# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import ast from dataclasses import replace from importlib.util import find_spec from typing import cast import numpy as np import torch import torch.nn as nn from vllm.config import ( CUDAGraphMode, 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.layers.attention_layer_base import AttentionLayerBase from vllm.model_executor.model_loader import get_model from vllm.model_executor.models import supports_multimodal from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache from vllm.model_executor.models.interfaces import SupportsMultiModal from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.platforms import current_platform from vllm.triton_utils import triton from vllm.utils.platform_utils import is_pin_memory_available from vllm.v1.attention.backend import ( AttentionMetadataBuilder, CommonAttentionMetadata, ) from vllm.v1.attention.backends.registry import AttentionBackendEnum from vllm.v1.attention.backends.tree_attn import ( TreeAttentionMetadata, TreeAttentionMetadataBuilder, ) from vllm.v1.attention.backends.triton_attn import TritonAttentionMetadata from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher from vllm.v1.kv_cache_interface import KVCacheConfig from vllm.v1.sample.metadata import SamplingMetadata from vllm.v1.sample.sampler import _SAMPLING_EPS from vllm.v1.spec_decode.metadata import SpecDecodeMetadata from vllm.v1.spec_decode.utils import ( PADDING_SLOT_ID, compute_new_slot_mapping, copy_and_expand_eagle_inputs_kernel, eagle_prepare_inputs_padded_kernel, eagle_prepare_next_token_padded_kernel, extend_all_queries_by_N, ) from vllm.v1.utils import CpuGpuBuffer from vllm.v1.worker.dp_utils import coordinate_batch_across_dp from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch logger = init_logger(__name__) class SpecDecodeBaseProposer: def __init__( self, vllm_config: VllmConfig, device: torch.device, pass_hidden_states_to_model: bool, runner=None, ): self.vllm_config = vllm_config assert vllm_config.speculative_config is not None self.speculative_config = vllm_config.speculative_config self.draft_model_config = self.speculative_config.draft_model_config self.method = self.speculative_config.method self.pass_hidden_states_to_model = pass_hidden_states_to_model self.runner = runner self.device = device self.dtype = vllm_config.model_config.dtype self.max_model_len = vllm_config.model_config.max_model_len self.dp_rank = vllm_config.parallel_config.data_parallel_rank self.num_speculative_tokens = self.speculative_config.num_speculative_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.inputs_embeds_size = self.draft_model_config.get_inputs_embeds_size() # Unifying eagle, draft model, and parallel drafting support self.parallel_drafting: bool = self.speculative_config.parallel_drafting self.extra_slots_per_request = ( 1 if not self.parallel_drafting else self.num_speculative_tokens ) self.net_num_new_slots_per_request = self.extra_slots_per_request - ( 1 if self.pass_hidden_states_to_model else 0 ) self.needs_extra_input_slots = self.net_num_new_slots_per_request > 0 self.parallel_drafting_token_id: int = 0 self.parallel_drafting_hidden_state_tensor: torch.Tensor | None = None if self.parallel_drafting: self._init_parallel_drafting_params() self.use_local_argmax_reduction: bool = ( self.speculative_config.use_local_argmax_reduction ) max_batch_size = vllm_config.scheduler_config.max_num_seqs self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens self.token_arange_np = np.arange(self.max_num_tokens) # Multi-modal data support self.mm_registry = MULTIMODAL_REGISTRY self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs( vllm_config.model_config ) self.attn_metadata_builder: AttentionMetadataBuilder | None = None self.draft_indexer_metadata_builder: AttentionMetadataBuilder | None = None self.attn_layer_names: list[str] = [] self.indexer_layer_names: list[str] = [] self.eagle3_use_aux_hidden_state: bool = ( self._get_eagle3_use_aux_hidden_state_from_config() ) self.compilation_config = self.vllm_config.compilation_config # Cudagraph dispatcher for PIECEWISE-only dispatching in eagle. # Keys are initialized later via initialize_cudagraph_keys() called from # gpu_model_runner._check_and_update_cudagraph_mode after # adjust_cudagraph_sizes_for_spec_decode is called. self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config) # persistent buffers for cuda graph self.input_ids = torch.zeros( self.max_num_tokens, dtype=torch.int32, device=device ) # Use draft model's M-RoPE setting, not target model's # Draft models may be text-only even if target is multimodal self.uses_mrope = self.draft_model_config.uses_mrope self.uses_xdrope_dim = self.vllm_config.model_config.uses_xdrope_dim self.draft_uses_xdrope_dim = self.draft_model_config.uses_xdrope_dim if self.uses_mrope: # NOTE: `mrope_positions` is implemented with one additional dummy # position on purpose to make it non-contiguous so that it can work # with torch compile. # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923 # NOTE: When M-RoPE is enabled, position ids are 3D regardless of # the modality of inputs. For text-only inputs, each dimension has # identical position IDs, making M-RoPE functionally equivalent to # 1D-RoPE. # See page 5 of https://arxiv.org/abs/2409.12191 self.mrope_positions = torch.zeros( (3, self.max_num_tokens + 1), dtype=torch.int64, device=device ) elif self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0: self.xdrope_positions = torch.zeros( (self.uses_xdrope_dim, self.max_num_tokens + 1), dtype=torch.int64, device=device, ) else: # RoPE need (max_num_tokens,) 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. max_num_slots_for_arange = max(max_batch_size + 1, self.max_num_tokens) self.arange = torch.arange( max_num_slots_for_arange, device=device, dtype=torch.int32 ) if self.needs_extra_input_slots: self._raise_if_padded_drafter_batch_disabled() self._raise_if_multimodal() self._raise_if_mrope() self.is_rejected_token_mask: torch.Tensor | None = None self.is_masked_token_mask: torch.Tensor | None = None if self.needs_extra_input_slots: # For draft models and parallel drafting, we need to keep track of # which tokens are rejected to update the slot mapping with padding slots. self.is_rejected_token_mask = torch.zeros( (self.max_num_tokens,), dtype=torch.bool, device=device ) # For parallel drafting, we also need to keep track of which tokens # are parallel-padding tokens used to sample at later positions. # We populate this tensor even when using draft models for simplicity. self.is_masked_token_mask = torch.zeros( (self.max_num_tokens,), dtype=torch.bool, device=device ) self.inputs_embeds = torch.zeros( (self.max_num_tokens, self.inputs_embeds_size), dtype=self.dtype, device=device, ) self.backup_next_token_ids = CpuGpuBuffer( max_batch_size, dtype=torch.int32, pin_memory=is_pin_memory_available(), device=device, with_numpy=True, ) self._slot_mapping_buffer = torch.zeros( self.max_num_tokens, dtype=torch.int64, device=device ) # Determine allowed attention backends once during initialization. self.allowed_attn_types: tuple | None = None if current_platform.is_rocm(): from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata rocm_types = [ TritonAttentionMetadata, RocmAttentionMetadata, ] # ROCM_AITER_FA is an optional backend # We check is_enabled() here to avoid importing the backend module during # auto-discovery when VLLM_ROCM_USE_AITER=0, which would trigger aiter # import and JIT compilation warnings. Explicit backend selection via # attention_config still works because the backend module is loaded # directly when selected, not through this auto-discovery path. # Check if backend module exists to allow explicit selection if find_spec( AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False) ): from vllm.v1.attention.backends.rocm_aiter_fa import ( AiterFlashAttentionMetadata, ) rocm_types.append(AiterFlashAttentionMetadata) # TRITON_MLA backend support for MLA models (e.g., DeepSeek) from vllm.model_executor.layers.attention.mla_attention import ( MLACommonMetadata, ) rocm_types.append(MLACommonMetadata) # FlexAttention backend support from vllm.v1.attention.backends.flex_attention import FlexAttentionMetadata rocm_types.append(FlexAttentionMetadata) self.allowed_attn_types = tuple(rocm_types) # Parse the speculative token tree. spec_token_tree = self.speculative_config.speculative_token_tree assert spec_token_tree is not None self.tree_choices: list[tuple[int, ...]] = ast.literal_eval(spec_token_tree) tree_depth = len(self.tree_choices[-1]) # Precompute per-level properties of the tree. num_drafts_per_level = [0] * tree_depth for node in self.tree_choices: num_drafts_per_level[len(node) - 1] += 1 self.cu_drafts_per_level = [num_drafts_per_level[0]] self.child_drafts_per_level = [num_drafts_per_level[0]] for level in range(1, tree_depth): self.cu_drafts_per_level.append( self.cu_drafts_per_level[-1] + num_drafts_per_level[level] ) self.child_drafts_per_level.append( num_drafts_per_level[level] // num_drafts_per_level[level - 1] ) # Precompute draft position offsets in flattened tree. self.tree_draft_pos_offsets = torch.arange( 1, len(self.tree_choices) + 1, device=device, dtype=torch.int32 ).repeat(max_batch_size, 1) def _raise_if_padded_drafter_batch_disabled(self): if self.speculative_config.disable_padded_drafter_batch: raise NotImplementedError( "Speculative Decoding with draft models or parallel drafting only " "supports padded drafter batch. Please unset " "disable_padded_drafter_batch in the speculative_config." ) def _raise_if_multimodal(self): if self.supports_mm_inputs: raise NotImplementedError( "Speculative Decoding with draft models or parallel drafting " "does not support multimodal models yet" ) def _raise_if_mrope(self): if self.draft_model_config.uses_mrope: raise NotImplementedError( "Speculative Decoding with draft models or parallel drafting " "does not support M-RoPE yet" ) def _init_parallel_drafting_params(self): # For parallel drafting, we need the token ID to use for masked slots # And for EAGLE + parallel drafting, we need the hidden state tensor to use # for those masked slots. model_hf_config = self.draft_model_config.hf_config if hasattr(model_hf_config, "pard_token"): self.parallel_drafting_token_id = model_hf_config.pard_token elif hasattr(model_hf_config, "ptd_token_id"): self.parallel_drafting_token_id = model_hf_config.ptd_token_id else: raise ValueError( "For parallel drafting, the draft model config must have " "`pard_token` or `ptd_token_id` specified in its config.json." ) if self.pass_hidden_states_to_model: self.parallel_drafting_hidden_state_tensor = torch.empty( self.hidden_size, dtype=self.dtype, device=self.device ) def _get_positions(self, num_tokens: int): if self.uses_mrope: return self.mrope_positions[:, :num_tokens] if self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0: return self.xdrope_positions[:, :num_tokens] return self.positions[:num_tokens] def _set_positions(self, num_tokens: int, positions: torch.Tensor): if self.uses_mrope: self.mrope_positions[:, :num_tokens] = positions elif self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0: self.xdrope_positions[:, :num_tokens] = positions else: # Convert M-RoPE positions if target model uses M-RoPE # but draft doesn't, For text inputs, all M-RoPE # dimensions are identical if self.vllm_config.model_config.uses_mrope: positions = positions[0] self.positions[:num_tokens] = positions def _get_slot_mapping( self, num_tokens: int, slot_mapping: torch.Tensor | None = None, ) -> dict[str, torch.Tensor]: """Return slot_mapping dict for EAGLE layers. If slot_mapping is provided, copies it into the buffer first. """ if slot_mapping is not None: num_actual = slot_mapping.shape[0] self._slot_mapping_buffer[:num_actual].copy_(slot_mapping) if num_tokens > num_actual: self._slot_mapping_buffer[num_actual:num_tokens].fill_(PADDING_SLOT_ID) view = self._slot_mapping_buffer[:num_tokens] return {name: view for name in self.attn_layer_names + self.indexer_layer_names} def initialize_cudagraph_keys(self, cudagraph_mode: CUDAGraphMode) -> None: """Initialize cudagraph dispatcher keys for eagle. Eagle only supports PIECEWISE cudagraphs (via mixed_mode). This should be called after adjust_cudagraph_sizes_for_spec_decode. """ if ( not self.speculative_config.enforce_eager and cudagraph_mode.mixed_mode() in [CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL] ): eagle_cudagraph_mode = CUDAGraphMode.PIECEWISE else: eagle_cudagraph_mode = CUDAGraphMode.NONE self.cudagraph_dispatcher.initialize_cudagraph_keys(eagle_cudagraph_mode) def _greedy_sample(self, hidden_states: torch.Tensor) -> torch.Tensor: """Greedy-sample draft tokens from hidden states.""" if self.use_local_argmax_reduction: return self.model.get_top_tokens(hidden_states) return self.model.compute_logits(hidden_states).argmax(dim=-1) def propose( self, # [num_tokens] target_token_ids: torch.Tensor, # [num_tokens] or [3, num_tokens] when M-RoPE is enabled target_positions: torch.Tensor, # [num_tokens, hidden_size] target_hidden_states: torch.Tensor, # [batch_size] next_token_ids: torch.Tensor, token_indices_to_sample: torch.Tensor | None, common_attn_metadata: CommonAttentionMetadata, sampling_metadata: SamplingMetadata, mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None, num_rejected_tokens_gpu: torch.Tensor | None = None, slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None, ) -> torch.Tensor: batch_size = common_attn_metadata.batch_size() 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 num_tokens, token_indices_to_sample, common_attn_metadata = ( self.set_inputs_first_pass( target_token_ids=target_token_ids, next_token_ids=next_token_ids, target_positions=target_positions, target_hidden_states=target_hidden_states, token_indices_to_sample=token_indices_to_sample, cad=common_attn_metadata, num_rejected_tokens_gpu=num_rejected_tokens_gpu, ) ) assert self.runner is not None if self.attn_metadata_builder is None: attn_metadata_builder = self._get_attention_metadata_builder() else: attn_metadata_builder = self.attn_metadata_builder attn_metadata = attn_metadata_builder.build_for_drafting( common_attn_metadata=common_attn_metadata, draft_index=0 ) # FIXME: support hybrid kv for draft model (remove separate indexer) if self.draft_indexer_metadata_builder: draft_indexer_metadata = ( self.draft_indexer_metadata_builder.build_for_drafting( common_attn_metadata=common_attn_metadata, draft_index=0, ) ) else: draft_indexer_metadata = None # 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 for layer_name in self.indexer_layer_names: assert draft_indexer_metadata is not None per_layer_attn_metadata[layer_name] = draft_indexer_metadata num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp( num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens ) cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch( num_tokens_dp_padded ) num_input_tokens = batch_desc.num_tokens if num_tokens_across_dp is not None: num_tokens_across_dp[self.dp_rank] = num_input_tokens if self.supports_mm_inputs: mm_embeds, is_mm_embed = mm_embed_inputs or (None, None) self.inputs_embeds[:num_tokens] = self.model.embed_input_ids( self.input_ids[:num_tokens], multimodal_embeddings=mm_embeds, is_multimodal=is_mm_embed, ) input_ids = None inputs_embeds = self.inputs_embeds[:num_input_tokens] else: input_ids = self.input_ids[:num_input_tokens] inputs_embeds = None model_kwargs = { "input_ids": input_ids, "positions": self._get_positions(num_input_tokens), "inputs_embeds": inputs_embeds, } if self.pass_hidden_states_to_model: model_kwargs["hidden_states"] = self.hidden_states[:num_input_tokens] with set_forward_context( per_layer_attn_metadata, self.vllm_config, num_tokens=num_input_tokens, num_tokens_across_dp=num_tokens_across_dp, cudagraph_runtime_mode=cudagraph_runtime_mode, slot_mapping=self._get_slot_mapping( num_input_tokens, common_attn_metadata.slot_mapping ), ): ret_hidden_states = self.model(**model_kwargs) if not self.model_returns_tuple(): last_hidden_states = ret_hidden_states hidden_states = last_hidden_states else: last_hidden_states, hidden_states = ret_hidden_states sample_hidden_states = last_hidden_states[token_indices_to_sample] # Early exit if there is only one draft token to be generated. if self.num_speculative_tokens == 1 or self.parallel_drafting: draft_token_ids = self._greedy_sample(sample_hidden_states) return draft_token_ids.view(-1, self.num_speculative_tokens) if self.uses_mrope: positions = self.mrope_positions[:, token_indices_to_sample] else: positions = self.positions[token_indices_to_sample] if self.method in ( "deepseek_mtp", "ernie_mtp", "longcat_flash_mtp", "pangu_ultra_moe_mtp", ): hidden_states = self.hidden_states[token_indices_to_sample] else: hidden_states = hidden_states[token_indices_to_sample] if isinstance(attn_metadata, TreeAttentionMetadata): # Draft using tree attention - requires full logits for top-k logits = self.model.compute_logits(sample_hidden_states) draft_token_ids_list = self.propose_tree( batch_size=batch_size, logits=logits, positions=positions, hidden_states=hidden_states, common_attn_metadata=common_attn_metadata, slot_mappings=slot_mappings, ) # [batch_size, num_tree_tokens] return torch.cat(draft_token_ids_list, dim=1) draft_token_ids = self._greedy_sample(sample_hidden_states) if self.allowed_attn_types is not None and not isinstance( attn_metadata, self.allowed_attn_types ): raise ValueError( f"Unsupported attention metadata type for speculative " "decoding with num_speculative_tokens > 1: " f"{type(attn_metadata)}. Supported types are: " f"{self.allowed_attn_types}" ) # Generate the remaining draft tokens. draft_token_ids_list = [draft_token_ids] batch_size_dp_padded, batch_size_across_dp = self._pad_batch_across_dp( num_tokens_unpadded=batch_size, num_tokens_padded=batch_size ) cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch( batch_size_dp_padded ) input_batch_size = batch_desc.num_tokens if batch_size_across_dp is not None: batch_size_across_dp[self.dp_rank] = input_batch_size common_attn_metadata.num_actual_tokens = batch_size common_attn_metadata.max_query_len = 1 common_attn_metadata.query_start_loc = self.arange[: batch_size + 1] common_attn_metadata.query_start_loc_cpu = torch.from_numpy( self.token_arange_np[: batch_size + 1] ).clone() # In padded drafter batch, we need to adjust the sequence lengths # to remove the "padding" (i.e. rejected tokens). # Only apply this adjustment when we have rejected tokens # (i.e., not the first proposal). if self.num_speculative_tokens > 1 and num_rejected_tokens_gpu is not None: common_attn_metadata.seq_lens -= num_rejected_tokens_gpu # Invalidate the CPU-side shadows to avoid H<>D sync. common_attn_metadata._seq_lens_cpu = None common_attn_metadata._num_computed_tokens_cpu = None for token_index 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() if self.uses_mrope: 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[0] >= 