eagle.py

# SPDX-License-Identifier: Apache-2.0
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 set_forward_context
from vllm.logger import init_logger
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.triton_utils import tl, triton
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
from vllm.v1.sample.metadata import SamplingMetadata

logger = init_logger(__name__)

PADDING_SLOT_ID = -1


class EagleProposer:

    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
    ):
        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.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.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)
        # 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)

    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,
        sampling_metadata: SamplingMetadata,
    ) -> 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()

        # 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,
        )
        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
        # copy inputs to buffer for cudagraph
        self.positions[:num_tokens] = target_positions
        self.hidden_states[:num_tokens] = target_hidden_states

        with set_forward_context(attn_metadata,
                                 self.vllm_config,
                                 num_tokens=num_input_tokens):
            last_hidden_states, hidden_states = self.model(
                input_ids=self.input_ids[:num_input_tokens],
                positions=self.positions[:num_input_tokens],
                hidden_states=self.hidden_states[:num_input_tokens],
            )
        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)

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

        positions = target_positions[last_token_indices]
        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]
        for _ 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)

            # Increment the sequence lengths.
            attn_metadata.max_seq_len += 1
            attn_metadata.seq_lens += 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

            # Run the model.
            with set_forward_context(attn_metadata,
                                     self.vllm_config,
                                     num_tokens=input_batch_size):
                last_hidden_states, hidden_states = self.model(
                    input_ids=self.input_ids[:input_batch_size],
                    positions=self.positions[:input_batch_size],
                    hidden_states=self.hidden_states[:input_batch_size],
                )
            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,
    ) -> 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.empty_like(cu_target_query_lens)
        torch.cumsum(num_tokens_per_req, dim=0, out=cu_num_tokens[1:])
        cu_num_tokens[0] = 0

        # FIXME(woosuk): Avoid synchronization.
        num_tokens = cu_num_tokens[-1].item()
        token_indices = torch.empty(
            num_tokens,
            dtype=torch.int32,
            device=cu_num_tokens.device,
        )

        batch_size = num_rejected_tokens.shape[0]
        BLOCK_SIZE = 1024
        prepare_input_kernel[(batch_size, )](
            token_indices,
            cu_target_query_lens,
            cu_num_tokens,
            BLOCK_SIZE=BLOCK_SIZE,
        )
        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())

        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)
        assert len(draft_attn_layer_names) == 1
        self.attn_layer_name = next(iter(draft_attn_layer_names))

        # share embed_tokens with the target model if needed
        if get_pp_group().world_size == 1:
            logger.info(
                "The EAGLE head shares the same vocab embedding" \
                " with the target model."
            )
            self.model.model.embed_tokens = target_model.model.embed_tokens
        else:
            logger.info(
                "Since PP > 1, the EAGLE head loaded its own vocab embedding" \
                " weights instead of sharing them with 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_model, "lm_head"):
            logger.info("Loading EAGLE LM head weights from the target model.")
            if supports_multimodal(target_model):
                self.model.lm_head = target_model.get_language_model().lm_head
            else:
                self.model.lm_head = target_model.lm_head

    @torch.inference_mode()
    def dummy_run(
        self,
        num_tokens: int,
    ) -> None:
        with set_forward_context(None, self.vllm_config,
                                 num_tokens=num_tokens):
            self.model(
                input_ids=self.input_ids[:num_tokens],
                positions=self.positions[:num_tokens],
                hidden_states=self.hidden_states[:num_tokens],
            )


# 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


@triton.jit
def prepare_input_kernel(
    out_ptr,
    cu_query_lens_ptr,
    cu_num_tokens_ptr,
    BLOCK_SIZE: tl.constexpr,
):
    pid = tl.program_id(0)

    # [start_pos, end_pos)
    start_pos = tl.load(cu_num_tokens_ptr + pid)
    end_pos = tl.load(cu_num_tokens_ptr + pid + 1)
    num_tokens = end_pos - start_pos

    index_start = tl.load(cu_query_lens_ptr + pid)

    num_blocks = tl.cdiv(num_tokens, BLOCK_SIZE)
    for i in tl.range(num_blocks):
        offset = i * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
        tl.store(
            out_ptr + start_pos + offset,
            index_start + offset,
            mask=offset < num_tokens,
        )