sdlm/models/utils.py

import os
from typing import Optional

import torch
from peft import LoraConfig, TaskType, get_peft_model
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from transformers.models.mistral import MistralConfig, MistralForCausalLM

from .ar_warp.ar_warper import GARDiffusionLM
from .cdcd.ar_warper import CDCDGARRobertaForDiffusionLM
from .cdcd.positionwise_warper_model import (
    PositionwiseCDCDRobertaConfig,
    PositionwiseCDCDRobertaForDiffusionLM,
)
from .cdcd.tokenwise_warper_model import TokenwiseCDCDRobertaForDiffusionLM
from .cdcd.warper_model import CDCDRobertaConfig, CDCDRobertaForDiffusionLM
from .confidence_tracker.confidence_tracker_model import (
    ConfidenceTrackerRobertaDiffusionLM,
)
from .llama.configuration_llama import LlamaDiffusionConfig
from .llama.modeling_llama import LlamaForDiffusionLM, LlamaForSeq2SeqLM
from .mistral.configuration_mistral import (
    CDCDMistralDiffusionConfig,
    MistralDiffusionConfig,
)
from .mistral.modeling_mistral import (
    CDCDMistralForDiffusionLM,
    MistralForDiffusionLM,
    MistralForSeq2SeqLM,
)
from .mixins.modeling_mixin import CDCDDiffusionModelMixin
from .roberta.configuration_roberta import RobertaDiffusionConfig
from .roberta.modeling_roberta import RobertaForDiffusionLM

def model_config_helper(
    model_name_or_path: str,
    use_model: str = "cdcd",
    is_diffusion: bool = True,
    conditional_generation: Optional[str] = None,
):
    if "llama" in model_name_or_path.lower():
        if conditional_generation == "seq2seq" and not is_diffusion:
            return LlamaDiffusionConfig, LlamaForSeq2SeqLM
        return LlamaDiffusionConfig, LlamaForDiffusionLM
    if "mistral" in model_name_or_path.lower():
        if conditional_generation == "seq2seq" and not is_diffusion:
            return MistralDiffusionConfig, MistralForSeq2SeqLM
        if conditional_generation is None and not is_diffusion:
            return MistralConfig, MistralForCausalLM
        if use_model == "cdcd":
            return CDCDMistralDiffusionConfig, CDCDMistralForDiffusionLM
        return MistralDiffusionConfig, MistralForDiffusionLM
    if "roberta" in model_name_or_path and use_model == "cdcd":
        return CDCDRobertaConfig, CDCDRobertaForDiffusionLM
    elif "roberta" in model_name_or_path and use_model == "tokenwise_cdcd":
        return CDCDRobertaConfig, TokenwiseCDCDRobertaForDiffusionLM
    elif "roberta" in model_name_or_path and use_model == "positionwise_cdcd":
        return PositionwiseCDCDRobertaConfig, PositionwiseCDCDRobertaForDiffusionLM
    elif "roberta" in model_name_or_path and use_model == "confidence":
        return RobertaDiffusionConfig, ConfidenceTrackerRobertaDiffusionLM
    elif "roberta" in model_name_or_path:
        print(
            f"Using RobertaDiffusionConfig and RobertaForDiffusionLM for {model_name_or_path}"
        )
        return RobertaDiffusionConfig, RobertaForDiffusionLM
    elif "roberta" in model_name_or_path and use_model == "cdcdgar":
        return CDCDRobertaConfig, CDCDGARRobertaForDiffusionLM
    # default to mistral
    if use_model == "cdcd":
        print(
            f"Using CDCDMistralDiffusionConfig and CDCDMistralForDiffusionLM for {model_name_or_path}"
        )
        return CDCDMistralDiffusionConfig, CDCDMistralForDiffusionLM
    print(
        f"Using MistralDiffusionConfig and MistralForDiffusionLM for {model_name_or_path}"
    )
    return MistralDiffusionConfig, MistralForDiffusionLM


def is_cdcd_check(model):
    return (
        isinstance(model, CDCDDiffusionModelMixin)
        or isinstance(model, CDCDMistralForDiffusionLM)
        or isinstance(model, CDCDRobertaForDiffusionLM)
        or isinstance(model, TokenwiseCDCDRobertaForDiffusionLM)
        or isinstance(model, PositionwiseCDCDRobertaForDiffusionLM)
        or isinstance(model, GARDiffusionLM)
        or isinstance(model, CDCDGARRobertaForDiffusionLM)
    )


def is_tokenwise_cdcd_check(model):
    return isinstance(model, TokenwiseCDCDRobertaForDiffusionLM) or isinstance(
        model, PositionwiseCDCDRobertaForDiffusionLM
    )


def freeze(module):
    for param in module.parameters():
        param.requires_grad = False


def get_torch_dtype(training_args):
    torch_dtype = torch.float32
    if training_args.bf16:
        torch_dtype = torch.bfloat16
    elif training_args.fp16:
        torch_dtype = torch.float16
    return torch_dtype