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.unsqueeze(0), torch.zeros_like(positions), positions, ) else: positions += 1 exceeds_max_model_len = positions >= self.max_model_len clamped_positions = torch.where(exceeds_max_model_len, 0, positions) # For data integrity when async scheduling, we shouldn't use in place # operations in case they are modified in next step's `prepare_input` # of main model. # Increment the sequence lengths. common_attn_metadata.seq_lens += 1 # For the requests that exceed the max model length, we set the # sequence length to 1 to minimize their overheads in attention. common_attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1) # Increment the maximum sequence length. We increment max_seq_len # unconditionally even though some seq_lens may have been capped above, # as max_seq_len serves as an upper bound for sequence lengths. common_attn_metadata.max_seq_len = min( common_attn_metadata.max_seq_len + 1, self.max_model_len ) # Also update the CPU-side shadow; NOTE: this is hacky and should be # removed in when common_attn_metadata.seq_lens_cpu is deprecated. if common_attn_metadata._seq_lens_cpu is not None: common_attn_metadata._seq_lens_cpu += 1 if common_attn_metadata._num_computed_tokens_cpu is not None: common_attn_metadata._num_computed_tokens_cpu += 1 # Compute the slot mapping. block_size = attn_metadata_builder.kv_cache_spec.block_size if self.uses_mrope: # all dimensions of positions are the same block_numbers = clamped_positions[0] // block_size else: block_numbers = clamped_positions // block_size block_ids = common_attn_metadata.block_table_tensor.gather( dim=1, index=block_numbers.view(-1, 1) ) block_ids = block_ids.view(-1) if self.uses_mrope: common_attn_metadata.slot_mapping = ( block_ids * block_size + clamped_positions[0] % block_size ) else: common_attn_metadata.slot_mapping = ( block_ids * block_size + clamped_positions % 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. common_attn_metadata.slot_mapping.masked_fill_( exceeds_max_model_len, PADDING_SLOT_ID ) # Rebuild attention metadata attn_metadata = attn_metadata_builder.build_for_drafting( # type: ignore common_attn_metadata=common_attn_metadata, draft_index=token_index + 1 ) for layer_name in self.attn_layer_names: per_layer_attn_metadata[layer_name] = attn_metadata # copy inputs to buffer for cudagraph self.input_ids[:batch_size] = input_ids self._set_positions(batch_size, clamped_positions) self.hidden_states[:batch_size] = hidden_states if self.supports_mm_inputs: self.inputs_embeds[:batch_size] = self.model.embed_input_ids(input_ids) input_ids = None inputs_embeds = self.inputs_embeds[:input_batch_size] else: input_ids = self.input_ids[:input_batch_size] inputs_embeds = None # Run the model. model_kwargs = { "input_ids": input_ids, "positions": self._get_positions(input_batch_size), "inputs_embeds": inputs_embeds, } if self.pass_hidden_states_to_model: model_kwargs["hidden_states"] = self.hidden_states[:input_batch_size] with set_forward_context( per_layer_attn_metadata, self.vllm_config, num_tokens=input_batch_size, num_tokens_across_dp=batch_size_across_dp, cudagraph_runtime_mode=cudagraph_runtime_mode, slot_mapping=self._get_slot_mapping( input_batch_size, common_attn_metadata.slot_mapping ), ): ret_hidden_states = self.model(**model_kwargs) if not self.model_returns_tuple(): last_hidden_states = ret_hidden_states hidden_states = ret_hidden_states else: last_hidden_states, hidden_states = ret_hidden_states hidden_states = hidden_states[:batch_size] draft_token_ids = self._greedy_sample(last_hidden_states[:batch_size]) 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 set_inputs_first_pass( self, target_token_ids: torch.Tensor, next_token_ids: torch.Tensor, target_positions: torch.Tensor, target_hidden_states: torch.Tensor, token_indices_to_sample: torch.Tensor | None, cad: CommonAttentionMetadata, num_rejected_tokens_gpu: torch.Tensor | None, ) -> tuple[int, torch.Tensor, CommonAttentionMetadata]: if not self.needs_extra_input_slots: # Default EAGLE pathway: no reshaping of input tensors needed. # Simply rotate the input ids and leave the positions unchanged, # Inserting the next token ids at the last slot in each request. if token_indices_to_sample is None: token_indices_to_sample = cad.query_start_loc[1:] - 1 num_tokens = target_token_ids.shape[0] # 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[token_indices_to_sample] = next_token_ids # copy inputs to buffer for cudagraph if self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim == 0: target_positions = target_positions[0] self._set_positions(num_tokens, target_positions) self.hidden_states[:num_tokens] = target_hidden_states return num_tokens, token_indices_to_sample, cad else: assert self.is_rejected_token_mask is not None assert self.is_masked_token_mask is not None # 1. # Call a custom triton kernel to copy input_ids and positions # into the correct slots in the preallocated buffers self.input_ids, # self.positions. batch_size = cad.batch_size() # Since we might have to copy a lot of data for prefills, we select the # block size based on the max query length and limit to max 256 slots/block. max_num_tokens_per_request = ( cad.max_query_len + self.net_num_new_slots_per_request ) BLOCK_SIZE_TOKENS = min( 256, triton.next_power_of_2(max_num_tokens_per_request) ) num_blocks = ( max_num_tokens_per_request + BLOCK_SIZE_TOKENS - 1 ) // BLOCK_SIZE_TOKENS total_num_input_tokens = target_token_ids.shape[0] total_num_output_tokens = total_num_input_tokens + ( self.net_num_new_slots_per_request * batch_size ) token_indices_to_sample = torch.empty( batch_size * self.extra_slots_per_request, dtype=torch.int32, device=self.device, ) # Destination indices to write target_hidden_states into drafting buffer. out_hidden_state_mapping = torch.empty( total_num_input_tokens, dtype=torch.int32, device=self.device ) # Kernel grid: one program per request (row) grid = (batch_size, num_blocks) query_start_loc = cad.query_start_loc query_end_loc = cad.query_start_loc[1:] - 1 if num_rejected_tokens_gpu is not None: query_end_loc = query_end_loc - num_rejected_tokens_gpu copy_and_expand_eagle_inputs_kernel[grid]( # (Padded) Inputs from the target model target_token_ids_ptr=target_token_ids, target_positions_ptr=target_positions, next_token_ids_ptr=next_token_ids, # sampled tokens, one per request # Outputs to the drafting