fish-speech/tools/llama/generate.py

import os
import queue
import threading
import time
from contextlib import nullcontext
from dataclasses import dataclass
from pathlib import Path
from typing import Literal, Optional, Tuple, Union

import click
import hydra
import numpy as np
import torch
import torch._dynamo.config
import torch._inductor.config
from loguru import logger
from tqdm import tqdm

from fish_speech.conversation import CODEBOOK_PAD_TOKEN_ID
from fish_speech.text import clean_text, split_text

os.environ["TOKENIZERS_PARALLELISM"] = "false"
torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.triton.unique_kernel_names = True

if hasattr(torch._inductor.config, "fx_graph_cache"):
    # Experimental feature to reduce compilation times, will be on by default in future
    torch._inductor.config.fx_graph_cache = True


from fish_speech.models.text2semantic.llama import (
    BaseTransformer,
    DualARTransformer,
    NaiveTransformer,
)


def multinomial_sample_one_no_sync(
    probs_sort,
):  # Does multinomial sampling without a cuda synchronization
    q = torch.empty_like(probs_sort).exponential_(1)
    return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)


def logits_to_probs(
    logits,
    previous_tokens: Optional[torch.Tensor] = None,
    temperature: torch.Tensor = 1.0,
    top_p: torch.Tensor = 1.0,
    repetition_penalty: torch.Tensor = 1.0,
) -> torch.Tensor:
    # Apply repetition penalty
    if previous_tokens is not None:
        previous_tokens = previous_tokens.long()
        score = torch.gather(logits, dim=0, index=previous_tokens)
        score = torch.where(
            score < 0, score * repetition_penalty, score / repetition_penalty
        )
        logits.scatter_(dim=0, index=previous_tokens, src=score)

    # Apply top-p sampling
    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
    cum_probs = torch.cumsum(torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1)
    sorted_indices_to_remove = cum_probs > top_p
    sorted_indices_to_remove[0] = False  # keep at least one option
    indices_to_remove = sorted_indices_to_remove.scatter(
        dim=0, index=sorted_indices, src=sorted_indices_to_remove
    )
    logits = logits.masked_fill(indices_to_remove, -float("Inf"))

    logits = logits / max(temperature, 1e-5)

    probs = torch.nn.functional.softmax(logits, dim=-1)
    return probs


def sample(
    logits,
    previous_tokens: Optional[torch.Tensor] = None,
    **sampling_kwargs,
) -> Tuple[torch.Tensor, torch.Tensor]:
    probs = logits_to_probs(
        logits=logits[0, -1], previous_tokens=previous_tokens, **sampling_kwargs
    )
    idx_next = multinomial_sample_one_no_sync(probs)
    return idx_next, probs


def decode_one_token_ar(
    model: DualARTransformer,
    x: torch.Tensor,
    input_pos: torch.Tensor,
    previous_tokens: torch.Tensor = None,
    **sampling_kwargs,
) -> torch.Tensor:
    x = model.forward_generate(x, input_pos)

    sampling_kwargs_main = sampling_kwargs.copy()
    sampling_kwargs_main["temperature"] = 0.1
    sampling_kwargs_main["top_p"] = 0.1
    sampling_kwargs_main["repetition_penalty"] = 1.0

    codebooks = [
        sample(
            x.logits,
            previous_tokens=None,  # Disable repetition penalty for the token codebook
            **sampling_kwargs_main,
        )[0]
    ]

    x = x.hidden_states

    # Cleanup the cache
    for layer in model.fast_layers:
        layer.attention.kv_cache.k_cache.fill_(0)
        layer.attention.kv_cache.v_cache.fill_(0)

    for codebook_idx in range(model.config.num_codebooks):
        input_pos = torch.tensor([codebook_idx], device=x.device, dtype=torch.long)
        logits = model.forward_generate_fast(x, input_pos)
        a = sample(
            logits,
            previous_tokens=(
                previous_tokens[codebook_idx + 1]
                if previous_tokens is not None
                else None
            ),
            **sampling_kwargs,
        )[0]
        x = model.fast_embeddings(a)
        codebooks.append(a)

    return torch.stack(codebooks, dim=0)


def decode_one_token_naive(
    model: NaiveTransformer,
    x: torch.Tensor,
    input_pos: torch.Tensor,
    previous_tokens: torch.Tensor = None,
    **sampling_kwargs,
) -> torch.Tensor:
    x = model.forward_generate(x, input_pos)

