eagle.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any, Optional
import numpy as np

import torch
import torch.nn as nn

from vllm.attention.layer import Attention
from vllm.config import (CompilationLevel, VllmConfig,
                         get_layers_from_vllm_config)
from vllm.distributed.parallel_state import get_pp_group
from vllm.forward_context import DPMetadata, set_forward_context
from vllm.logger import init_logger
import vllm.envs as envs
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models import supports_multimodal
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM

from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
from vllm.v1.attention.backends.mla.common import MLACommonMetadata, MLACommonDecodeMetadata

from vllm.v1.kv_cache_interface import KVCacheConfig
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.spec_decode.utils import prepare_eagle_input_kernel

logger = init_logger(__name__)

PADDING_SLOT_ID = -1


class EagleProposer:

    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
        runner=None,
    ):
        self.vllm_config = vllm_config
        self.speculative_config = vllm_config.speculative_config
        self.draft_model_config = self.speculative_config.draft_model_config
        self.method = self.speculative_config.method

        self.runner = runner

        self.dtype = vllm_config.model_config.dtype
        self.max_model_len = vllm_config.model_config.max_model_len
        self.block_size = vllm_config.cache_config.block_size
        self.num_speculative_tokens = (
            self.speculative_config.num_speculative_tokens)
        self.max_num_tokens = (
            vllm_config.scheduler_config.max_num_batched_tokens)
        # We need to get the hidden size from the draft model config because
        # the draft model's hidden size can be different from the target model's
        # hidden size (e.g., Llama 3.3 70B).
        self.hidden_size = self.draft_model_config.get_hidden_size()

        self.use_cuda_graph = (self.vllm_config.compilation_config.level
                               == CompilationLevel.PIECEWISE and
                               not self.vllm_config.model_config.enforce_eager)
        self.use_full_cuda_graph = (
            self.use_cuda_graph
            and vllm_config.compilation_config.full_cuda_graph)
        self.cudagraph_batch_sizes = list(
            reversed(
                self.vllm_config.compilation_config.cudagraph_capture_sizes))

        # persistent buffers for cuda graph
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=device)
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=device)
        self.hidden_states = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=device)

        # attention metadata captured in full cudagraph mode
        self.attn_metadata_cudagraph = None
        # We need +1 here because the arange is used to set query_start_loc,
        # which has one more element than batch_size.
        self.arange = torch.arange(vllm_config.scheduler_config.max_num_seqs +
                                   1,
                                   device=device,
                                   dtype=torch.int32)
        
        self.dp_size = self.vllm_config.parallel_config.data_parallel_size
        self.enable_expert_parallel = vllm_config.parallel_config.enable_expert_parallel

    def propose(
        self,
        # [num_tokens]
        target_token_ids: torch.Tensor,
        # [num_tokens]
        target_positions: torch.Tensor,
        # [num_tokens, hidden_size]
        target_hidden_states: torch.Tensor,
        # [num_tokens]
        target_slot_mapping: torch.Tensor,
        # [batch_size]
        next_token_ids: torch.Tensor,
        # [batch_size + 1] starting with 0
        cu_num_tokens: torch.Tensor,
        # [batch_size, max_num_blocks_per_req]
        block_table: torch.Tensor,
        # [batch_size]
        sampling_metadata: SamplingMetadata,
        decoding: bool = False,
    ) -> torch.Tensor:
        num_tokens = target_token_ids.shape[0]
        batch_size = next_token_ids.shape[0]
        last_token_indices = cu_num_tokens[1:] - 1

        if self.method == "eagle3":
            assert isinstance(self.model, Eagle3LlamaForCausalLM)
            target_hidden_states = self.model.combine_hidden_states(
                target_hidden_states)
            assert target_hidden_states.shape[-1] == self.hidden_size

        # Shift the input ids by one token.
        # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
        self.input_ids[:num_tokens - 1] = target_token_ids[1:]
        # Replace the last token with the next token.
        # E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
        self.input_ids[last_token_indices] = next_token_ids

