from accelerate import Accelerator
from transformers import TextIteratorStreamer
from threading import Thread
from .tree_utils import full_func_head, grab_before_comments

def combine_generation_kwargs(temperature=2.0, max_new_tokens=512, top_p=0.95, repetition_penalty=1.2):
    """
    Combines the generation kwargs into a single dict.
    """
    gen_kwargs = {}
    gen_kwargs["do_sample"] = True
    gen_kwargs["temperature"] = temperature
    gen_kwargs["max_new_tokens"] = max_new_tokens
    gen_kwargs["top_p"] = top_p
    gen_kwargs["repetition_penalty"] = repetition_penalty
    return gen_kwargs


def stream_generation(prompt:str, pipe, gen_kwargs:dict):
    accelerator = Accelerator()
    device = accelerator.device
    """
    Text generation function
    Args:
        prompt (str): The context to start generation from.
        pipe (Pipeline): The pipeline to use for generation (we take the model and tokenizer form it)
        gen_kwargs (dict): The generation kwargs.
    Returns:
        str: The generated text. (it iterates over time)
    """
    # Tokenize the model_context
    model_inputs = pipe.tokenizer(prompt, return_tensors="pt")
    model_inputs.to(device)
    model = pipe.model.to(device) #is this also required?

    # Start generation on a separate thread, so that we don't block the UI. The text is pulled from the streamer
    # in the main thread. Adds timeout to the streamer to handle exceptions in the generation thread.
    streamer = TextIteratorStreamer(pipe.tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=45.0) #IPEX takes a bit on first inference, to avoid an error with the empty queue timeout on the first time, we just wait longer.
    generate_kwargs = dict(model_inputs, streamer=streamer, **gen_kwargs)
    t = Thread(target=pipe.model.generate, kwargs=generate_kwargs)
    t.start()

    # Pull the generated text from the streamer, and update the model output.
    model_output = ""
    for new_text in streamer:
        # print("step", end="")
        model_output += new_text
        yield model_output
    streamer.on_finalized_text("stream reached the end.")
    return model_output #is this ever reached?

def construct_model_context(func_node, prompt=""):
    """
    Constructs the model context from a function node.
    returns: model_context, start_byte
    """
    model_context, start_byte = grab_before_comments(func_node)
    model_context += full_func_head(func_node)
    if prompt != "":
        model_context = "//Title: " + prompt + "\n" + model_context #prepend user prompt/title
        model_context = "//Language: Shadertoy GLSL fragment shader\n" + model_context #prepend system prompt, language hint
    return model_context, start_byte