Update app.py

app.py

@@ -3,8 +3,7 @@ import gradio as gr
 import requests
 import inspect
 import pandas as pd
-from smolagents import tool, Tool, CodeAgent, DuckDuckGoSearchTool, VisitWebpageTool, SpeechToTextTool, FinalAnswerTool
-#from smolagents import tool, Tool, CodeAgent, DuckDuckGoSearchTool, HfApiModel, VisitWebpageTool, SpeechToTextTool, FinalAnswerTool
+from smolagents import tool, Tool, CodeAgent, DuckDuckGoSearchTool, HfApiModel, VisitWebpageTool, SpeechToTextTool, FinalAnswerTool
 from dotenv import load_dotenv
 import heapq
 from collections import Counter
@@ -15,8 +14,7 @@ from langchain_community.tools.tavily_search import TavilySearchResults
 from langchain_community.document_loaders import WikipediaLoader
 from langchain_community.utilities import WikipediaAPIWrapper
 from langchain_community.document_loaders import ArxivLoader
-from langchain_community.llms import HfApiModel
-
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 # (Keep Constants as is)
 # --- Constants ---
@@ -184,12 +182,22 @@ class VideoTranscriptionTool(Tool):
 
 class BasicAgent:
     def __init__(self):
-        token = os.environ.get("HF_API_TOKEN")
-        self.model = HfApiModel(
-            "google/gemini-2.5-flash",
-            temperature=0.1,
-            token=token
+        # Configuration for Qwen2.5-Coder-32B-Instruct
+        model_name = "Qwen/Qwen2.5-Coder-32B-Instruct"
+
+        # Load the model and tokenizer directly using Hugging Face Transformers
+        # This will download the model weights and load them onto your device (GPU if available)
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            torch_dtype="auto", # Uses bfloat16 or float16 if supported, otherwise float32
+            device_map="auto" # Automatically maps model layers to available devices (e.g., GPU(s), CPU)
         )
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+        # Note: You won't use 'token' for direct Hugging Face model loading unless
+        # the model is private and requires authentication. For public models like Qwen,
+        # it's usually not needed for loading.
+        token = os.environ.get("HF_API_TOKEN") # This line might not be needed now
 
         search_tool = DuckDuckGoSearchTool()
         wiki_search_tool = WikiSearchTool()
@@ -210,8 +218,49 @@ If the answer is a string, do not use articles or abbreviations (e.g., for citie
 If the answer is a comma-separated list, apply the above rules for each element based on whether it is a number or a string.
 """
 
+        # Here, you might need to adapt how CodeAgent expects the model.
+        # If CodeAgent is built to work with LangChain's LLM instances,
+        # you'll need to wrap your Qwen model with a custom LangChain LLM
+        # or adjust CodeAgent to accept direct Hugging Face model/tokenizer.
+        # For simplicity, if CodeAgent can take a callable for 'model',
+        # you could define a simple wrapper.
+        # Otherwise, you might need to write a custom LangChain LLM class.
+
+        # For demonstration, let's assume CodeAgent can handle a custom callable
+        # that performs inference using your loaded model and tokenizer.
+        # This is a simplification and might require adjustment to CodeAgent.
+        class CustomQwenLLM:
+            def __init__(self, model, tokenizer):
+                self.model = model
+                self.tokenizer = tokenizer
+
+            def __call__(self, prompt: str) -> str:
+                messages = [
+                    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
+                    {"role": "user", "content": prompt}
+                ]
+                text = self.tokenizer.apply_chat_template(
+                    messages,
+                    tokenize=False,
+                    add_generation_prompt=True
+                )
+                model_inputs = self.tokenizer([text], return_tensors="pt").to(self.model.device)
+
+                generated_ids = self.model.generate(
+                    **model_inputs,
+                    max_new_tokens=512,
+                    do_sample=True, # Added for better response quality
+                    temperature=0.7 # Added for better response quality
+                )
+                input_length = model_inputs.input_ids.shape[1]
+                generated_text = self.tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0]
+                return generated_text
+
+        self.llm_for_agent = CustomQwenLLM(self.model, self.tokenizer)
+
+
         self.agent = CodeAgent(
-            model=self.model,
+            model=self.llm_for_agent, # Pass the custom wrapper
             tools=[search_tool, wiki_search_tool, str_reverse_tool, keywords_extract_tool, speech_to_text_tool, visit_webpage_tool, final_answer_tool, parse_excel_to_json, video_transcription_tool],
             add_base_tools=True
         )
@@ -223,6 +272,7 @@ If the answer is a comma-separated list, apply the above rules for each element
         print(f"Agent returning answer: {answer}")
         return answer
 
+
 def run_and_submit_all( profile: gr.OAuthProfile | None):
     """
     Fetches all questions, runs the BasicAgent on them, submits all answers,