Update app.py

app.py

@@ -1,121 +1,83 @@
-import torch
-torch._dynamo.config.disable = True
-from collections.abc import Iterator
-from transformers import (
-    Gemma3ForConditionalGeneration,
-    TextIteratorStreamer,
-    Gemma3Processor,
-    Gemma3nForConditionalGeneration,
-)
-import gradio as gr
-import os
-import spaces
+import os, torch, gradio as gr, spaces
+from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
 
-# Load environment variables
-model_3n_id = os.getenv("MODEL_3N_ID", "JDhruv14/merged_model")
+MODEL_ID = os.getenv("MODEL_ID", "JDhruv14/merged_model")
 
-# Load model and processor
-model_3n = Gemma3nForConditionalGeneration.from_pretrained(
-    model_3n_id,
-    dtype=torch.bfloat16,
+# Load once (CPU until first call; device_map will move to GPU on first run)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
     device_map="auto",
-    attn_implementation="eager"
+    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else "auto",
+    trust_remote_code=True,
 )
-input_processor = Gemma3Processor.from_pretrained(model_3n_id)
 
-def infer_text(messages, max_new_tokens=300, temperature=1.0, top_p=0.95, top_k=64, repetition_penalty=1.1):
-    chat_template = []
-    for turn in messages:
-        if turn[0]:
-            chat_template.append({"role": "user", "content": [{"type": "text", "text": turn[0]}]})
-        if turn[1]:
-            chat_template.append({"role": "assistant", "content": [{"type": "text", "text": turn[1]}]})
-    chat_template.append({"role": "assistant", "content": [{"type": "text", "text": ""}]})
+def _msgs_from_history(history, system_text):
+    msgs = []
+    if system_text:
+        msgs.append({"role": "system", "content": system_text})
+    for user, assistant in history:
+        if user:
+            msgs.append({"role": "user", "content": user})
+        if assistant:
+            msgs.append({"role": "assistant", "content": assistant})
+    return msgs
 
-    inputs = input_processor.apply_chat_template(
-        chat_template,
-        add_generation_prompt=True,
-        tokenize=True,
-        return_dict=True,
-        return_tensors="pt",
-    ).to(device=model_3n.device, dtype=torch.bfloat16)
+def _eos_ids(tok):
+    ids = {tok.eos_token_id}
+    im_end = tok.convert_tokens_to_ids("<|im_end|>")
+    if im_end is not None:
+        ids.add(im_end)
+    return list(ids)
 
-    with torch.no_grad():
-        output_tokens = model_3n.generate(
-            **inputs,
-            max_new_tokens=max_new_tokens,
-            temperature=temperature,
-            top_p=top_p,
-            top_k=top_k,
-            repetition_penalty=repetition_penalty,
-            do_sample=True,
-        )
+@spaces.GPU(duration=120)  # REQUIRED for ZeroGPU; remove if using standard GPU hardware
+def chat_fn(message, history, system_text, temperature, top_p, max_new, min_new):
+    msgs = _msgs_from_history(history, system_text) + [{"role": "user", "content": message}]
+    prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
 
-    generated_text = input_processor.batch_decode(
-        output_tokens[:, inputs['input_ids'].shape[1]:], skip_special_tokens=True
-    )[0]
-    return generated_text.strip()
+    gen_cfg = GenerationConfig(
+        do_sample=True,
+        temperature=float(temperature),
+        top_p=float(top_p),
+        max_new_tokens=int(max_new),
+        min_new_tokens=int(min_new),
+        repetition_penalty=1.02,
+        no_repeat_ngram_size=3,
+        eos_token_id=_eos_ids(tokenizer),
+        pad_token_id=tokenizer.eos_token_id,
+    )
+    with torch.no_grad():
+        out = model.generate(**inputs, generation_config=gen_cfg)
 
-@spaces.GPU()
-def gradio_fn(message, history):
-    response = infer_text(history + [(message, None)])
-    return response
+    # slice off the prompt so we show only the assistant reply
+    new_tokens = out[:, inputs["input_ids"].shape[1]:]
+    reply = tokenizer.batch_decode(new_tokens, skip_special_tokens=True)[0].strip()
+    return reply
 
-with gr.Blocks(css="""
-    .gradio-container {
-        max-width: 600px;
-        margin: auto;
-        padding: 20px;
-        font-family: sans-serif;
-        position: relative;
-        }
-    .chatbot {
-        height: 500px !important;
-        overflow-y: auto;
-        }
-    .corner {
-       position: fixed;
-       bottom: 2px;
-       z-index: 9999;
-       pointer-events: none;
-        } 
-    #left { left: 2px; }
-    #right { right: 2px; }
-    .corner img {
-       height: 500px;  /* fixed height */
-       width: auto;    /* auto to keep aspect ratio */
-        }
-    
-    """) as demo:
+with gr.Blocks() as demo:
     gr.Markdown(
-    """
-        <div style='text-align: center; padding: 10px;'>
-        <h1 style='font-size: 2.2em; margin-bottom: 0.2em;'>🤖 <span style='color: #4F46E5;'>kRISHNA.ai</span></h1>
-        <p style='font-size: 1.1em; color: #555;'>5000-Years of Ancient WISDOM with Modern AI ✨</p>
-        </div>
-    """,
-    elem_id="header"
+        "<h1 style='text-align:center'>Gita Assistant (Qwen2.5-3B Fine-tuned)</h1>"
+        "<p style='text-align:center'>Ask in English / हिंदी / ગુજરાતી. The assistant cites verses when relevant.</p>"
     )
-    chat = gr.ChatInterface(
-        fn=gradio_fn,
-        examples=[
-            "Hello!",
-            "How can I overcome fear of failure?",
-            "How do I forgive someone who hurt me deeply?",
-            "What can I do to stop overthinking?"
-        ],
-        chatbot=gr.Chatbot(elem_classes="chatbot"),
-        theme="compact",
+    system_box = gr.Textbox(
+        value="Reply in the user’s language with 2–3 concise points (200–400 words); cite Gita verses when relevant.",
+        label="System prompt",
     )
-    gr.HTML(f"""
-      <div id="left" class="corner">
-        <img src="https://huggingface.co/spaces/p2kalita/kRISHNA.ai/resolve/main/assets/Arjun.png" alt="Arjun">
-      </div>
-      <div id="right" class="corner">
-        <img src="https://huggingface.co/spaces/p2kalita/kRISHNA.ai/resolve/main/assets/Krishna.png" alt="Krishna">
-      </div>
-    """)
+    temperature = gr.Slider(0.1, 1.2, value=0.7, step=0.05, label="temperature")
+    top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="top_p")
+    max_new = gr.Slider(64, 1024, value=512, step=16, label="max_new_tokens")
+    min_new = gr.Slider(0, 512, value=160, step=8, label="min_new_tokens")
 
+    chat = gr.ChatInterface(
+        fn=lambda m, h: chat_fn(m, h, system_box.value, temperature.value, top_p.value, max_new.value, min_new.value),
+        title=None,
+        additional_inputs=[system_box, temperature, top_p, max_new, min_new],
+        retry_btn="Regenerate",
+        undo_btn="Undo Last",
+        clear_btn="Clear",
+        queue=True,  # queue is recommended (and required for ZeroGPU concurrency)
+    )
 
 if __name__ == "__main__":
     demo.launch()