Create app.py

app.py

@@ -0,0 +1,84 @@
+# app.py
+import gradio as gr, torch, threading
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
+
+# --- Model config ---
+MODEL_ID = "md-nishat-008/TigerLLM-1B-it"  # Bangla-optimized, instruction-tuned
+DEVICE = 0 if torch.cuda.is_available() else -1
+
+SYSTEM_BN = "তুমি একজন সহায়ক বাংলা সহকারী। সংক্ষিপ্ত, স্পষ্ট এবং ভদ্রভাবে উত্তর দাও।"
+SYSTEM_HI = "तुम एक सहायक हिन्दी सहायक हो। स्पष्ट और संक्षिप्त उत्तर दो। (यह मॉडल बांग्ला-केंद्रित है, इसलिए हिन्दी गुणवत्ता सीमित हो सकती है।)"
+
+# --- Load once at startup (reduces first-turn stall) ---
+tok = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    torch_dtype=torch.float32,   # CPU-friendly on free Spaces
+    low_cpu_mem_usage=True,
+    device_map="auto" if DEVICE != -1 else None,
+)
+if DEVICE != -1:
+    model = model.to("cuda")
+
+def build_prompt(history, lang):
+    sys = SYSTEM_BN if lang == "bn" else SYSTEM_HI
+    parts = [f"<<SYS>> {sys} </SYS>"]
+    for user, bot in history:
+        if user:
+            parts.append(f"<<USER>> {user}")
+        if bot:
+            parts.append(f"<<ASSISTANT>> {bot}")
+    parts.append("<<ASSISTANT>>")
+    return "\n".join(parts)
+
+def on_user(user_text, chat_state):
+    if not user_text.strip():
+        return "", chat_state
+    return "", chat_state + [(user_text, None)]
+
+def on_bot_stream(chat_state, lang_label):
+    # lang_label is "বাংলা" or "हिन्दी"
+    lang = "bn" if lang_label == "বাংলা" else "hi"
+    last_user, _ = chat_state[-1]
+
+    prompt = build_prompt(chat_state, lang)
+    inputs = tok(prompt, return_tensors="pt")
+    if DEVICE != -1:
+        inputs = {k: v.to("cuda") for k, v in inputs.items()}
+
+    streamer = TextIteratorStreamer(tok, skip_prompt=True, skip_special_tokens=True)
+    gen_kwargs = dict(
+        **inputs,
+        max_new_tokens=160,      # keep short for speed
+        do_sample=True,
+        temperature=0.6,
+        top_p=0.9,
+        repetition_penalty=1.05,
+        eos_token_id=tok.eos_token_id,
+        pad_token_id=tok.eos_token_id,
+        streamer=streamer,
+    )
+
+    thread = threading.Thread(target=model.generate, kwargs=gen_kwargs)
+    thread.start()
+
+    partial = ""
+    for new_text in streamer:
+        partial += new_text
+        clean = partial.replace("<<SYS>>", "").replace("<<USER>>", "").replace("<<ASSISTANT>>", "")
+        chat_state[-1] = (last_user, clean[:2000])
+        yield chat_state
+
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("### Suman AI Project — বাংলা / हिंदी")
+    lang = gr.Radio(choices=["বাংলা", "हिन्दी"], value="বাংলা", label="Language")
+    # 'tuples' shows a deprecation warning but works fine on Gradio v5
+    chat = gr.Chatbot(height=420, type="tuples")
+    txt = gr.Textbox(placeholder="বাংলা বা हिन्दी লিখুন…", autofocus=True)
+    send = gr.Button("Send")
+
+    txt.submit(on_user, [txt, chat], [txt, chat]).then(on_bot_stream, [chat, lang], [chat])
+    send.click(on_user, [txt, chat], [txt, chat]).then(on_bot_stream, [chat, lang], [chat])
+
+# Gradio v5: queue() takes no kwargs; just enable and launch
+demo.queue().launch()