        # FA requires seq_len to have dtype int32.
        seq_lens = (target_positions[last_token_indices] + 1).int()

        if self.method in ["eagle", "eagle3"]:
            # FIXME(woosuk): The below two ops cause synchronization. Optimize.
            max_seq_len = seq_lens.max().item()
            max_num_tokens = (cu_num_tokens[1:] -
                              cu_num_tokens[:-1]).max().item()
            attn_metadata = FlashAttentionMetadata(
                num_actual_tokens=num_tokens,
                max_query_len=max_num_tokens,
                query_start_loc=cu_num_tokens,
                max_seq_len=max_seq_len,
                seq_lens=seq_lens,
                block_table=block_table,
                slot_mapping=target_slot_mapping,
                # TODO(woosuk): Support cascade attention.
                use_cascade=False,
                common_prefix_len=0,
                cu_prefix_query_lens=None,
                prefix_kv_lens=None,
                suffix_kv_lens=None,
            )
        elif self.method == "deepseek_mtp":
            query_lens = cu_num_tokens[1:] - cu_num_tokens[:-1]
            max_query_len = query_lens.max().item()

            common_attn_metadata = CommonAttentionMetadata(
                query_start_loc=cu_num_tokens,
                seq_lens=seq_lens,
                num_reqs=batch_size,
                num_actual_tokens=num_tokens,
                max_query_len=max_query_len,
                slot_mapping=target_slot_mapping,
                spec_layer_decoding=decoding
            )

            assert self.runner is not None

            # FIXME: need to consider multiple kv_cache_groups
            attn_metadata = self.runner.attn_metadata_builders[0].build(
                common_prefix_len=0,
                common_attn_metadata=common_attn_metadata
            )
        else:
            raise ValueError(f"Unsupported method: {self.method}")

        # At this moment, we assume all eagle layers belong to the same KV
        # cache group, thus using the same attention metadata.
        per_layer_attn_metadata = {}
        for layer_name in self.attn_layer_names:
            per_layer_attn_metadata[layer_name] = attn_metadata
        if self.use_cuda_graph and \
            num_tokens <= self.cudagraph_batch_sizes[-1]:
            num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
        else:
            num_input_tokens = num_tokens
        

        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad
        # copy inputs to buffer for cudagraph
        self.positions[:num_tokens] = target_positions
        self.hidden_states[:num_tokens] = target_hidden_states

        if (decoding and self.use_full_cuda_graph
                and num_tokens <= self.cudagraph_batch_sizes[-1]):
            assert self.attn_metadata_cudagraph
            if self.method in ["eagle", "eagle3"]:
                self.attn_metadata_cudagraph.seq_lens[:batch_size] = (
                    attn_metadata.seq_lens)
                self.attn_metadata_cudagraph.slot_mapping[:num_tokens] = (
                    attn_metadata.slot_mapping)
                self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
                    attn_metadata.query_start_loc)
                self.attn_metadata_cudagraph.block_table[:batch_size] = (
                    attn_metadata.block_table)
            elif self.method == "deepseek_mtp":
                self.attn_metadata_cudagraph.num_actual_tokens = (
                    attn_metadata.num_actual_tokens)
                self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
                    attn_metadata.query_start_loc)
                self.attn_metadata_cudagraph.slot_mapping[:num_tokens] = (
                    attn_metadata.slot_mapping)
                self.attn_metadata_cudagraph.num_decodes = (
                    attn_metadata.num_decodes)
                self.attn_metadata_cudagraph.num_decode_tokens = (
                    attn_metadata.num_decode_tokens)
                self.attn_metadata_cudagraph.num_prefills = (
                    attn_metadata.num_prefills)

                if attn_metadata.decode is not None:
                    self.attn_metadata_cudagraph.decode.block_table[:attn_metadata.num_decode_tokens] = (
                            attn_metadata.decode.block_table)
                    self.attn_metadata_cudagraph.decode.seq_lens[:attn_metadata.num_decode_tokens] = (
                        attn_metadata.decode.seq_lens)

        with set_forward_context(per_layer_attn_metadata,
                                 self.vllm_config,
                                 num_tokens=num_input_tokens,
                                 num_tokens_across_dp=num_tokens_across_dp):
                                 #skip_cuda_graphs=not decoding):
            ret_hidden_states = self.model(
                self.input_ids[:num_input_tokens],
                self.positions[:num_input_tokens],
                self.hidden_states[:num_input_tokens],
            )
            if self.method == "deepseek_mtp":
                last_hidden_states = ret_hidden_states
            else:
                last_hidden_states, hidden_states = ret_hidden_states
        sample_hidden_states = last_hidden_states[last_token_indices]
        logits = self.model.compute_logits(sample_hidden_states, None)

        draft_token_ids = logits.argmax(dim=-1)

