topk_topp_sampler.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from typing import Optional

import torch
import torch.nn as nn

from vllm import envs
from vllm.logger import init_logger
from vllm.platforms import current_platform

logger = init_logger(__name__)

try:
    import flashinfer.sampling
    is_flashinfer_available = True
except ImportError:
    is_flashinfer_available = False


class TopKTopPSampler(nn.Module):
    """
    Module that performs optional top-k and top-p filtering followed by
    weighted random sampling of logits.

    Implementations may update the logits tensor in-place.
    """

    def __init__(self):
        super().__init__()
        if current_platform.is_cuda():
            if is_flashinfer_available:
                flashinfer_version = flashinfer.__version__
                if flashinfer_version < "0.2.3":
                    logger.warning(
                        "FlashInfer version >= 0.2.3 required. "
                        "Falling back to default sampling implementation.")
                    self.forward = self.forward_native
                elif envs.VLLM_USE_FLASHINFER_SAMPLER is not False:
                    # NOTE(woosuk): The V0 sampler doesn't use FlashInfer for
                    # sampling unless VLLM_USE_FLASHINFER_SAMPLER=1 (i.e., by
                    # default it is unused). For backward compatibility, we set
                    # `VLLM_USE_FLASHINFER_SAMPLER` as None by default and
                    # interpret it differently in V0 and V1 samplers: In V0,
                    # None means False, while in V1, None means True. This is
                    # why we use the condition
                    # `envs.VLLM_USE_FLASHINFER_SAMPLER is not False` here.
                    logger.info("Using FlashInfer for top-p & top-k sampling.")
                    self.forward = self.forward_cuda
                else:
                    logger.warning(
                        "FlashInfer is available, but it is not enabled. "
                        "Falling back to the PyTorch-native implementation of "
                        "top-p & top-k sampling. For the best performance, "
                        "please set VLLM_USE_FLASHINFER_SAMPLER=1.")
                    self.forward = self.forward_native
            else:
                logger.warning(
                    "FlashInfer is not available. Falling back to the PyTorch-"
                    "native implementation of top-p & top-k sampling. For the "
                    "best performance, please install FlashInfer.")
                self.forward = self.forward_native
        elif current_platform.is_tpu():
            self.forward = self.forward_tpu
        else:
            self.forward = self.forward_native

    def forward_native(
        self,
        logits: torch.Tensor,
        generators: dict[int, torch.Generator],
        k: Optional[torch.Tensor],
        p: Optional[torch.Tensor],
        *,
        max_top_k: Optional[int] = None,
        has_any_no_top_k: bool = False,
    ) -> torch.Tensor:
        """
        PyTorch-native implementation of top-k and top-p sampling.

        The logits tensor may be updated in-place.
        """
        # Fast path: when top-k is enabled, avoid full-vocab sort/softmax by
        # sampling only from the top-k candidates (and applying top-p within
        # that set). This is especially important on ROCm where the PyTorch
        # native sort path can be very expensive.
        #
        # NOTE: Do not branch on device tensors here; doing so triggers
        # `aten::is_nonzero` and synchronizes the CPU with GPU.
        if (envs.VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER and k is not None
                and max_top_k is not None and not has_any_no_top_k
                and max_top_k <= 4096):
            try:
                return sample_top_k_top_p_reduced(logits,
                                                  generators,
                                                  k,
                                                  p,
                                                  max_top_k=max_top_k)
            except Exception:
                # Fall back to the reference implementation for safety.
                pass
        logits = apply_top_k_top_p(logits, k, p)
        probs = logits.softmax(dim=-1, dtype=torch.float32)
        return random_sample(probs, generators)

    def forward_cuda(
        self,
        logits: torch.Tensor,
        generators: dict[int, torch.Generator],
        k: Optional[torch.Tensor],
        p: Optional[torch.Tensor],
        *,
        max_top_k: Optional[int] = None,
        has_any_no_top_k: bool = False,
    ) -> torch.Tensor:
        """More optimized implementation for top-k and top-p sampling."""
        if k is None and p is None:
            # We prefer `random_sample` over `flashinfer_sample` when sorting is
            # not needed. This is because `random_sample` does not require
            # CPU-GPU synchronization while `flashinfer_sample` does.
            probs = logits.softmax(dim=-1, dtype=torch.float32)
            return random_sample(probs, generators)
        if generators:
            logger.warning("FlashInfer 0.2.3+ does not support "
                           "per-request generators. Falling back to "
                           "PyTorch-native implementation.")
            return self.forward_native(logits,
                                       generators,
                                       k,
                                       p,
                                       max_top_k=max_top_k,
                                       has_any_no_top_k=has_any_no_top_k)
        # flashinfer sampling functions expect contiguous logits.
        # In flex_attn/triton_attn fp32 inference, logits can be non-contiguous
        # because of slicing operation in logits_processor.
        return flashinfer_sample(logits.contiguous(), k, p, generators)

