Indus 3.0 Demo

app.py

@@ -1,78 +1,63 @@
-import os, tarfile
+import os
+import tarfile
 from huggingface_hub import hf_hub_download
 from transformers import AutoTokenizer, AutoModelForCausalLM
 import gradio as gr
 
-# At startup, download and extract model
+# 1. Download the tar.gz archive
 archive_path = hf_hub_download(
     repo_id="SamOrion/Llama_3.2_3b_Hindi_Pruned",
     filename="llama-3.2-3b-hindi-pruned.tar.gz",
-    repo_type="model"
+    repo_type="model",
 )
+
+# 2. Extract into './model', stripping the top-level folder
 extract_dir = "./model"
 os.makedirs(extract_dir, exist_ok=True)
 
-# Extract and list contents
 with tarfile.open(archive_path, "r:gz") as tar:
-    tar.extractall(path=extract_dir)
-    print("Archive members:")
     for member in tar.getmembers():
-        print(f"  {member.name}")
+        # Skip the first path component (the folder name)
+        parts = member.name.split("/", 1)
+        if len(parts) == 2:
+            member.name = parts[1]
+            tar.extract(member, path=extract_dir)
 
-# Check what files were actually created
-print("\nExtracted files:")
-for root, dirs, files in os.walk(extract_dir):
-    for file in files:
-        full_path = os.path.join(root, file)
-        print(f"  {full_path}")
-        
-# Check if config.json exists and show its contents
+# 3. Verify that config.json is at ./model/config.json
 config_path = os.path.join(extract_dir, "config.json")
-if os.path.exists(config_path):
-    import json
-    with open(config_path, 'r') as f:
-        config = json.load(f)
-    print(f"\nconfig.json contents: {config}")
-else:
-    print(f"\nconfig.json not found at {config_path}")
-
-model_path = None
-for root, dirs, files in os.walk(extract_dir):
-    if "config.json" in files:
-        model_path = root
-        break
-
-if model_path is None:
-    raise FileNotFoundError("config.json not found after extraction")
+if not os.path.isfile(config_path):
+    raise FileNotFoundError(f"config.json not found in {extract_dir}")
 
-# Load from the extracted folder
-tokenizer = AutoTokenizer.from_pretrained(model_path)
-model     = AutoModelForCausalLM.from_pretrained(
-    model_path,
+# 4. Load tokenizer and model straight from './model'
+tokenizer = AutoTokenizer.from_pretrained(extract_dir)
+model = AutoModelForCausalLM.from_pretrained(
+    extract_dir,
     torch_dtype="auto",
     device_map="auto",
-    low_cpu_mem_usage=True
+    low_cpu_mem_usage=True,
 )
 
+# 5. Define chat function using OpenAI-style messages
 def chat_fn(prompt, history):
     history = history or []
     history.append({"role": "user", "content": prompt})
+
     inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
     outputs = model.generate(**inputs, max_new_tokens=100)
     response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
     history.append({"role": "assistant", "content": response})
     return history, ""
 
+# 6. Build Gradio interface
 with gr.Blocks() as demo:
-    gr.Markdown("## 🌐 Indus 3.0 Demo")
+    gr.Markdown("## 🌐 Indus 3.0 Hindi LLM Demo")
     chat = gr.Chatbot(type="messages")
-    msg = gr.Textbox(placeholder="Type here...")
+    msg = gr.Textbox(placeholder="Type here…")
     clear = gr.Button("Clear")
     msg.submit(chat_fn, [msg, chat], [chat, msg])
-    clear.click(lambda: None, None, chat)
+    clear.click(lambda: ([], ""), None, [chat, msg])
 
+# 7. Launch without `share=True` on Spaces
 if __name__ == "__main__":
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860
-    )
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_api=False, api_open=False)