        # Early exit if there is only one draft token to be generated.
        if self.num_speculative_tokens == 1:
            # [batch_size, 1]
            return draft_token_ids.view(-1, 1)

        # TODO: Currently, MTP module released by deepseek only has
        # one layer. Adapt this code to support multiple layers once
        # there's a multi-layer MTP module.

        # Generate the remaining draft tokens.
        draft_token_ids_list = [draft_token_ids]

        positions = target_positions[last_token_indices]

        if self.method == "deepseek_mtp":
            hidden_states = last_hidden_states[last_token_indices]
        else:
            hidden_states = hidden_states[last_token_indices]

        if self.use_cuda_graph and \
                batch_size <= self.cudagraph_batch_sizes[-1]:
            input_batch_size = self.vllm_config.pad_for_cudagraph(batch_size)
        else:
            input_batch_size = batch_size
        attn_metadata.num_actual_tokens = batch_size
        attn_metadata.max_query_len = 1
        attn_metadata.query_start_loc = self.arange[:batch_size + 1]

        if isinstance(attn_metadata, MLACommonMetadata):
            attn_metadata.num_decodes = batch_size
            attn_metadata.num_decode_tokens = batch_size
            attn_metadata.num_prefills = 0
            block_table = self.runner.attn_metadata_builders[0].block_table.get_device_tensor()[:batch_size, ...]
            attn_metadata.decode = self.runner.attn_metadata_builders[0]._build_decode(
                block_table_tensor=block_table,
                seq_lens=seq_lens,
            )

        for i in range(self.num_speculative_tokens - 1):
            # Update the inputs.
            # cast to int32 is crucial when eagle model is compiled.
            # tensor.argmax() returns int64 by default.
            input_ids = draft_token_ids_list[-1].int()
            positions += 1

            # NOTE(woosuk): We should handle the case where the draft model
            # generates tokens beyond the max model length. Since it is complex
            # to remove such requests from the batch, we keep them in the batch
            # but adjust the position ids and slot mappings to avoid the
            # out-of-range access during the model execution. The draft tokens
            # generated with this adjustment should be ignored.
            exceeds_max_model_len = positions >= self.max_model_len
            # Mask out the position ids that exceed the max model length.
            # Otherwise, we may get out-of-range error in RoPE.
            clamped_positions = torch.where(exceeds_max_model_len, 0,
                                            positions)

            if isinstance(attn_metadata, MLACommonMetadata):
                attn_metadata.decode.seq_lens += 1
            else:
                attn_metadata.seq_lens += 1

                # Increment the sequence lengths.
                attn_metadata.max_seq_len += 1
                # Consider max model length.
                attn_metadata.max_seq_len = min(attn_metadata.max_seq_len,
                                                self.max_model_len)

                # For the requests that exceed the max model length, we set the
                # sequence length to 1 to minimize their overheads in attention.
                attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)

            # Compute the slot mapping.
            block_numbers = clamped_positions // self.block_size
            block_ids = block_table.gather(dim=1,
                                        index=block_numbers.view(-1, 1))
            block_ids = block_ids.view(-1)
            attn_metadata.slot_mapping = (block_ids * self.block_size +
                                        clamped_positions % self.block_size)
            # Mask out the slot mappings that exceed the max model length.
            # Otherwise, the KV cache will be inadvertently updated with the
            # padding tokens.
            attn_metadata.slot_mapping.masked_fill_(exceeds_max_model_len,
                                                    PADDING_SLOT_ID)