def load_model(model_args, data_args, training_args, diffusion_args, logger):
    config_kwargs = {
        "cache_dir": model_args.cache_dir,
        "revision": model_args.model_revision,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    cfg_cls, model_cls = model_config_helper(
        model_args.model_name_or_path,
        use_model=model_args.use_model,
        is_diffusion=diffusion_args.num_diffusion_steps > 0,
        conditional_generation=data_args.conditional_generation,
    )
    config = cfg_cls.from_pretrained(
        model_args.model_name_or_path,
        self_condition=diffusion_args.self_condition,
        self_condition_zeros_after_softmax=diffusion_args.self_condition_zeros_after_softmax,
        deepmind_conditional=diffusion_args.deepmind_conditional,
        classifier_free_simplex_inputs=diffusion_args.classifier_free_simplex_inputs,
        classifier_free_uncond_input=diffusion_args.classifier_free_uncond_input,
        self_condition_mlp_projection=diffusion_args.self_condition_mlp_projection,
        self_condition_mix_before_weights=diffusion_args.self_condition_mix_before_weights,
        self_condition_mix_logits_before_weights=diffusion_args.self_condition_mix_logits_before_weights,
        empty_token_be_mask=diffusion_args.empty_token_be_mask,
        is_causal=model_args.is_causal,
        mask_padding_in_loss=training_args.mask_padding_in_loss,
        padding_side=model_args.tokenizer_padding_side,
        token=os.environ.get("HF_TOKEN", None),
        **config_kwargs,
    )
    tokenizer_kwargs = {
        "cache_dir": model_args.cache_dir,
        "use_fast": model_args.use_fast_tokenizer,
        "revision": model_args.model_revision,
        "padding_side": model_args.tokenizer_padding_side,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.tokenizer_name,
            token=os.environ.get("HF_TOKEN", None),
            **tokenizer_kwargs,
        )
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path,
            token=os.environ.get("HF_TOKEN", None),
            **tokenizer_kwargs,
        )
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
        )

    try:
        tokenizer.add_eos_token = True
    except AttributeError:
        # roberta does not have this
        pass

    if model_args.model_name_or_path and not model_args.from_scratch:
        model = model_cls.from_pretrained(
            model_args.model_name_or_path,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            config=config,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
            torch_dtype=get_torch_dtype(training_args),
            token=os.environ.get("HF_TOKEN", None),
            attn_implementation="flash_attention_2"
            if model_args.use_flash_attention2
            else "eager",
        ).to("cuda")
        if model_args.freeze_embedding:
            model.get_input_embeddings().requires_grad = False
        if model_args.freeze_model:
            freeze(model)
    else:
        logger.warning("Training new model from scratch")
        model = model_cls._from_config(config)
        model.init_weights()

    if not tokenizer.pad_token_id:
        tokenizer.add_special_tokens({"pad_token": "[PAD]"})

    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
    # on a small vocab and want a smaller embedding size, remove this test.
    vocab_size = model.get_input_embeddings().weight.shape[0]
    if len(tokenizer) > vocab_size:
        model.resize_token_embeddings(len(tokenizer))
        model.config.pad_token_id = tokenizer.pad_token_id

    # if peft, apply it here
    if model_args.use_lora:
        peft_config = LoraConfig(
            task_type=TaskType.CAUSAL_LM,
            inference_mode=False,
            r=model_args.lora_rank,
            lora_alpha=model_args.lora_alpha,
            lora_dropout=model_args.lora_dropout,
        )
        # we just peft the internal model.
        # a little hacky, remove the task type wrapper class
        # TODO: does this cook anything?
        model.model = get_peft_model(model.model, peft_config).base_model

    # apply liger monkey patching
    if model_args.use_liger_kernel:
        from liger_kernel.transformers import apply_liger_kernel_to_mistral
        apply_liger_kernel_to_mistral()

    return tokenizer, model


def load_classifier(classifier_model_name_or_path: str):
    tokenizer = AutoTokenizer.from_pretrained(classifier_model_name_or_path)
    model = AutoModelForSequenceClassification.from_pretrained(
        classifier_model_name_or_path,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
    ).eval()
    model.gradient_checkpointing_enable()
    # NOTE: for quick testing (reduce vram req)
    # model.model.layers = torch.nn.ModuleList([model.model.layers[0]])
    freeze(model)
    # from liger_kernel.transformers import apply_liger_kernel_to_mistral
    # apply_liger_kernel_to_mistral()
    return tokenizer, model


def check_tokenizer_equal(tokenizer1, tokenizer2):
    # check class
    assert tokenizer1.__class__ is tokenizer2.__class__
    # check vocab size
    assert tokenizer1.vocab_size == tokenizer2.vocab_size
    # check special tokens size
    assert len(tokenizer1.special_tokens_map) == len(tokenizer2.special_tokens_map)
    # check special tokens
    for special_token in ("bos", "eos", "unk", "pad"):
        attr = f"{special_token}_token_id"
        assert getattr(tokenizer1, attr) == getattr(tokenizer2, attr)
    # full decoding check
    for i in range(tokenizer1.vocab_size + len(tokenizer1.special_tokens_map)):
        decoded1 = tokenizer1.decode([i])
        decoded2 = tokenizer2.decode([i])
        assert decoded1 == decoded2