Update app.py

app.py

@@ -1,64 +1,97 @@
+import gc
+import threading
 import gradio as gr
-from huggingface_hub import InferenceClient
+import torch
+from transformers import pipeline, TextIteratorStreamer, AutoTokenizer
+import spaces
 
-"""
-For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-"""
-client = InferenceClient("google/medgemma-27b-text-it")
+# Single-model configuration
+MODEL_REPO = "google/medgemma-27b-text-it"
 
+tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO)
+# Choose best dtype available
+dtype = (torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported()
+         else torch.float16 if torch.cuda.is_available()
+         else torch.float32)
+pipe = pipeline(
+    task="text-generation",
+    model=MODEL_REPO,
+    tokenizer=tokenizer,
+    trust_remote_code=True,
+    torch_dtype=dtype,
+    device_map="auto" if torch.cuda.is_available() else None
+)
+pipe.to('cuda')
 
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    messages = [{"role": "system", "content": system_message}]
-
-    for val in history:
-        if val[0]:
-            messages.append({"role": "user", "content": val[0]})
-        if val[1]:
-            messages.append({"role": "assistant", "content": val[1]})
+# Cancellation event for streaming
+cancel_event = threading.Event()
+def cancel_generation():
+    cancel_event.set()
+    return "Generation cancelled."
 
-    messages.append({"role": "user", "content": message})
+# Streaming chat response without web search
+@spaces.GPU
+def chat_response(user_msg, chat_history, max_tokens, temperature, top_k, top_p, repetition_penalty):
+    cancel_event.clear()
+    history = list(chat_history or [])
+    history.append({"role": "user", "content": user_msg})
+    # Build prompt
+    prompt = ""
+    for msg in history:
+        prefix = "User: " if msg["role"] == "user" else "Assistant: "
+        prompt += prefix + msg["content"].strip() + "\n"
+    if not prompt.rstrip().endswith("Assistant:"):
+        prompt += "Assistant:"
+    
+    # Set up streamer
+    streamer = TextIteratorStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+    t = threading.Thread(
+        target=pipe,
+        args=(prompt,),
+        kwargs={
+            'max_new_tokens': max_tokens,
+            'temperature': temperature,
+            'top_k': top_k,
+            'top_p': top_p,
+            'repetition_penalty': repetition_penalty,
+            'streamer': streamer,
+            'return_full_text': False,
+        }
+    )
+    t.start()
 
+    # Stream tokens into chat history
     response = ""
-
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
-
+    history.append({"role": "assistant", "content": ""})
+    for token in streamer:
+        if cancel_event.is_set():
+            break
         response += token
-        yield response
-
-
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a helpful medical assistant.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
+        history[-1]["content"] = response
+        yield history
 
+    t.join()
+    gc.collect()
+    yield history
 
-if __name__ == "__main__":
+# Build Gradio interface
+with gr.Blocks(title="MedGemma Chat") as demo:
+    gr.Markdown("## Chat with Google/MedGemma-27B")
+    chat = gr.Chatbot()
+    txt = gr.Textbox(placeholder="Type your message and press Enter...")
+    with gr.Row():
+        max_tok = gr.Slider(64, 4096, value=1024, step=64, label="Max Tokens")
+        temp = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature")
+        k = gr.Slider(1, 100, value=50, step=1, label="Top-K")
+        p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-P")
+        rp = gr.Slider(1.0, 2.0, value=1.2, step=0.1, label="Repetition Penalty")
+        cancel_btn = gr.Button("Cancel Generation")
+    cancel_btn.click(cancel_generation)
+    txt.submit(
+        fn=chat_response,
+        inputs=[txt, chat, max_tok, temp, k, p, rp],
+        outputs=chat
+    )
     demo.launch()