app.py

# # import gradio as gr
# # from huggingface_hub import InferenceClient

# # """
# # 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("HuggingFaceH4/zephyr-7b-beta")


# # 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]})

# #     messages.append({"role": "user", "content": message})

# #     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

# #         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 friendly Chatbot.", 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)",
# #         ),
# #     ],
# # )


# # if __name__ == "__main__":
# #     demo.launch()

# import torch
# import gradio as gr
# from transformers import AutoModelForCausalLM, AutoTokenizer
# import os

# # Define model names
# MODEL_1_PATH = "./adapter_model.safetensors"  # Local path inside Space
# MODEL_2_NAME = "sarvamai/sarvam-1"  # The base model on Hugging Face Hub

# # Load the tokenizer (same for both models)
# TOKENIZER_NAME = "sarvamai/sarvam-1"
# tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_NAME)
# def fix_checkpoint(model_path):
#     """Fixes the model checkpoint by adjusting mismatched weight dimensions."""
#     checkpoint_file = os.path.join(model_path, "pytorch_model.bin")
#     fixed_checkpoint_file = os.path.join(model_path, "pytorch_model_fixed.bin")

#     if not os.path.exists(checkpoint_file):
#         raise FileNotFoundError(f"Checkpoint file not found at: {checkpoint_file}")

#     print("Loading checkpoint for fixing...")
#     checkpoint = torch.load(checkpoint_file, map_location="cpu")

#     # Adjust weights (truncate the last token if mismatch)
#     if "base_model.model.lm_head.base_layer.weight" in checkpoint:
#         checkpoint["base_model.model.lm_head.base_layer.weight"] = checkpoint["base_model.model.lm_head.base_layer.weight"][:-1]

#     if "base_model.model.lm_head.lora_B.default.weight" in checkpoint:
#         checkpoint["base_model.model.lm_head.lora_B.default.weight"] = checkpoint["base_model.model.lm_head.lora_B.default.weight"][:-1]

#     # Save the fixed checkpoint
#     print("Saving fixed checkpoint...")
#     torch.save(checkpoint, fixed_checkpoint_file)

#     return fixed_checkpoint_file  # Return the new file path

# # Function to load a model
# def load_model(model_choice):
#     if model_choice == "Hugging face dataset":
#         model = AutoModelForCausalLM.from_pretrained("./", torch_dtype=torch.float16, device_map="auto")
#         model.load_adapter(MODEL_1_PATH, "safe_tensors")  # Load safetensors adapter
#     else:
#         model = AutoModelForCausalLM.from_pretrained(MODEL_2_NAME)
#     model.eval()
#     return model

# # Load default model on startup
# current_model = load_model("Hugging face dataset")

# # Chatbot response function
# def respond(message, history, model_choice, max_tokens, temperature, top_p):
#     global current_model
    
#     # Switch model if user selects a different one
#     if (model_choice == "Hugging face dataset" and current_model is not None and current_model.config.name_or_path != MODEL_1_PATH) or \
#        (model_choice == "Proprietary dataset1" and current_model is not None and current_model.config.name_or_path != MODEL_2_NAME):
#         current_model = load_model(model_choice)

#     # Convert chat history to format
#     messages = [{"role": "system", "content": "You are a friendly AI assistant."}]
#     for val in history:
#         if val[0]:
#             messages.append({"role": "user", "content": val[0]})
#         if val[1]:
#             messages.append({"role": "assistant", "content": val[1]})
#     messages.append({"role": "user", "content": message})

#     # Tokenize and generate response
#     inputs = tokenizer.apply_chat_template(messages, tokenize=False)
#     input_tokens = tokenizer(inputs, return_tensors="pt").to("cuda" if torch.cuda.is_available() else "cpu")

#     output_tokens = current_model.generate(
#         **input_tokens,
#         max_new_tokens=max_tokens,
#         temperature=temperature,
#         top_p=top_p,
#         pad_token_id=tokenizer.pad_token_id,
#         eos_token_id=tokenizer.eos_token_id,
#     )

#     response = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
#     return response

# # Define Gradio Chat Interface
# demo = gr.ChatInterface(
#     fn=respond,
#     additional_inputs=[
#         gr.Dropdown(choices=["Hugging face dataset", "Proprietary dataset1"], value="Fine-Tuned Model", label="Select Model"),
#         gr.Slider(minimum=1, maximum=1024, value=256, step=1, label="Max Tokens"),
#         gr.Slider(minimum=0.1, maximum=2.0, 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"),
#     ],
# )

# if __name__ == "__main__":
#     demo.launch()

import torch
import os
from transformers import AutoModelForCausalLM, AutoTokenizer

# Define model and tokenizer paths
MODEL_1_PATH = "Priyanka6/fine-tuning-inference"
TOKENIZER_NAME = "sarvam/sarvam-1"  # Keep this unchanged if tokenizer hasn't changed

def trim_adapter_weights(model_path):
    """
    Trims the last token from the adapter's lm_head.lora_B.default.weight 
    if there is a mismatch with the base model.
    """
    adapter_file = os.path.join(model_path, "adapter_model.safetensors")
    
    if not os.path.exists(adapter_file):
        raise FileNotFoundError(f"Adapter file not found: {adapter_file}")
    
    checkpoint = torch.load(adapter_file, map_location="cpu")

    key_to_trim = "lm_head.lora_B.default.weight"
    
    if key_to_trim in checkpoint:
        original_size = checkpoint[key_to_trim].shape[0]
        expected_size = original_size - 1  # Removing last token
        
        print(f"Trimming {key_to_trim}: {original_size} -> {expected_size}")

        checkpoint[key_to_trim] = checkpoint[key_to_trim][:-1]  # Trim the last row

        # Save the modified adapter
        trimmed_adapter_path = os.path.join(model_path, "adapter_model_trimmed.safetensors")
        torch.save(checkpoint, trimmed_adapter_path)
        return trimmed_adapter_path
    
    return adapter_file

# Before loading the adapter, trim it if necessary
trimmed_adapter_path = trim_adapter_weights(MODEL_1_PATH)

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_NAME)

# Load the model
model = AutoModelForCausalLM.from_pretrained(
    MODEL_1_PATH, torch_dtype=torch.float16, device_map="auto"
)

# Load the trimmed adapter
model.load_adapter(trimmed_adapter_path, "safe_tensors")  

# Chat function
def chat(query):
    inputs = tokenizer(query, return_tensors="pt").to("cuda")
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=100)
    return tokenizer.decode(output[0], skip_special_tokens=True)

# Test the chatbot
if __name__ == "__main__":
    while True:
        query = input("User: ")
        if query.lower() in ["exit", "quit"]:
            break
        response = chat(query)
        print(f"Bot: {response}")