    sampling_kwargs_main = sampling_kwargs.copy()
    sampling_kwargs_main["temperature"] = 0.1
    sampling_kwargs_main["top_p"] = 0.1
    sampling_kwargs_main["repetition_penalty"] = 1.0

    codebooks = [
        sample(
            x.logits,
            previous_tokens=None,  # Disable repetition penalty for the token codebook
            **sampling_kwargs_main,
        )[0]
    ]

    for i in range(model.config.num_codebooks):
        codebooks.append(
            sample(
                x.codebook_logits[:, :, i],
                previous_tokens=(
                    previous_tokens[i + 1] if previous_tokens is not None else None
                ),
                **sampling_kwargs,
            )[0]
        )

    return torch.stack(codebooks, dim=0)


def decode_n_tokens(
    model: NaiveTransformer,
    cur_token: torch.Tensor,
    input_pos: torch.Tensor,
    num_new_tokens: int,
    im_end_id: int = 4,
    decode_one_token=decode_one_token_naive,
    **sampling_kwargs,
):
    previous_tokens = torch.zeros(
        (model.config.num_codebooks + 1, model.config.max_seq_len),
        dtype=torch.int,
        device=cur_token.device,
    )

    for i in tqdm(range(num_new_tokens)):
        # We need to get windowed repeat penalty
        win_size = 16
        if i < win_size:
            window = previous_tokens[:, :win_size]
        else:
            window = previous_tokens[:, i - win_size : i]

        with (
            torch.backends.cuda.sdp_kernel(
                enable_flash=False, enable_mem_efficient=False, enable_math=True
            )
            if torch.cuda.is_available()
            else nullcontext()
        ):  # Actually better for Inductor to codegen attention here
            next_token = decode_one_token(
                model=model,
                x=cur_token,
                input_pos=input_pos,
                previous_tokens=window,
                **sampling_kwargs,
            )

        input_pos += 1
        cur_token = next_token.view(1, model.config.num_codebooks + 1, -1)
        previous_tokens[:, i : i + 1] = next_token.view(
            model.config.num_codebooks + 1, -1
        )

        if cur_token[0, 0, -1] == im_end_id:
            break

    return previous_tokens[:, : i + 1]


@torch.no_grad()
@torch.inference_mode()
def generate(
    *,
    model: NaiveTransformer,
    prompt: torch.Tensor,
    max_new_tokens: int,
    im_end_id: int = 4,
    decode_one_token=decode_one_token_naive,
    **sampling_kwargs,
) -> torch.Tensor:
    """
    Takes a conditioning sequence (prompt) as input and continues to generate as many tokens as requested.
    """

    # create an empty tensor of the expected final shape and fill in the current tokens
    T = prompt.size(1)

    device, dtype = prompt.device, prompt.dtype

    codebook_dim = 1 + model.config.num_codebooks
    # create an empty tensor of the expected final shape and fill in the current tokens
    empty = torch.empty((codebook_dim, max_new_tokens), dtype=dtype, device=device)
    empty[:, :T] = prompt
    seq = empty
    input_pos = torch.arange(0, T, device=device)

    # Use non-accelerated version for now, to avoid compilation overhead
    prefill_decode = (
        decode_one_token_naive
        if isinstance(model, NaiveTransformer)
        else decode_one_token_ar
    )

    next_token = prefill_decode(
        model, prompt.view(1, codebook_dim, -1), input_pos, **sampling_kwargs
    )
    seq[:, T : T + 1] = next_token

    input_pos = torch.tensor([T], device=device, dtype=torch.int)
    x = decode_n_tokens(
        model,
        next_token.view(1, codebook_dim, -1),
        input_pos,
        max_new_tokens - 1,
        im_end_id=im_end_id,
        decode_one_token=decode_one_token,
        **sampling_kwargs,
    )
    # x = torch.cat(generated_tokens, dim=1)
    seq = seq[:, : T + 1 + x.size(1)]
    seq[:, T + 1 :] = x

    return seq


def encode_tokens(
    tokenizer,
    string,
    device="cuda",
    prompt_tokens=None,
    num_codebooks=4,
):
    string = clean_text(string)
    string = f"<|im_start|>user\n{string}<|im_end|><|im_start|>assistant\n"

    new_tokens = tokenizer.encode(
        string,
        add_special_tokens=False,
        max_length=10**6,
        truncation=False,
    )
    tokens = torch.tensor([new_tokens], dtype=torch.int, device=device)

