# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project from typing import 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 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.utils import is_pin_memory_available from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata from vllm.v1.attention.backends.utils import CommonAttentionMetadata from vllm.v1.kv_cache_interface import KVCacheConfig 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, 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) self.token_arange_np = np.arange(self.max_num_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.is_multimodal_model = vllm_config.model_config \ .is_multimodal_model 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) self.inputs_embeds = torch.zeros( (self.max_num_tokens, self.hidden_size), dtype=self.dtype, device=device) 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, # [batch_size] next_token_ids: torch.Tensor, common_attn_metadata: CommonAttentionMetadata, sampling_metadata: SamplingMetadata, mm_embeds: Optional[list[torch.Tensor]] = None, ) -> torch.Tensor: num_tokens = target_token_ids.shape[0] batch_size = next_token_ids.shape[0] last_token_indices = common_attn_metadata.query_start_loc[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 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, fast_build=True, ) # 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 # copy inputs to buffer for cudagraph self.positions[:num_tokens] = target_positions self.hidden_states[:num_tokens] = target_hidden_states if self.is_multimodal_model: input_ids = self.input_ids[:num_tokens] inputs_embeds = self.model.get_input_embeddings( input_ids, multimodal_embeddings=mm_embeds or None, ) self.inputs_embeds[:num_tokens] = inputs_embeds inputs_embeds = self.inputs_embeds[:num_input_tokens] input_ids = None else: inputs_embeds = None input_ids = self.input_ids[:num_input_tokens] with set_forward_context(per_layer_attn_metadata, self.vllm_config, num_tokens=num_input_tokens): ret_hidden_states = self.model( input_ids=input_ids, positions=self.positions[:num_input_tokens], hidden_states=self.hidden_states[:num_input_tokens], inputs_embeds=inputs_embeds, ) 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. # Currently FlashAttention is the only backend that supports # multi-token eagle spec decode. This is because the code below # makes assumptions about attn_metadata attributes available. assert isinstance(attn_metadata, FlashAttentionMetadata) # 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 = attn_metadata.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.is_multimodal_model: inputs_embeds = self.model.get_input_embeddings(input_ids) self.inputs_embeds[:batch_size] = inputs_embeds inputs_embeds = self.inputs_embeds[:input_batch_size] input_ids = None else: inputs_embeds = None input_ids = self.input_ids[:input_batch_size] # Run the model. with set_forward_context(per_layer_attn_metadata, self.vllm_config, num_tokens=input_batch_size): last_hidden_states, hidden_states = self.model( input_ids=input_ids, positions=self.positions[:input_batch_size], hidden_states=self.hidden_states[:input_batch_size], inputs_embeds=inputs_embeds, ) 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 def prepare_inputs( self, common_attn_metadata: CommonAttentionMetadata, # [batch_size] num_rejected_tokens: torch.Tensor ) -> tuple[CommonAttentionMetadata, torch.Tensor]: """ This function is used to prepare the inputs for the spec decode. It updates to the common_attn_metadata to account for the rejected tokens (and newly sampled tokens). It also returns the token indices of the tokens that should be fed to the speculator. """ # E.g. # common_attn_metadata.query_start_loc{_cpu}: # [0, q1, q1 + q2, q1 + q2 + q3] # common_attn_metadata.seq_lens{_cpu}: [s1, s2, s3] # num_rejected_tokens: [n1, n2, n3] # This function computes the intermediate values: # num_tokens_per_req: [q1 - n1, q2 - n2, q3 - n3] # And returns: # common_attn_metadata.query_start_loc{_cpu}: # [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3] # common_attn_metadata.seq_lens{_cpu}: # [s1 - n1 + 1, s2 - n2 + 1, s3 - n3 + 1] # token_indices: [0, 1, ..., q1 - n1 - 1, # q1, q1 + 1, ..., q1 + q2 - n2 - 1, # q1 + q2, q1 + q2 + 1, ..., q1 + q2 + q3 - n3 - 1] device = common_attn_metadata.query_start_loc.device query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu new_seq_lens_cpu = common_attn_metadata.seq_lens_cpu \ - num_rejected_tokens # [0, q1, q1 + q2, q1 + q2 + q3] -> [q1, q2, q3] new_query_len_per_req = (query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]) # [q1, q2, q3] -> [q1 - n1, q2 - n2, q3 - n3] new_num_tokens_per_req = new_query_len_per_req - num_rejected_tokens new_num_tokens_per_req_np = new_num_tokens_per_req.numpy() # [q1 - n1, q2 - n2, q3 - n3] -> # [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3] new_query_start_loc_cpu = torch.zeros( query_start_loc_cpu.shape, dtype=torch.int32, pin_memory=is_pin_memory_available()) new_query_start_loc_np = new_query_start_loc_cpu.numpy() np.cumsum(new_num_tokens_per_req_np, out=new_query_start_loc_np[1:]) total_num_tokens = new_query_start_loc_np[-1] # Example assuming num_tokens_per_req_np = [2, 4, 3] # this implies that `new_query_start_locs` is: # [0, 2, 6, 9] -> # [0, 0, 2, 2, 2, 2, 6, 6, 6] # _r1_ ____r2____ ___r3__ new_query_start_locs_expanded = np.repeat(new_query_start_loc_np[:-1], new_num_tokens_per_req_np) # [0, 1, 2, 3, 4, 5, 6, 7, 8] -> # [0, 1, 0, 1, 2, 3, 0, 1, 2] # _r1_ ____r2____ ___r3__ token_offests = self.token_arange_np[:total_num_tokens] \ - new_query_start_locs_expanded # Expand starting positions to match token pattern # [0, q1, q1 + q2] -> # [0, 0, q1, q1, q1, q1, q1 + q2, q1 + q2, q1 + q2] # _r1_ _____r2_______ ___________r3____________ old_query_start_locs_expanded = np.repeat( query_start_loc_cpu[:-1].numpy(), new_num_tokens_per_req_np) # Final token indices are: # [0, 1, // req 1 # q1 + 0, q1 + 1, q1 + 2, q1 + 3, // req 2 # q1 + q2 + 0, q1 + q2 + 1, q1 + q2 + 2] // req 3 token_indices_np = token_offests + old_query_start_locs_expanded token_indices = torch.from_numpy(token_indices_np).to( device, non_blocking=True) spec_common_attn_metadata = CommonAttentionMetadata( query_start_loc=new_query_start_loc_cpu.to(device, non_blocking=True), seq_lens=new_seq_lens_cpu.to(device, non_blocking=True), query_start_loc_cpu=new_query_start_loc_cpu, seq_lens_cpu=new_seq_lens_cpu, num_computed_tokens_cpu=common_attn_metadata. num_computed_tokens_cpu, num_reqs=common_attn_metadata.num_reqs, num_actual_tokens=total_num_tokens, max_query_len=new_query_len_per_req.max().item(), block_table_tensor=common_attn_metadata.block_table_tensor, slot_mapping=common_attn_metadata.slot_mapping[token_indices], causal=True, ) return spec_common_attn_metadata, 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.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 @torch.inference_mode() def dummy_run( self, num_tokens: int, ) -> None: with set_forward_context(None, self.vllm_config, num_tokens=num_tokens): if self.is_multimodal_model: input_ids = None inputs_embeds = self.inputs_embeds[:num_tokens] else: input_ids = self.input_ids[:num_tokens] inputs_embeds = None self.model( input_ids=input_ids, positions=self.positions[:num_tokens], hidden_states=self.hidden_states[:num_tokens], inputs_embeds=inputs_embeds, ) 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