buffers out_input_ids_ptr=self.input_ids, out_positions_ptr=self.positions, # Doesn't support mrope for now out_is_rejected_token_mask_ptr=self.is_rejected_token_mask, out_is_masked_token_mask_ptr=self.is_masked_token_mask, out_new_token_indices_ptr=token_indices_to_sample, out_hidden_state_mapping_ptr=out_hidden_state_mapping, # Input metadata query_start_loc_ptr=query_start_loc, query_end_loc_ptr=query_end_loc, padding_token_id=0, parallel_drafting_token_id=self.parallel_drafting_token_id, # Sizing info # Note that we can deduce batch_size for free from the grid size total_input_tokens=total_num_input_tokens, num_padding_slots_per_request=self.extra_slots_per_request, shift_input_ids=self.pass_hidden_states_to_model, BLOCK_SIZE_TOKENS=BLOCK_SIZE_TOKENS, ) if self.pass_hidden_states_to_model: assert self.parallel_drafting_hidden_state_tensor is not None self.hidden_states[out_hidden_state_mapping] = target_hidden_states # Use torch.where to avoid DtoH sync from boolean indexing mask = self.is_masked_token_mask[:total_num_output_tokens] torch.where( mask.unsqueeze(1), self.parallel_drafting_hidden_state_tensor, self.hidden_states[:total_num_output_tokens], out=self.hidden_states[:total_num_output_tokens], ) # 2. # Recompute the slot mapping based on the new positions and # rejection mask. builder = ( self._get_attention_metadata_builder() if self.attn_metadata_builder is None else self.attn_metadata_builder ) new_slot_mapping = compute_new_slot_mapping( cad=cad, new_positions=self.positions[:total_num_output_tokens], is_rejected_token_mask=self.is_rejected_token_mask[ :total_num_output_tokens ], block_size=builder.kv_cache_spec.block_size, num_new_tokens=self.net_num_new_slots_per_request, max_model_len=self.max_model_len, ) # 3. Update the common attention metadata with the new (meta)data new_cad = extend_all_queries_by_N( cad, N=self.net_num_new_slots_per_request, arange=self.arange, new_slot_mapping=new_slot_mapping, ) return total_num_output_tokens, token_indices_to_sample, new_cad def model_returns_tuple(self) -> bool: return self.method not in ("mtp", "draft_model") def prepare_next_token_ids_cpu( self, sampled_token_ids: list[list[int]], requests: dict[str, CachedRequestState], gpu_input_batch: InputBatch, num_scheduled_tokens: dict[str, int], ) -> torch.Tensor: """ This function is used to prepare the inputs for speculative decoding. It calculates the next token ids for each request based on the sampled token ids from the CPU. If a request has no sampled token ids (e.g., during the initial decoding steps), it falls back to using the request state to get the next token id. """ req_ids = gpu_input_batch.req_ids next_token_ids: list[int] = [] for i, token_ids in enumerate(sampled_token_ids): if token_ids: # Common case. next_token_id = token_ids[-1] else: # Partial prefill (rare case). # Get the next token id from the request state. req_id = req_ids[i] req_state = requests[req_id] seq_len = req_state.num_computed_tokens + num_scheduled_tokens[req_id] next_token_id = req_state.get_token_id(seq_len) next_token_ids.append(next_token_id) next_token_ids = torch.tensor( next_token_ids, dtype=torch.int32, device=self.input_ids.device ) return next_token_ids def prepare_next_token_ids_padded( self, common_attn_metadata: CommonAttentionMetadata, sampled_token_ids: torch.Tensor, requests: dict[str, CachedRequestState], gpu_input_batch: InputBatch, discard_request_mask: torch.Tensor, ) -> tuple[torch.Tensor, torch.Tensor]: """ This function is used to prepare the inputs for speculative decoding. It calculates the next token ids and the number of valid sampled tokens for each request, considering the "discarded" requests whose next token is not sampled and comes from `request.get_token_id()` instead. This is denoted the "backup" token id. It also counts rejected tokens via `sampled_token_ids`. """ # Precompute get_token_id for when there is no valid next token num_reqs = gpu_input_batch.num_reqs self.backup_next_token_ids.np[:num_reqs] = np.array( [ requests[gpu_input_batch.req_ids[i]].get_token_id( common_attn_metadata.seq_lens_cpu[i].item() ) for i in range(num_reqs) ], dtype=np.int32, ) self.backup_next_token_ids.copy_to_gpu(num_reqs) backup_tokens_gpu = self.backup_next_token_ids.gpu batch_size, num_tokens = sampled_token_ids.shape device = sampled_token_ids.device assert discard_request_mask.dtype == torch.bool assert backup_tokens_gpu.dtype == torch.int32 next_token_ids = torch.empty(batch_size, dtype=torch.int32, device=device) valid_sampled_tokens_count = next_token_ids.new_empty(batch_size) # Kernel grid: one program per request (row) grid = (batch_size,) # Find the next power of 2 for block sizes BLOCK_SIZE_TOKENS = triton.next_power_of_2(num_tokens) eagle_prepare_next_token_padded_kernel[grid]( sampled_token_ids, discard_request_mask, backup_tokens_gpu, next_token_ids, valid_sampled_tokens_count, gpu_input_batch.vocab_size, num_tokens, batch_size, sampled_token_ids.stride(0), BLOCK_SIZE_TOKENS=BLOCK_SIZE_TOKENS, ) return next_token_ids, valid_sampled_tokens_count def prepare_inputs_padded( self, common_attn_metadata: CommonAttentionMetadata, spec_decode_metadata: SpecDecodeMetadata, valid_sampled_tokens_count: torch.Tensor, ) -> tuple[CommonAttentionMetadata, torch.Tensor, torch.Tensor]: """ This function is used to prepare the inputs for speculative decoding It updates the common_attn_metadata for speculative decoding, but does not consider the rejected tokens. Instead, all tokens are included as inputs to the speculator, with the rejected tokens used as padding and filtered out later by `token_indices_to_sample`. No blocking CPU operations should be introduced in this function. """ num_reqs = common_attn_metadata.num_reqs device = valid_sampled_tokens_count.device token_indices_to_sample = torch.empty( (num_reqs,), dtype=torch.int32, device=device ) num_rejected_tokens_gpu = torch.empty( (num_reqs,), dtype=torch.int32, device=device ) grid = (num_reqs,) eagle_prepare_inputs_padded_kernel[grid]( spec_decode_metadata.cu_num_draft_tokens, valid_sampled_tokens_count, common_attn_metadata.query_start_loc, token_indices_to_sample, num_rejected_tokens_gpu, num_reqs, ) query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] total_num_tokens = query_start_loc_cpu[-1].item() spec_common_attn_metadata = CommonAttentionMetadata( query_start_loc=common_attn_metadata.query_start_loc, seq_lens=common_attn_metadata.seq_lens, query_start_loc_cpu=query_start_loc_cpu, _seq_lens_cpu=common_attn_metadata._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(), max_seq_len=common_attn_metadata.seq_lens_cpu.max().item(), block_table_tensor=common_attn_metadata.block_table_tensor, slot_mapping=common_attn_metadata.slot_mapping[:total_num_tokens], causal=True, dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens, ) return ( spec_common_attn_metadata, token_indices_to_sample, num_rejected_tokens_gpu, ) def propose_tree( self, batch_size: int, # [num_tokens, vocab_size] logits: torch.Tensor, # [num_tokens] positions: torch.Tensor, # [num_tokens, hidden_size] hidden_states: torch.Tensor, common_attn_metadata: CommonAttentionMetadata, slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None, ) -> list[torch.Tensor]: tree_attn_metadata_builder = self.runner.attn_groups[0][ 0 ].get_metadata_builder() assert isinstance(tree_attn_metadata_builder, TreeAttentionMetadataBuilder) total_num_drafts = self.cu_drafts_per_level[0] level_num_drafts = total_num_drafts # Sample a draft token for each child at the tree root level. num_children = self.child_drafts_per_level[0] if num_children == 1: draft_token_ids = logits.argmax(dim=-1).view(batch_size, -1) else: draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view( batch_size, -1 ) draft_token_ids_list = [draft_token_ids] draft_hidden_states = hidden_states.view(batch_size, 1, -1) # Initialize empty tensors for concatenation with the level outputs. tree_input_ids = torch.empty( 0, device=self.input_ids.device, dtype=self.input_ids.dtype ) tree_positions = torch.empty( 0, device=self.positions.device, dtype=self.positions.dtype ) tree_hidden_states = torch.empty( 0, device=self.hidden_states.device, dtype=self.hidden_states.dtype ) # Precompute the draft token positions. flattened_draft_positions = ( positions.view(batch_size, -1) + self.tree_draft_pos_offsets[:batch_size, :] ) tree_depth = len(self.cu_drafts_per_level) for level in range(tree_depth - 1): # Get draft positions for RoPE. draft_positions = positions + (level + 1) exceeds_max_model_len = (positions + total_num_drafts) >= 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. draft_positions = torch.where( exceeds_max_model_len, 0, draft_positions, ).view(batch_size, -1) if level_num_drafts > 1: # Repeat the positions for each draft at this level. draft_positions = draft_positions.repeat_interleave( level_num_drafts, dim=1 ) if num_children > 1: # Repeat draft hidden states for each child. draft_hidden_states = draft_hidden_states.repeat_interleave( num_children, dim=1 ) # Concatenate the draft tokens, positions, and hidden states. tree_input_ids = torch.cat([tree_input_ids, draft_token_ids], dim=1) tree_positions = torch.cat([tree_positions, draft_positions], dim=1) tree_hidden_states = torch.cat( [tree_hidden_states, draft_hidden_states], dim=1 ) # Build new attention metadata for the next level of drafts. # This is necessary to support tree attention. query_len = total_num_drafts common_attn_metadata = replace( common_attn_metadata, query_start_loc=query_len * self.arange[: batch_size + 1], seq_lens=common_attn_metadata.seq_lens + level_num_drafts, num_actual_tokens=batch_size * query_len, max_query_len=query_len, ) attn_metadata = tree_attn_metadata_builder.build_for_drafting( common_attn_metadata=common_attn_metadata, draft_index=level + 1 ) # Apply new attention metadata to all layers. per_layer_attn_metadata = {} for layer_name in self.attn_layer_names: per_layer_attn_metadata[layer_name] = attn_metadata # 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_size = tree_attn_metadata_builder.kv_cache_spec.block_size query_positions = flattened_draft_positions[:, level : level + query_len] block_numbers = query_positions // block_size block_ids = attn_metadata.block_table.gather(dim=1, index=block_numbers) slot_mapping = block_ids * block_size + query_positions % 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. slot_mapping[exceeds_max_model_len] = PADDING_SLOT_ID attn_metadata.slot_mapping = slot_mapping.view(-1) # Copy inputs to buffer for cudagraph. num_tokens = attn_metadata.num_actual_tokens input_ids = tree_input_ids.view(-1) self.input_ids[:num_tokens] = input_ids self.positions[:num_tokens] = tree_positions.view(-1) self.hidden_states[:num_tokens] = tree_hidden_states.view(num_tokens, -1) cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch( num_tokens ) num_input_tokens = batch_desc.num_tokens # Run the model. with set_forward_context( per_layer_attn_metadata, self.vllm_config, num_tokens=num_input_tokens, cudagraph_runtime_mode=cudagraph_runtime_mode, slot_mapping=self._get_slot_mapping( num_input_tokens, attn_metadata.slot_mapping ), ): 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], inputs_embeds=None, ) # Get the output hidden states for the draft tokens. draft_hidden_states = hidden_states[:num_tokens].view( batch_size, query_len, -1 )[:, -level_num_drafts:] draft_last_hidden_states = last_hidden_states[:num_tokens].view( batch_size, query_len, -1 )[:, -level_num_drafts:] # Get the output logits for the draft tokens. logits = self.model.compute_logits( draft_last_hidden_states.reshape(batch_size * level_num_drafts, -1) ) # Sample a draft token for each child at the next tree level. num_children = self.child_drafts_per_level[level + 1] if num_children == 1: draft_token_ids = logits.argmax(dim=-1).view(batch_size, -1) else: draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view( batch_size, -1 ) draft_token_ids_list.append(draft_token_ids) # Update the # drafts counters for the next tree level. level_num_drafts = self.cu_drafts_per_level[level + 1] - total_num_drafts total_num_drafts = self.cu_drafts_per_level[level + 1] return draft_token_ids_list def prepare_inputs( self, common_attn_metadata: CommonAttentionMetadata, sampled_token_ids: list[list[int]], num_draft_tokens: list[int], ) -> tuple[CommonAttentionMetadata, torch.Tensor]: """ This function is used to prepare the inputs for speculative decoding. 