT_PER_1M + (completion_tokens / 1e6) * OPENAI_OUTPUT_PER_1M
+    kalpana_cost = (total_tokens / 1e6) * KALPANA_SALE_PER_1M
+    savings_pct = round(((openai_cost - kalpana_cost) / openai_cost) * 100, 1) if openai_cost > 0 else 0.0
+
+    return {
+        "model": MODEL_ID,
+        "response": generated_text,
+        "generation_time_sec": round(t1 - t0, 3),
+        "usage": {"prompt_tokens": prompt_tokens, "completion_tokens": completion_tokens, "total_tokens": total_tokens},
+        "cost_comparison": {
+            "openai_gpt4o_cost": f"${openai_cost:.6f}",
+            "kalpana_cost": f"${kalpana_cost:.6f}",
+            "savings": f"{savings_pct}%"
+        }
+    }
 
 demo = gr.Interface(
-    fn=ui_predict,
+    fn=gradio_inference,
     inputs=[
         gr.Textbox(label="Message", value="Explain quantum computing in simple terms."),
         gr.Slider(minimum=10, maximum=512, value=100, step=1, label="Max Tokens"),
@@ -495,6 +549,7 @@ app = gr.mount_gradio_app(app, demo, path="/ui")
 def root_redirect():
     return RedirectResponse(url="/docs")
 
+# ZeroGPU REQUIRES demo.launch() — not uvicorn.run()
 if __name__ == "__main__":
-    import uvicorn
-    uvicorn.run(app, host="0.0.0.0", port=7860)
+    demo.launch(server_name="0.0.0.0", server_port=7860)
+

requirements.txt

@@ -10,4 +10,5 @@ protobuf
 pypdf
 python-multipart
 torch
+ninja
 huggingface_hub<0.25