            # copy inputs to buffer for cudagraph
            self.input_ids[:batch_size] = input_ids
            self.positions[:batch_size] = clamped_positions
            self.hidden_states[:batch_size] = hidden_states

            if (self.use_full_cuda_graph
                    and batch_size <= self.cudagraph_batch_sizes[-1]):
                assert self.attn_metadata_cudagraph
                if self.method in ["eagle", "eagle3"]:
                    self.attn_metadata_cudagraph.seq_lens[:batch_size] = (
                        attn_metadata.seq_lens)
                    self.attn_metadata_cudagraph.slot_mapping[:batch_size] = (
                        attn_metadata.slot_mapping)
                    if i == 0:
                        self.attn_metadata_cudagraph.query_start_loc[:batch_size +
                                                                    1] = (
                                                                        attn_metadata
                                                                        .
                                                                        query_start_loc
                                                                    )
                        self.attn_metadata_cudagraph.block_table[:batch_size] = (
                            attn_metadata.block_table)
                elif self.method == "deepseek_mtp":
                    self.attn_metadata_cudagraph.num_actual_tokens = (
                        attn_metadata.num_actual_tokens)
                    self.attn_metadata_cudagraph.slot_mapping[:attn_metadata.num_decode_tokens] = (
                        attn_metadata.slot_mapping)
                    self.attn_metadata_cudagraph.num_decodes = (
                        attn_metadata.num_decodes)
                    self.attn_metadata_cudagraph.num_decode_tokens = (
                        attn_metadata.num_decode_tokens)
                    self.attn_metadata_cudagraph.num_prefills = (
                        attn_metadata.num_prefills)
                    self.attn_metadata_cudagraph.decode.seq_lens[:attn_metadata.num_decode_tokens] = (
                        attn_metadata.decode.seq_lens)

                    if i == 0:
                        self.attn_metadata_cudagraph.query_start_loc[:batch_size + 1] = (
                            attn_metadata.query_start_loc)
                        self.attn_metadata_cudagraph.decode.block_table[:attn_metadata.num_decode_tokens] = (
                            attn_metadata.decode.block_table)

            # Run the model.
            with set_forward_context(per_layer_attn_metadata,
                                     self.vllm_config,
                                     num_tokens=input_batch_size,
                                     num_tokens_across_dp=num_tokens_across_dp):
                ret_hidden_states = self.model(
                    self.input_ids[:input_batch_size],
                    self.positions[:input_batch_size],
                    self.hidden_states[:input_batch_size],
                )
                if self.method == "deepseek_mtp":
                    last_hidden_states = ret_hidden_states
                    hidden_states = last_hidden_states[:batch_size]
                else:
                    last_hidden_states, hidden_states = ret_hidden_states
                    hidden_states = hidden_states[:batch_size]

            logits = self.model.compute_logits(last_hidden_states[:batch_size],
                                               None)

            # TODO(wenlong): get more than one token for tree attention
            draft_token_ids = logits.argmax(dim=-1)
            draft_token_ids_list.append(draft_token_ids)

        # [batch_size, num_speculative_tokens]
        draft_token_ids = torch.stack(draft_token_ids_list, dim=1)

        return draft_token_ids

    # @staticmethod
    # def prepare_inputs(
    #     # [batch_size + 1]
    #     cu_target_query_lens: torch.Tensor,
    #     # [batch_size]
    #     num_rejected_tokens: torch.Tensor,
    #     num_tokens: int,
    # ) -> tuple[torch.Tensor, torch.Tensor]:
    #     # cu_target_query_lens: [0, a, a + b, a + b + c]
    #     # num_rejected_tokens: [n1, n2, n3]
    #     # num_tokens_per_req: [a - n1, b - n2, c - n3]
    #     # cu_num_tokens: [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
    #     # token_indices: [0, 1, ..., a - n1 - 1,
    #     #                 a, a + 1, ..., a + b - n2 - 1,
    #     #                 a + b, a + b + 1, ..., a + b + c - n3 - 1]

    #     # [0, a, a + b, a + b + c] -> [a, b, c]
    #     query_len_per_req = (cu_target_query_lens[1:] -
    #                          cu_target_query_lens[:-1])
    #     # [a, b, c] -> [a - n1, b - n2, c - n3]
    #     num_tokens_per_req = query_len_per_req - num_rejected_tokens