    # Codebooks
    zeros = (
        torch.ones((num_codebooks, tokens.size(1)), dtype=torch.int, device=device)
        * CODEBOOK_PAD_TOKEN_ID
    )
    prompt = torch.cat((tokens, zeros), dim=0)

    if prompt_tokens is None:
        return prompt

    # Get prompt tokens
    if prompt_tokens.ndim == 3:
        assert (
            prompt_tokens.shape[0] == 1
        ), f"3 dim prompt tokens should have shape (1, num_codebooks, seq_len)"
        prompt_tokens = prompt_tokens[0]

    assert prompt_tokens.ndim == 2
    data = prompt_tokens + 1

    if prompt_tokens.shape[0] > num_codebooks:
        logger.warning(
            f"Prompt tokens shape {prompt_tokens.shape} is larger than num_codebooks {num_codebooks}, getting first {num_codebooks} codebooks"
        )
        data = data[:num_codebooks]

    # Add pad token for each codebook
    data = torch.cat(
        (data, torch.zeros((data.size(0), 1), dtype=torch.int, device=device)),
        dim=1,
    )

    # Since 1.0, we use <|semantic|>
    s0_token_id = tokenizer.convert_tokens_to_ids("<|semantic|>")
    end_token_id = tokenizer.convert_tokens_to_ids("<|im_end|>")
    main_token_ids = (
        torch.ones((1, data.size(1)), dtype=torch.int, device=device) * s0_token_id
    )
    main_token_ids[0, -1] = end_token_id

    data = torch.cat((main_token_ids, data), dim=0)
    prompt = torch.cat((prompt, data), dim=1)

    return prompt


def load_model(checkpoint_path, device, precision, compile=False):
    model: Union[NaiveTransformer, DualARTransformer] = BaseTransformer.from_pretrained(
        checkpoint_path, load_weights=True
    )

    model = model.to(device=device, dtype=precision)
    logger.info(f"Restored model from checkpoint")

    if isinstance(model, DualARTransformer):
        decode_one_token = decode_one_token_ar
        logger.info("Using DualARTransformer")
    else:
        decode_one_token = decode_one_token_naive
        logger.info("Using NaiveTransformer")

    if compile:
        logger.info("Compiling function...")
        decode_one_token = torch.compile(
            decode_one_token,
            fullgraph=True,
            backend="inductor" if torch.cuda.is_available() else "aot_eager",
            mode="reduce-overhead" if torch.cuda.is_available() else None,
        )

    return model.eval(), decode_one_token


@dataclass
class GenerateResponse:
    action: Literal["sample", "next"]
    codes: Optional[torch.Tensor] = None
    text: Optional[str] = None


def generate_long(
    *,
    model,
    device: str | torch.device,
    decode_one_token: callable,
    text: str,
    num_samples: int = 1,
    max_new_tokens: int = 0,
    top_p: int = 0.7,
    repetition_penalty: float = 1.5,
    temperature: float = 0.7,
    compile: bool = False,
    iterative_prompt: bool = True,
    max_length: int = 2048,
    chunk_length: int = 150,
    prompt_text: Optional[str | list[str]] = None,
    prompt_tokens: Optional[torch.Tensor | list[torch.Tensor]] = None,
):
    assert 0 < top_p <= 1, "top_p must be in (0, 1]"
    assert 0 < repetition_penalty < 2, "repetition_penalty must be in (0, 2)"
    assert 0 < temperature < 2, "temperature must be in (0, 2)"

    use_prompt = prompt_text is not None and prompt_tokens is not None
    if use_prompt and isinstance(prompt_text, str):
        prompt_text = [prompt_text]
        prompt_tokens = [prompt_tokens]

    assert use_prompt is False or len(prompt_text) == len(
        prompt_tokens
    ), "Prompt text and tokens must have the same length"

    model_size = sum(p.numel() for p in model.parameters() if p.requires_grad)
    tokenizer = model.tokenizer
    im_end_id = tokenizer.convert_tokens_to_ids("<|im_end|>")

