custom_generate/generate.py

import torch
from transformers import Cache, DynamicCache
from transformers.generation.utils import ModelOutput
from typing import Optional, Any

def prepare_inputs_for_generation(
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ):
        input_ids = input_ids[:, cache_position].clone(memory_format=torch.contiguous_format)
        model_inputs = {"cache_position": cache_position, 
                        "past_key_values": None, 
                        "input_ids": input_ids, 
                        "inputs_embeds": None,
                        "attention_mask": attention_mask,
                        }
        if attention_mask is not None and kwargs.get("position_ids") is None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            kwargs["position_ids"] = position_ids
        model_inputs.update({k: v for k, v in kwargs.items() if k not in model_inputs})
        return model_inputs

def update_model_kwargs_for_generation(
        outputs: ModelOutput,
        model_kwargs: dict[str, Any],
        num_new_tokens: int = 1,
    ) -> dict[str, Any]:
        if "token_type_ids" in model_kwargs:
            token_type_ids = model_kwargs["token_type_ids"]
            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1)
        if "attention_mask" in model_kwargs:
            attention_mask = model_kwargs["attention_mask"]
            model_kwargs["attention_mask"] = torch.cat(
                [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
            )
        past_positions = model_kwargs.pop("cache_position")
        new_positions = torch.arange(
            past_positions[-1] + 1, past_positions[-1] + num_new_tokens + 1, dtype=past_positions.dtype
        ).to(past_positions.device)
        model_kwargs["cache_position"] = torch.cat((past_positions, new_positions))
        return model_kwargs


def next_logits_with_cache_update(model, model_kwargs, input_ids):
    """
    Gets the next token logits and updates the KV cache:
    - Runs the model forward pass
    - Extracts logits for the last token
    - Updates the KV cache
    - Returns updated `model_kwargs` and `logits`

    Args:
        model: The language model
        model_kwargs: Model keyword arguments including KV cache
        input_ids: Current input token IDs

    Returns:
        Updated model_kwargs, logits for the next token
    """
    model_inputs = prepare_inputs_for_generation(input_ids, **model_kwargs)
    with torch.no_grad():
        outputs = model(**model_inputs, return_dict=True)

    logits = outputs.logits[:, -1].detach()
    model_kwargs = update_model_kwargs_for_generation(outputs, model_kwargs)
    del outputs
    return model_kwargs, logits


def init_gen(model_kwargs, model, max_new_tokens, bos_token_id):
    """
    Initializes the generation process and prepares the KV cache:
    - Sets up input sequences and model inputs
    - Prepares the KV cache for generation
    - Returns updated `model_kwargs` and `input_ids`

    Args:
        model_kwargs: Model keyword arguments
        model: The language model
        max_new_tokens: Maximum number of new tokens to generate
        bos_token_id: Beginning-of-sequence token ID

    Returns:
        Model keyword arguments and input token IDs
    """
    input_ids = model_kwargs.pop("input_ids")
    model_kwargs["past_key_values"] = None
    model_kwargs["cache_position"] = torch.ones(input_ids.shape[1], dtype=torch.int64, device=input_ids.device).cumsum(0) - 1
    return model_kwargs, input_ids


def _apply_top_k(ps, model):
    """Apply top-k filtering to probabilities."""
    if not hasattr(model, "generation_config") or not hasattr(
        model.generation_config, "top_k"
    ):
        return ps

    top_k = model.generation_config.top_k
    if top_k is None or top_k >= ps.size(-1):
        return ps

    indices_to_remove = ps < torch.topk(ps, top_k)[0][..., -1, None]
    ps[indices_to_remove] = 0.0
    return ps / ps.sum(dim=-1, keepdim=True)


def _apply_top_p(ps, model):
    """Apply top-p (nucleus) filtering to probabilities."""
    if not hasattr(model, "generation_config") or not hasattr(
        model.generation_config, "top_p"
    ):
        return ps

    top_p = model.generation_config.top_p
    if top_p is None or top_p >= 1.0:
        return ps

    sorted_probs, sorted_indices = torch.sort(ps, descending=True)
    cumulative_probs = torch.cumsum(sorted_probs, dim=-1)

    sorted_indices_to_remove = cumulative_probs > top_p
    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
    sorted_indices_to_remove[..., 0] = 0

    indices_to_remove = sorted_indices_to_remove.scatter(
        1, sorted_indices, sorted_indices_to_remove
    )
    ps[indices_to_remove] = 0.0
    return ps / ps.sum(dim=-1, keepdim=True)


def sampling(
    model_kwargs,
    model,
    eos_token_ids,
    pad_token_id,
    bos_token_id,
    do_sample=True,
    max_new_tokens=20,
    temperature=1.0,
):
    """
    Sampling implementation with proper KV caching.