    #     # [a - n1, b - n2, c - n3] ->
    #     # [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
    #     cu_num_tokens = torch.zeros_like(cu_target_query_lens)
    #     torch.cumsum(num_tokens_per_req, dim=0, out=cu_num_tokens[1:])
    #     token_indices = torch.empty(
    #         num_tokens,
    #         dtype=torch.int32,
    #         device=cu_target_query_lens.device,
    #     )
    #     batch_size = num_rejected_tokens.shape[0]
    #     BLOCK_SIZE = 1024
    #     prepare_eagle_input_kernel[(batch_size, )](
    #         token_indices,
    #         cu_target_query_lens,
    #         cu_num_tokens,
    #         BLOCK_SIZE=BLOCK_SIZE,
    #     )
    #     return cu_num_tokens, token_indices

    @staticmethod
    def prepare_inputs(
        # [batch_size + 1]
        cu_target_query_lens: torch.Tensor,
        # [batch_size]
        num_accepted_tokens_tensor: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        cu_num_tokens = torch.arange(cu_target_query_lens.shape[0], device=cu_target_query_lens.device, dtype=torch.int32)
        token_indices = num_accepted_tokens_tensor + cu_target_query_lens[:-1]
        return cu_num_tokens, token_indices

    def load_model(self, target_model: nn.Module) -> None:
        draft_model_config = \
            self.vllm_config.speculative_config.draft_model_config
        target_attn_layer_names = set(
            get_layers_from_vllm_config(self.vllm_config, Attention).keys())

        from vllm.compilation.backends import set_model_tag
        with set_model_tag("eagle_head"):
            self.model = get_model(vllm_config=self.vllm_config,
                                   model_config=draft_model_config)

        draft_attn_layer_names = (
            get_layers_from_vllm_config(self.vllm_config, Attention).keys() -
            target_attn_layer_names)

        self.attn_layer_names = list(draft_attn_layer_names)

        if supports_multimodal(target_model):
            # handle multimodality
            self.model.config.image_token_index = (
                target_model.config.image_token_index)
            target_language_model = target_model.get_language_model()
        else:
            target_language_model = target_model
        # share embed_tokens with the target model if needed
        if get_pp_group().world_size == 1 \
            and self.method != "deepseek_mtp" \
            and self.model.model.embed_tokens.weight.shape \
                == target_language_model.model.embed_tokens.weight.shape:
            logger.info(
                "Assuming the EAGLE head shares the same vocab embedding" \
                " with the target model."
            )
            del self.model.model.embed_tokens
            self.model.model.embed_tokens = (
                target_language_model.model.embed_tokens)
        else:
            logger.info(
                "The EAGLE head's vocab embedding will be loaded separately" \
                " from the target model."
            )

        # share lm_head with the target model if needed
        # some model definition do not define lm_head explicitly
        # and reuse embed_tokens for lm_head, e.g., CohereForCausalLM
        if self.vllm_config.speculative_config.method != "eagle3" and \
                hasattr(target_language_model, "lm_head"):
            logger.info("Loading EAGLE LM head weights from the target model.")
            self.model.lm_head = target_language_model.lm_head

    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank

        # For DP: Don't pad when setting enforce_eager.
        # This lets us set enforce_eager on the prefiller in a P/D setup and
        # still use CUDA graphs (enabled by this padding) on the decoder.
        #
        # TODO(tms) : There are many cases where padding is enabled for
        # prefills, causing unnecessary and excessive padding of activations.

        if dp_size == 1 or self.vllm_config.model_config.enforce_eager or envs.VLLM_ALL2ALL_BACKEND != 'naive':
            # auto 
            if not envs.VLLM_ALL2ALL_BACKEND == "deepep_auto":
                # Early exit.
                return 0, None

        try:
            num_tokens_across_dp = DPMetadata.num_tokens_across_dp(
                num_tokens, dp_size, dp_rank)
            max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp).item()
            num_tokens_after_padding = torch.tensor([max_tokens_across_dp_cpu] *
                                                    dp_size,
                                                    device="cpu",
                                                    dtype=torch.int32)
            return max_tokens_across_dp_cpu - num_tokens, num_tokens_after_padding
        except (RuntimeError, AttributeError) as e:
            # DP group may not be initialized yet during dummy run
            # Skip padding in this case
            logger.debug(
                "Skipping DP padding in eagle get_dp_padding due to: %s", e)
            return 0, None

