Qwen2-1.5B-wina

use the following script to load the model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# --- Critical: Import the custom functions needed to build the model ---
from wina_utils import apply_wina_to_model

# --- Configuration ---
# The path to your locally saved model, or its ID on the Hugging Face Hub
MODEL_PATH = "moelanoby/Qwen2-1.5B-wina"
# Example after upload: 
SPARSITY_LEVEL = 0.65 # THIS MUST STAY 0.65 OR 65% ANYTHING ELSE MIGHT CAUSE ERRORS

def load_custom_wina_model(model_path: str, sparsity: float):
    """
    Loads a custom WINA-modified model correctly.
    """
    print(f"Loading WINA model from: {model_path}")

    # --- Step 1: Load the tokenizer ---
    # trust_remote_code=True is needed here because the model folder contains wina_utils.py
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

    # --- Step 2: Load the BASE model architecture ---
    # We load the original model structure first. It will have the standard MLP layers.
    # trust_remote_code=True allows it to execute the original model's code if needed.
    print("Loading the base model architecture...")
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.float16,
        device_map="auto",
        trust_remote_code=True
    )

    # --- Step 3: Manually apply the WINA transformation to the architecture ---
    # This swaps the standard MLP layers with our WinaMLP layers, creating the
    # correct structure in memory to receive the saved weights.
    print("Applying WINA transformation to the loaded architecture...")
    apply_wina_to_model(model, sparsity_level=sparsity)

    print("\n" + "="*50)
    print("SUCCESS: Custom WINA model loaded correctly.")
    print("="*50)
    return model, tokenizer

if __name__ == "__main__":
    wina_model, wina_tokenizer = load_custom_wina_model(MODEL_PATH, SPARSITY_LEVEL)

    # You can now use the model for inference
    print("\nVerifying model architecture (you should see 'WinaMLP' layers):")
    print(wina_model)