LOAD_THIS_MODEL.py

from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig, BitsAndBytesConfig




import torch
n_ahead_talk_global = 4
n_passes_global = 2
n_ahead_global = 8
n_examples = 0

def model_init(params):
    original = False
    if params is None:
        params = {}
    else:
        params = params.params
    # save params to file
    n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
    n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
    n_passes = params.get("n_passes", n_passes_global if not original else 1)
    gumbel_temperature = params.get("gumbel_temperature", 1)
    use_start_thought_token = params.get("use_start_thought_token", True)
    use_end_thought_token = params.get("use_end_thought_token", True)
    include_policy_loss = params.get("include_policy_loss", True)
    gumbel_detach = params.get("gumbel_detach", True)
    merged_talk_heads = params.get("merged_talk_heads", True)
    residual_think_head = params.get("residual_think_head", False)
    optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)

    model_id = "LeroyDyer/_Spydaz_Web_AI_V2_Aligned"
    tokenizer_id = model_id
    print("Loading model")

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        max_thoughts=n_ahead + n_ahead_talk + 1,
        merged_talk_heads=merged_talk_heads,
        merged_lm_and_talk_heads=False,
        merged_lm_and_think_heads=True,
        use_concat_talk_head=True,
        use_shallow_think=True,
        use_shallow_talk=False,
        use_complex_think_head=False,
        use_complex_talk_head=True,
        use_weighted_talk_head=True,
        trust_remote_code=True,
        device_map="auto",
    )
    print("Loaded model")

    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding_side="right")
    tokenizer.pad_token_id = tokenizer.eos_token_id

    special_tokens_to_add = []
    if model.use_start_thought_token:
        special_tokens_to_add.append("<|startthought|>")
    if model.use_end_thought_token:
        special_tokens_to_add.append("<|endthought|>")
    if special_tokens_to_add:
        tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
        model.resize_token_embeddings(len(tokenizer))
    model.tokenizer = tokenizer
    for name, module in model.named_modules():
        if "embed" in name:
            print(module, flush=True)

    model.gumbel_detach = gumbel_detach
    model.include_policy_loss = include_policy_loss
    model.use_end_thought_token = use_end_thought_token
    model.use_start_thought_token = use_start_thought_token
    model.n_ahead = n_ahead
    model.n_ahead_talk = n_ahead_talk
    model.n_passes = n_passes
    model.residual_think_head = residual_think_head
    model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
    model.gumbel_temperature = gumbel_temperature
    model.original_mode = original
    model.config_params = params
    model.run_start = int(time.time())
    model.train()
    return model,tokenizer

model,tokenizer = model_init(None)
tokenizer.save_pretrained("IModel")
model.save_pretrained("IModel")





from datasets import load_dataset

### DATA SET FOR TRAINING


alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""


EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
def formatting_prompts_func(examples):
    instructions = examples["instruction"]
    inputs       = examples["input"]
    outputs      = examples["output"]
    texts = []
    for instruction, input, output in zip(instructions, inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }
pass

from datasets import load_dataset
dataset = load_dataset("gate369/Alpaca-Star", split = "train[:2000]")
dataset = dataset.shuffle(seed=3704)
dataset = dataset.map(formatting_prompts_func, batched = True,)