    Args:
        prompts: List of input prompts
        model: The language model
        max_new_tokens: Maximum number of new tokens to generate
        eos_token_ids: List of end-of-sequence token IDs
        pad_token_id: Padding token ID
        bos_token_id: Beginning-of-sequence token ID
        max_new_tokens: Maximum number of new tokens to generate

    Returns:
        Generated sequences, log probabilities, and metadata
    """
    # Initialize the generation process and prepare the KV cache
    model_kwargs, input_ids = init_gen(
        model_kwargs, model, max_new_tokens, bos_token_id
    )
    batch_size, _ = input_ids.shape

    # Keeps track of which sequences are finished and their lengths
    active_seqs = input_ids.new_ones((batch_size, 1), dtype=torch.bool)
    # Modified log probabilities of the sequences
    scores = torch.zeros((batch_size, max_new_tokens), dtype=model.dtype)
    # Unfiltered sequence log probabilities (temperature=1, no sampling processors applied)
    logprobs = torch.zeros((batch_size, max_new_tokens), dtype=model.dtype)

    for i in range(max_new_tokens):
        # Get the next token probabilities and update the KV cache
        model_kwargs, logits = next_logits_with_cache_update(
            model, model_kwargs, input_ids
        )
        # Store original model probabilities (temperature=1, no sampling processors applied)
        model_ps = logits.softmax(-1)

        # Logit processors (temperature, top-k, top-p). We can chain these!
        ps = (logits / temperature).softmax(-1)
        ps = _apply_top_k(ps, model)
        ps = _apply_top_p(ps, model)

        # Sample the next token and gather the log probabilities
        if do_sample:  # Sampling
            next_token_ids = (
                torch.multinomial(ps, 1) * active_seqs + pad_token_id * ~active_seqs
            )
        else:  # Greedy decoding
            next_token_ids = (
                torch.argmax(ps, dim=-1).unsqueeze(-1) * active_seqs
                + pad_token_id * ~active_seqs
            )
        next_token_logprobs = ps.gather(-1, next_token_ids).log()
        next_token_model_logprobs = model_ps.gather(-1, next_token_ids).log()

        input_ids = torch.cat([input_ids, next_token_ids], dim=-1)
        scores[:, i] = (next_token_logprobs * active_seqs).squeeze()
        logprobs[:, i] = (next_token_model_logprobs * active_seqs).squeeze()

        active_seqs &= ~torch.isin(next_token_ids, eos_token_ids)
        if active_seqs.sum() == 0:
            break
    return input_ids.detach().cpu(), scores[:, : i + 1], logprobs[:, : i + 1]


def generate(model, **kwargs):
    """
    Sampling strategy - multinomial sampling with temperature and optional top-k/top-p filtering.
    Simple implementation with proper KV caching support.

    Args:
        model: The language model
        model_kwargs: Model keyword arguments from the tokenizer
        generation_config: Generation configuration
        temperature: Sampling temperature (higher = more random)
        top_k: Only consider top-k most probable tokens
        top_p: Only consider tokens with cumulative probability <= top_p
        **kwargs: Additional arguments

    Returns:
        Generated token IDs
    """
    generation_config = model.generation_config
    max_new_tokens = kwargs.get("max_new_tokens", generation_config.max_new_tokens)
    max_new_tokens = 512 if max_new_tokens is None else max_new_tokens
    do_sample = kwargs.get("do_sample", True)
    eos_token_ids = kwargs.get("eos_token_ids", generation_config.eos_token_id)
    if eos_token_ids is None:
        raise ValueError(
            "Model generation config does not have an EOS token id. You must provide it to generate() with the eos_token_ids argument."
        )
    eos_token_ids = torch.as_tensor(eos_token_ids, device=model.device)
    if eos_token_ids is not None and eos_token_ids.ndim == 0:
        eos_token_ids = eos_token_ids.unsqueeze(0)

    pad_token_id = kwargs.get(
        "pad_token_id",
        generation_config.pad_token_id
        if generation_config.pad_token_id is not None
        else eos_token_ids[0],
    )
    bos_token_id = kwargs.get("bos_token_id", generation_config.bos_token_id)
    if bos_token_id is None:
        raise ValueError(
            "Model generation config does not have a BOS token id. You must provide it to generate() with the bos_token_id argument."
        )
    temperature = kwargs.get("temperature", 1.0)
    return_dict = kwargs.get("return_dict_in_generate", False)

    generated_ids, scores, logprobs = sampling(
        model_kwargs=kwargs,
        model=model,
        eos_token_ids=eos_token_ids,
        pad_token_id=pad_token_id,
        bos_token_id=bos_token_id,
        do_sample=do_sample,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
    )

    if return_dict:
        return {
            "sequences": generated_ids,
            "scores": scores,
            "logprobs": logprobs,
        }
    else:
        return generated_ids