    @torch.inference_mode()
    def dummy_run(
        self,
        num_tokens: int,
        attn_metadata: Optional[dict[str, Any]] = None,
    ) -> None:
        if attn_metadata is not None and self.attn_metadata_cudagraph is None:
            self.attn_metadata_cudagraph = attn_metadata[
                self.attn_layer_names[0]]
        
        # Padding for DP
        num_input_tokens = num_tokens
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_input_tokens += num_pad
        
        with set_forward_context(attn_metadata,
                                 self.vllm_config,
                                 num_tokens=num_input_tokens,
                                 num_tokens_across_dp=num_tokens_across_dp):
            self.model(
                self.input_ids[:num_input_tokens],
                self.positions[:num_input_tokens],
                self.hidden_states[:num_input_tokens],
            )

        if self.dp_size > 1 and self.enable_expert_parallel and self.num_speculative_tokens > 1:
            num_tokens = 1
            for _ in range(self.num_speculative_tokens - 1):
                with set_forward_context(attn_metadata,
                                 self.vllm_config,
                                 num_tokens=num_tokens,):
                    self.model(
                        self.input_ids[:num_tokens],
                        self.positions[:num_tokens],
                        self.hidden_states[:num_tokens],
                    )

    def validate_same_kv_cache_group(self,
                                     kv_cache_config: KVCacheConfig) -> None:
        """
        Validate that all eagle layers belong to the same KVCacheGroup.
        Need this assumption to ensure all eagle layers can use the
        same AttentionMetadata.
        May extend to multiple AttentionMetadata in the future.
        """
        kv_cache_groups: dict[str, int] = {}
        for id, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            for layer_name in kv_cache_group.layer_names:
                kv_cache_groups[layer_name] = id
        assert len(
            set([
                kv_cache_groups[layer_name]
                for layer_name in self.attn_layer_names
            ])
        ) == 1, "All eagle layers should belong to the same kv cache group"


# NOTE(woosuk): Currently, the below code is not used and we always use argmax
# to sample the draft tokens. We will use this after we find a way to manage
# the draft prob tensor.
# Refer to https://github.com/vllm-project/vllm/pull/16899 for the details.
# FIXME(woosuk): The logic here is duplicated with the main sampling code.
# We should refactor this to reuse the same sampling implementation.
def compute_probs_and_sample_next_token(
    logits: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> tuple[torch.Tensor, torch.Tensor]:
    if sampling_metadata.all_greedy:
        # For greedy requests, draft_probs is not used in rejection sampling.
        # Therefore, we can just return the logits.
        probs = logits
        next_token_ids = logits.argmax(dim=-1)
        return next_token_ids, probs

    is_greedy = sampling_metadata.temperature == -1
    temperature = torch.where(is_greedy, 1.0, sampling_metadata.temperature)
    logits.div_(temperature.view(-1, 1))
    probs = logits.softmax(dim=-1, dtype=torch.float32)

    # NOTE(woosuk): Currently, we ignore most of the sampling parameters in
    # generating the draft tokens. We only use the temperature. While this
    # could degrade the acceptance rate, it does not affect the distribution
    # of the generated tokens after rejection sampling.

    # TODO(woosuk): Consider seeds.
    q = torch.empty_like(probs)
    q.exponential_()
    # NOTE(woosuk): We shouldn't use `probs.div_(q)` because the draft_probs
    # will be used later for rejection sampling.
    next_token_ids = probs.div(q).argmax(dim=-1).view(-1)
    if not sampling_metadata.all_random:
        greedy_token_ids = probs.argmax(dim=-1)
        next_token_ids = torch.where(
            is_greedy,
            greedy_token_ids,
            next_token_ids,
        )
    return next_token_ids, probs