    def forward_tpu(
        self,
        logits: torch.Tensor,
        generators: dict[int, torch.Generator],
        k: Optional[torch.Tensor],
        p: Optional[torch.Tensor],
    ) -> torch.Tensor:
        logits = apply_top_k_top_p_tpu(logits, k, p)
        probs = logits.softmax(dim=-1, dtype=torch.float32)
        return random_sample(probs, generators)


def apply_top_k_top_p_tpu(
    logits: torch.Tensor,
    k: torch.Tensor,
    p: torch.Tensor,
) -> torch.Tensor:
    """
    Apply top-k and top-p optimized for TPU.

    This algorithm avoids using torch.scatter which is extremely slow on TPU.
    This is achieved by finding a "cut-off" element in the original logit, and
    after thresholding the logit using this cut-off, the remaining elements
    shall constitute the top-p set.

    Note: in the case of tie (i.e. multipple cut-off elements present in the
    logit), all tie elements are included in the top-p set. In other words,
    this function does not break ties. Instead, these tie tokens have equal
    chance of being chosen during final sampling, so we can consider the tie
    being broken then.
    """
    probs = logits.softmax(dim=-1)
    probs_sort, _ = probs.sort(dim=-1, descending=False)

    if k is not None:
        top_k_count = probs_sort.size(1) - k.to(torch.long)  # shape: (batch, )
        top_k_count = top_k_count.unsqueeze(dim=1)
        top_k_cutoff = probs_sort.gather(-1, top_k_count)

        # Make sure the no top-k rows are no-op.
        no_top_k_mask = (k == logits.shape[1]).unsqueeze(dim=1)
        top_k_cutoff.masked_fill_(no_top_k_mask, -float("inf"))

        elements_to_discard = probs < top_k_cutoff
        logits.masked_fill_(elements_to_discard, -float("inf"))

    if p is not None:
        cumprob = torch.cumsum(probs_sort, dim=-1)
        top_p_mask = cumprob <= 1 - p.unsqueeze(dim=1)
        top_p_mask[:, -1] = False  # at least one

        top_p_count = top_p_mask.sum(dim=-1).unsqueeze(1)
        top_p_cutoff = probs_sort.gather(-1, top_p_count)
        elements_to_discard = probs < top_p_cutoff
        logits.masked_fill_(elements_to_discard, -float("inf"))

    return logits


def apply_top_k_top_p(
    logits: torch.Tensor,
    k: Optional[torch.Tensor],
    p: Optional[torch.Tensor],
) -> torch.Tensor:
    """Apply top-k and top-p masks to the logits.

    If a top-p is used, this function will sort the logits tensor,
    which can be slow for large batches.

    The logits tensor may be updated in-place.
    """
    if p is None:
        if k is None:
            return logits

        # Avoid sorting vocab for top-k only case.
        return apply_top_k_only(logits, k)

    logits_sort, logits_idx = logits.sort(dim=-1, descending=False)

    if k is not None:
        # Apply top-k.
        top_k_mask = logits_sort.size(1) - k.to(torch.long)  # shape: B
        # Get all the top_k values.
        top_k_mask = logits_sort.gather(1, top_k_mask.unsqueeze(dim=1))
        top_k_mask = logits_sort < top_k_mask
        logits_sort.masked_fill_(top_k_mask, -float("inf"))

    if p is not None:
        # Apply top-p.
        probs_sort = logits_sort.softmax(dim=-1)
        probs_sum = torch.cumsum(probs_sort, dim=-1, out=probs_sort)
        top_p_mask = probs_sum <= 1 - p.unsqueeze(dim=1)
        # at least one
        top_p_mask[:, -1] = False
        logits_sort.masked_fill_(top_p_mask, -float("inf"))

    # Re-sort the probabilities.
    logits = logits_sort.scatter(dim=-1, index=logits_idx, src=logits_sort)
    return logits


def apply_top_k_only(
    logits: torch.Tensor,
    k: torch.Tensor,
) -> torch.Tensor:
    """
    Apply top-k mask to the logits.

    This implementation doesn't involve sorting the entire vocab.