    encoded = []
    texts = split_text(text, chunk_length) if iterative_prompt else [text]
    encoded_prompts = []

    if use_prompt:
        for idx, (t, c) in enumerate(zip(prompt_text, prompt_tokens)):
            encoded_prompts.append(
                encode_tokens(
                    tokenizer,
                    string=t,
                    device=device,
                    prompt_tokens=c,
                    num_codebooks=model.config.num_codebooks,
                )
            )

    for idx, text in enumerate(texts):
        encoded.append(
            encode_tokens(
                tokenizer,
                string=text,
                device=device,
                num_codebooks=model.config.num_codebooks,
            )
        )
        logger.info(f"Encoded text: {text}")

    # Move temperature, top_p, repetition_penalty to device
    # This is important so that changing params doesn't trigger recompile
    temperature = torch.tensor(temperature, device=device, dtype=torch.float)
    top_p = torch.tensor(top_p, device=device, dtype=torch.float)
    repetition_penalty = torch.tensor(
        repetition_penalty, device=device, dtype=torch.float
    )

    for sample_idx in range(num_samples):
        if torch.cuda.is_available():
            torch.cuda.synchronize()

        global_encoded = []
        seg_idx = 0

        while seg_idx < len(encoded):
            logger.info(
                f"Generating sentence {seg_idx + 1}/{len(encoded)} of sample {sample_idx + 1}/{num_samples}"
            )

            seg = encoded[seg_idx]
            global_encoded.append(seg)

            lengths = reversed([seg.size(1) for seg in global_encoded])

            # Pick last 2000 tokens
            count = 0
            for i, length in enumerate(lengths):
                count += length
                if count + length > max_length - 1024 - sum(
                    t.shape[1] for t in encoded_prompts
                ):
                    break

            if i != 0 and i % 2 == 0:
                i -= 1

            # Rotate the list, always make sure first segment is included to avoid drift
            if i < len(global_encoded) - 2:
                partial_encoded = global_encoded[:2] + global_encoded[-i:]
            else:
                partial_encoded = global_encoded

            if use_prompt:
                partial_encoded = encoded_prompts + partial_encoded

            cat_encoded = torch.cat(partial_encoded, dim=1)
            prompt_length = cat_encoded.size(1)

            t0 = time.perf_counter()
            y = generate(
                model=model,
                prompt=cat_encoded,
                max_new_tokens=max_new_tokens,
                im_end_id=im_end_id,
                decode_one_token=decode_one_token,
                temperature=temperature,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
            )

            if sample_idx == 0 and seg_idx == 0 and compile:
                logger.info(f"Compilation time: {time.perf_counter() - t0:.2f} seconds")

            if torch.cuda.is_available():
                torch.cuda.synchronize()

            t = time.perf_counter() - t0

            tokens_generated = y.size(1) - prompt_length
            tokens_sec = tokens_generated / t
            logger.info(
                f"Generated {tokens_generated} tokens in {t:.02f} seconds, {tokens_sec:.02f} tokens/sec"
            )
            logger.info(
                f"Bandwidth achieved: {model_size * tokens_sec / 1e9:.02f} GB/s"
            )

            if torch.cuda.is_available():
                logger.info(
                    f"GPU Memory used: {torch.cuda.max_memory_reserved() / 1e9:.02f} GB"
                )

            # Put the generated tokens
            # since there is <im_end> and <eos> tokens, we remove last 2 tokens
            codes = y[1:, prompt_length:-1].clone()
            codes = codes - 1
            assert (codes >= 0).all(), f"Negative code found"

            decoded = y[:, prompt_length:-1].clone()
            # But for global encoding, we should keep the <im_end> token

            global_encoded.append(decoded)
            assert (codes >= 0).all(), f"Negative code found: {codes}"
            yield GenerateResponse(action="sample", codes=codes, text=texts[seg_idx])
            seg_idx += 1

        # This indicates the end of the current sample
        yield GenerateResponse(action="next")


@dataclass
class WrappedGenerateResponse:
    status: Literal["success", "error"]
    response: Optional[GenerateResponse | Exception] = None


