import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)
model.eval()

def blend_generate(prompt, wa, wb):
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)

    with torch.no_grad():
        output_a = model_a(input_ids)
        output_b = model_b(input_ids)

    logits_a = output_a.logits[:, -1, :]
    logits_b = output_b.logits[:, -1, :]

    # Weighted sum of raw logits (before softmax)
    blended_logits = wa * logits_a + wb * logits_b

    # Apply softmax safely to get valid probability distribution
    probs = torch.softmax(blended_logits, dim=-1)

    # Sample token from valid probability distribution
    token = torch.multinomial(probs, 1)
    next_token_id = token.item()
    next_token = tokenizer.decode([next_token_id])

    return next_token

with gr.Blocks() as demo:
    gr.Markdown("## Blended Prompt Chat (TinyLlama)")
    sysA = gr.Textbox(label="System Prompt A", value="You are assistant A.")
    sysB = gr.Textbox(label="System Prompt B", value="You are assistant B.")
    wA = gr.Slider(-5, 5, value=1.0, step=0.1, label="Weight A")
    wB = gr.Slider(-5, 5, value=1.0, step=0.1, label="Weight B")
    user_msg = gr.Textbox(label="User Message")
    temp = gr.Slider(0.1, 2.0, value=1.0, step=0.1, label="Temperature")
    top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
    max_tokens = gr.Slider(1, 200, value=100, step=1, label="Max New Tokens")
    output = gr.Textbox(label="Response")

    btn = gr.Button("Generate")
    btn.click(
        blend_generate,
        [sysA, sysB, wA, wB, user_msg, max_tokens, temp, top_p],
        output,
        show_progress=True,
    )

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