    The logits tensor may be updated in-place.
    """
    no_top_k_mask = k == logits.shape[1]
    # Set non-top-k rows to 1 so that we can gather.
    k = k.masked_fill(no_top_k_mask, 1)
    max_top_k = k.max()
    # topk.values tensor has shape [batch_size, max_top_k].
    # Convert top k to 0-based index in range [0, max_top_k).
    k_index = k.sub_(1).unsqueeze(1)
    top_k_mask = logits.topk(max_top_k, dim=1).values.gather(1, k_index.long())
    # Handle non-topk rows.
    top_k_mask.masked_fill_(no_top_k_mask.unsqueeze(1), -float("inf"))
    logits.masked_fill_(logits < top_k_mask, -float("inf"))
    return logits


def random_sample(
    probs: torch.Tensor,
    generators: dict[int, torch.Generator],
) -> torch.Tensor:
    """Randomly sample from the probabilities.

    We use this function instead of torch.multinomial because torch.multinomial
    causes CPU-GPU synchronization.
    """
    q = torch.empty_like(probs)
    # NOTE(woosuk): To batch-process the requests without their own seeds,
    # which is the common case, we first assume that every request does
    # not have its own seed. Then, we overwrite the values for the requests
    # that have their own seeds.
    if len(generators) != probs.shape[0]:
        q.exponential_()
    if generators:
        # TODO(woosuk): This can be slow because we handle each request
        # one by one. Optimize this.
        for i, generator in generators.items():
            q[i].exponential_(generator=generator)
    return probs.div_(q).argmax(dim=-1).view(-1)


def sample_top_k_top_p_reduced(
    logits: torch.Tensor,
    generators: dict[int, torch.Generator],
    k: torch.Tensor,
    p: Optional[torch.Tensor],
    *,
    max_top_k: int,
) -> torch.Tensor:
    """Sample from logits using only the top-k candidates.

    This avoids full-vocab sorting and full-vocab softmax/exponential kernels.
    Semantics match applying top-k then top-p (if provided) and sampling from
    the resulting distribution.
    """
    vocab_size = logits.shape[-1]
    # Cap for safety; very large top-k values may be expensive or defeat the
    # purpose of the reduced path.
    if max_top_k <= 0 or max_top_k >= vocab_size:
        masked_logits = apply_top_k_top_p(logits, k, p)
        probs = masked_logits.softmax(dim=-1, dtype=torch.float32)
        return random_sample(probs, generators)

    topk = logits.topk(max_top_k, dim=-1)
    topk_logits = topk.values
    topk_indices = topk.indices

    # Apply per-row top-k (some rows may have smaller k).
    # topk_logits is sorted descending by default.
    k = k.to(torch.long)
    arange_k = torch.arange(max_top_k, device=logits.device).unsqueeze(0)
    keep_k = arange_k < k.unsqueeze(1)
    topk_logits = topk_logits.masked_fill(~keep_k, -float("inf"))

    # Convert to probabilities over the reduced candidate set.
    probs = topk_logits.softmax(dim=-1, dtype=torch.float32)

    if p is not None:
        # Apply top-p within the reduced set. Since candidates are already
        # sorted by descending logit, we can do cumulative top-p on this order.
        # Keep tokens until cumprob exceeds p, inclusive of the boundary token.
        cumprob = torch.cumsum(probs, dim=-1)
        cumprob_prev = cumprob - probs
        keep_p = cumprob_prev <= p.unsqueeze(1)
        probs = probs * keep_p

    # Sample a position within the reduced set and map it back to vocab ids.
    pos = random_sample(probs, generators)
    return topk_indices.gather(1, pos.unsqueeze(1)).squeeze(1)


def flashinfer_sample(
    logits: torch.Tensor,
    k: Optional[torch.Tensor],
    p: Optional[torch.Tensor],
    generators: dict[int, torch.Generator],
) -> torch.Tensor:
    """Sample from the logits using FlashInfer.

    Statistically, this function is equivalent to the `random_sample` function.
    However, this function is faster because it avoids sorting the logits tensor
    via rejection sampling.

    NOTE: The outputs of this function do not necessarily match the outputs of
    the `random_sample` function. It only guarantees that the outputs are
    statistically equivalent.

    NOTE: This function includes CPU-GPU synchronization, while `random_sample`
    does not. Call this function at the end of the forward pass to minimize
    the synchronization overhead.
    """
    assert not (k is None and p is None)
    if k is None:
        # Top-p only.
        probs = logits.softmax(dim=-1, dtype=torch.float32)
        next_token_ids = flashinfer.sampling.top_p_sampling_from_probs(
            probs, p, deterministic=True)
    elif p is None:
        # Top-k only.
        probs = logits.softmax(dim=-1, dtype=torch.float32)
        next_token_ids = flashinfer.sampling.top_k_sampling_from_probs(
            probs, k, deterministic=True)
    else:
        # Both top-k and top-p.
        next_token_ids = flashinfer.sampling.top_k_top_p_sampling_from_logits(
            logits, k, p, deterministic=True)

    return next_token_ids.view(-1)