@dataclass
class GenerateRequest:
    request: dict
    response_queue: queue.Queue


def launch_thread_safe_queue(
    checkpoint_path,
    device,
    precision,
    compile: bool = False,
):
    input_queue = queue.Queue()
    init_event = threading.Event()

    def worker():
        model, decode_one_token = load_model(
            checkpoint_path, device, precision, compile=compile
        )
        with torch.device(device):
            model.setup_caches(
                max_batch_size=1, max_seq_len=2048, dtype=next(model.parameters()).dtype
            )
        init_event.set()

        while True:
            item: GenerateRequest | None = input_queue.get()
            if item is None:
                break

            kwargs = item.request
            response_queue = item.response_queue

            try:
                for chunk in generate_long(
                    model=model, decode_one_token=decode_one_token, **kwargs
                ):
                    response_queue.put(
                        WrappedGenerateResponse(status="success", response=chunk)
                    )
            except Exception as e:
                response_queue.put(WrappedGenerateResponse(status="error", response=e))

    threading.Thread(target=worker, daemon=True).start()
    init_event.wait()

    return input_queue


@click.command()
@click.option(
    "--text",
    type=str,
    default="你说的对, 但是原神是一款由米哈游自主研发的开放世界手游.",
)
@click.option("--prompt-text", type=str, default=None, multiple=True)
@click.option(
    "--prompt-tokens",
    type=click.Path(path_type=Path, exists=True),
    default=None,
    multiple=True,
)
@click.option("--num-samples", type=int, default=1)
@click.option("--max-new-tokens", type=int, default=1024)
@click.option("--top-p", type=float, default=0.7)
@click.option("--repetition-penalty", type=float, default=1.2)
@click.option("--temperature", type=float, default=0.7)
@click.option(
    "--checkpoint-path",
    type=click.Path(path_type=Path, exists=True),
    default="checkpoints/fish-speech-1.4",
)
@click.option("--device", type=str, default="cuda")
@click.option("--compile/--no-compile", default=False)
@click.option("--seed", type=int, default=42)
@click.option("--half/--no-half", default=False)
@click.option("--iterative-prompt/--no-iterative-prompt", default=True)
@click.option("--chunk-length", type=int, default=100)
def main(
    text: str,
    prompt_text: Optional[list[str]],
    prompt_tokens: Optional[list[Path]],
    num_samples: int,
    max_new_tokens: int,
    top_p: int,
    repetition_penalty: float,
    temperature: float,
    checkpoint_path: Path,
    device: str,
    compile: bool,
    seed: int,
    half: bool,
    iterative_prompt: bool,
    chunk_length: int,
) -> None:

    precision = torch.half if half else torch.bfloat16

    if prompt_text is not None and len(prompt_text) != len(prompt_tokens):
        raise ValueError(
            f"Number of prompt text ({len(prompt_text)}) and prompt tokens ({len(prompt_tokens)}) should be the same"
        )

    logger.info("Loading model ...")
    t0 = time.time()
    model, decode_one_token = load_model(
        checkpoint_path, device, precision, compile=compile
    )
    with torch.device(device):
        model.setup_caches(
            max_batch_size=1, max_seq_len=2048, dtype=next(model.parameters()).dtype
        )
    if torch.cuda.is_available():
        torch.cuda.synchronize()

    logger.info(f"Time to load model: {time.time() - t0:.02f} seconds")

    if prompt_tokens is not None:
        prompt_tokens = [torch.from_numpy(np.load(p)).to(device) for p in prompt_tokens]

    torch.manual_seed(seed)

    if torch.cuda.is_available():
        torch.cuda.manual_seed(seed)

    generator = generate_long(
        model=model,
        device=device,
        decode_one_token=decode_one_token,
        text=text,
        num_samples=num_samples,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        temperature=temperature,
        compile=compile,
        iterative_prompt=iterative_prompt,
        chunk_length=chunk_length,
        prompt_text=prompt_text,
        prompt_tokens=prompt_tokens,
    )

    idx = 0
    codes = []

    for response in generator:
        if response.action == "sample":
            codes.append(response.codes)
            logger.info(f"Sampled text: {response.text}")
        elif response.action == "next":
            if codes:
                np.save(f"codes_{idx}.npy", torch.cat(codes, dim=1).cpu().numpy())
                logger.info(f"Saved codes to codes_{idx}.npy")
            logger.info(f"Next sample")
            codes = []
            idx += 1
        else:
            logger.error(f"Error: {response}")


if __name__ == "__main__":
    main()