vllm.v1.sample.ops.topk_topp_sampler

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 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_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 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)

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)