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] num_rejected_tokens = [ n + 1 - len(sampled_token_ids[i]) if n > 0 else 0 for i, n in enumerate(num_draft_tokens) ] num_rejected_tokens = torch.tensor(num_rejected_tokens, dtype=torch.int32) 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_offsets = ( 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_offsets + 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(), max_seq_len=new_seq_lens_cpu.max().item(), block_table_tensor=common_attn_metadata.block_table_tensor, slot_mapping=common_attn_metadata.slot_mapping[token_indices], causal=True, dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens, ) return spec_common_attn_metadata, token_indices def get_model_name(self, model: nn.Module) -> str: if hasattr(model, "module"): # multi-GPU model = model.module return model.__class__.__name__ def _get_model(self) -> nn.Module: """ Default method to call get_model(). Can be overridden by subclasses which need to customize model loading. """ from vllm.compilation.backends import set_model_tag with set_model_tag("eagle_head"): model = get_model( vllm_config=self.vllm_config, model_config=self.speculative_config.draft_model_config, load_config=self.speculative_config.draft_load_config, ) return model def load_model(self, target_model: nn.Module) -> None: target_attn_layer_names = set( get_layers_from_vllm_config( self.vllm_config, AttentionLayerBase, # type: ignore[type-abstract] ).keys() ) # FIXME: support hybrid kv for draft model target_indexer_layer_names = set( get_layers_from_vllm_config( self.vllm_config, DeepseekV32IndexerCache ).keys() ) self.model = self._get_model() draft_attn_layer_names = ( get_layers_from_vllm_config( self.vllm_config, AttentionLayerBase, # type: ignore[type-abstract] ).keys() - target_attn_layer_names ) indexer_layers = get_layers_from_vllm_config( self.vllm_config, DeepseekV32IndexerCache ) draft_indexer_layer_names = indexer_layers.keys() - target_indexer_layer_names self.attn_layer_names = list(draft_attn_layer_names - draft_indexer_layer_names) self.indexer_layer_names = list(draft_indexer_layer_names) if self.indexer_layer_names: first_layer = self.indexer_layer_names[0] self.draft_indexer_metadata_builder = ( indexer_layers[first_layer] .get_attn_backend() .get_builder_cls()( indexer_layers[first_layer].get_kv_cache_spec(self.vllm_config), self.indexer_layer_names, self.vllm_config, self.device, ) ) else: self.draft_indexer_metadata_builder = None if self.supports_mm_inputs: # Even if the target model is multimodal, we can also use # text-only draft models try: dummy_input_ids = torch.tensor([[1]], device=self.input_ids.device) self.model.embed_input_ids(dummy_input_ids, multimodal_embeddings=None) except (NotImplementedError, AttributeError, TypeError): logger.warning( "Draft model does not support multimodal inputs, " "falling back to text-only mode" ) self.supports_mm_inputs = False if supports_multimodal(target_model): # handle multimodality assert hasattr(target_model, "config") if self.get_model_name(target_model) in [ "Qwen2_5_VLForConditionalGeneration", "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "HunYuanVLForConditionalGeneration", "GlmOcrForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration", ]: self.model.config.image_token_index = target_model.config.image_token_id elif self.get_model_name(target_model) == "PixtralForConditionalGeneration": self.model.config.image_token_index = ( target_model.config.vision_config.image_token_id ) else: self.model.config.image_token_index = ( target_model.config.image_token_index ) target_language_model = cast( SupportsMultiModal, target_model ).get_language_model() else: target_language_model = target_model self._maybe_share_embeddings(target_language_model) self._maybe_share_lm_head(target_language_model) if self.parallel_drafting and self.pass_hidden_states_to_model: assert self.parallel_drafting_hidden_state_tensor is not None self.parallel_drafting_hidden_state_tensor.copy_( self.model.combine_hidden_states( self.model.mask_hidden.view(3 * self.hidden_size) ) if self.eagle3_use_aux_hidden_state else self.model.mask_hidden.view(self.hidden_size) ) def _maybe_share_embeddings(self, target_language_model: nn.Module) -> None: """ Some draft models may not have their own embedding layers, and some may have a duplicate copy of the target model's embedding layers. In these cases, we share the target model's embedding layers with the draft model to save memory. """ if get_pp_group().world_size == 1: inner_model = getattr(target_language_model, "model", None) if inner_model is None: raise AttributeError("Target model does not have 'model' attribute") if hasattr(inner_model, "embed_tokens"): target_embed_tokens = inner_model.embed_tokens elif hasattr(inner_model, "embedding"): target_embed_tokens = inner_model.embedding else: raise AttributeError( "Target model does not have 'embed_tokens' or 'embedding' attribute" ) share_embeddings = False if hasattr(self.model, "has_own_embed_tokens"): # EAGLE model if not self.model.has_own_embed_tokens: share_embeddings = True logger.info( "Detected EAGLE model without its own embed_tokens in the" " checkpoint. Sharing target model embedding weights with the" " draft model." ) elif ( isinstance(target_embed_tokens.weight, torch.Tensor) and isinstance(self.model.model.embed_tokens.weight, torch.Tensor) # TODO: Offload to CPU for comparison to avoid extra GPU memory # usage in CI testing environments with limited GPU memory and torch.equal( target_embed_tokens.weight.cpu(), self.model.model.embed_tokens.weight.cpu(), ) ): share_embeddings = True logger.info( "Detected EAGLE model with embed_tokens identical to the target" " model. Sharing target model embedding weights with the draft" " model." ) else: logger.info( "Detected EAGLE model with distinct embed_tokens weights. " "Keeping separate embedding weights from the target model." ) else: # MTP model share_embeddings = True logger.info( "Detected MTP model. " "Sharing target model embedding weights with the draft model." ) if share_embeddings: if hasattr(self.model.model, "embed_tokens"): del self.model.model.embed_tokens self.model.model.embed_tokens = target_embed_tokens else: logger.info( "The draft model's vocab embedding will be loaded separately" " from the target model." ) def _maybe_share_lm_head(self, target_language_model: nn.Module) -> None: """ Some draft models may not have their own LM head, and some may have a duplicate copy of the target model's LM head. In these cases, we share the target model's LM head with the draft model to save memory. """ share_lm_head = False if hasattr(self.model, "has_own_lm_head"): # EAGLE model if not self.model.has_own_lm_head: share_lm_head = True logger.info( "Detected EAGLE model without its own lm_head in the checkpoint. " "Sharing target model lm_head weights with the draft model." ) elif ( hasattr(target_language_model, "lm_head") and isinstance(target_language_model.lm_head.weight, torch.Tensor) and isinstance(self.model.lm_head.weight, torch.Tensor) # TODO: Offload to CPU for comparison to avoid extra GPU memory # usage in CI testing environments with limited GPU memory and torch.equal( target_language_model.lm_head.weight.cpu(), self.model.lm_head.weight.cpu(), ) ): share_lm_head = True logger.info( "Detected EAGLE model with lm_head identical to the target model. " "Sharing target model lm_head weights with the draft model." ) else: logger.info( "Detected EAGLE model with distinct lm_head weights. " "Keeping separate lm_head weights from the target model." ) else: # MTP model share_lm_head = True logger.info( "Detected MTP model. " "Sharing target model lm_head weights with the draft model." ) if share_lm_head and hasattr(target_language_model, "lm_head"): if hasattr(self.model, "lm_head"): del self.model.lm_head self.model.lm_head = target_language_model.lm_head # MTP models call compute_logits via shared_head.head (a # ParallelLMHead inside each MTP layer), not self.model.lm_head. # If the checkpoint omits a copy of the lm_head weights at the # MTP layer path, shared_head.head stays uninitialised and # produces NaN logits. Always share it explicitly. inner = getattr(self.model, "model", None) layers = getattr(inner, "layers", None) if inner else None if layers is not None: items = layers.values() if isinstance(layers, nn.ModuleDict) else layers for layer in items: sh = getattr(layer, "shared_head", None) if sh is not None and hasattr(sh, "head"): del sh.head sh.head = target_language_model.lm_head logger.info( "Shared target model lm_head with MTP shared_head.head." ) if self.use_local_argmax_reduction: if not hasattr(self.model, "get_top_tokens"): raise ValueError( "use_local_argmax_reduction is enabled but draft model " f"{self.model.__class__.__name__} does not implement " "get_top_tokens()." ) # Warn if draft model has vocab remapping, which forces fallback # to the full-logits path (negating the optimization). if ( hasattr(self.model, "draft_id_to_target_id") and self.model.draft_id_to_target_id is not None ): logger.warning( "use_local_argmax_reduction is enabled but draft model " "uses draft_id_to_target_id vocab remapping. The " "optimization will be bypassed (falling back to full " "logits gather + argmax)." ) else: logger.info( "Using local argmax reduction for draft token generation " "(communication: O(2*tp_size) vs O(vocab_size))." ) @torch.inference_mode() def dummy_run( self, num_tokens: int, use_cudagraphs: bool = True, is_graph_capturing: bool = False, slot_mappings: dict[str, torch.Tensor] | None = None, ) -> None: # FIXME: when using tree-based specdec, adjust number of forward-passes # according to the depth of the tree. for fwd_idx in range( self.num_speculative_tokens if not is_graph_capturing else 1 ): if fwd_idx <= 1: num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp( num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens ) if use_cudagraphs: cudagraph_runtime_mode, batch_desc = ( self.cudagraph_dispatcher.dispatch(num_tokens_dp_padded) ) num_input_tokens = batch_desc.num_tokens else: cudagraph_runtime_mode = CUDAGraphMode.NONE num_input_tokens = num_tokens_dp_padded if num_tokens_across_dp is not None: num_tokens_across_dp[self.dp_rank] = num_input_tokens # Make sure to use EAGLE's own buffer during cudagraph capture. if ( self.attn_layer_names and slot_mappings is not None and self.attn_layer_names[0] in slot_mappings ): slot_mapping_dict = self._get_slot_mapping(num_input_tokens) else: slot_mapping_dict = slot_mappings or {} with set_forward_context( None, self.vllm_config, num_tokens=num_input_tokens, num_tokens_across_dp=num_tokens_across_dp, cudagraph_runtime_mode=cudagraph_runtime_mode, slot_mapping=slot_mapping_dict, ): if self.supports_mm_inputs: input_ids = None inputs_embeds = self.inputs_embeds[:num_input_tokens] else: input_ids = self.input_ids[:num_input_tokens] inputs_embeds = None kwargs = dict( input_ids=input_ids, positions=self._get_positions(num_input_tokens), inputs_embeds=inputs_embeds, ) if self.pass_hidden_states_to_model: kwargs["hidden_states"] = self.hidden_states[:num_input_tokens] self.model(**kwargs) def _get_attention_metadata_builder(self) -> AttentionMetadataBuilder: """Find and return the attention metadata builders for EAGLE layers. Returns: The metadata builders for EAGLE layers. Raises: AssertionError: If no metadata builders are found for EAGLE layers. """ builder = None chosen_layer = self.attn_layer_names[0] for kv_cache_group in self.runner.attn_groups: for attn_group in kv_cache_group: if chosen_layer in attn_group.layer_names: builder = attn_group.get_metadata_builder() break if builder is not None: break assert builder is not None, ( "Failed to find attention metadata builder for EAGLE layers." ) return builder def _get_eagle3_use_aux_hidden_state_from_config(self) -> bool: """ Some eagle3 heads (e.g., nvidia/gpt-oss-120b-Eagle3-v2) do not use auxiliary hidden states and directly uses the last layer output just like eagle1. They might indicate this by setting "use_aux_hidden_state" to False inside the "eagle_config" dict of their hf_config. """ if self.method != "eagle3": return False # Assume that eagle3 heads use aux hidden states by default use_aux_hidden_state = True eagle_config = getattr(self.draft_model_config.hf_config, "eagle_config", None) if eagle_config is not None: use_aux_hidden_state = eagle_config.get("use_aux_hidden_state", True) return use_aux_hidden_state def validate_same_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None: """ Validate that all drafting layers belong to the same KVCacheGroup. Need this assumption to ensure all drafting 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 drafting layers should belong to the same kv cache group" def _pad_batch_across_dp( self, num_tokens_unpadded: int, num_tokens_padded: int, ) -> tuple[int, torch.Tensor]: # TODO(Flechman): support DBO ubatching should_ubatch, num_toks_across_dp, _ = coordinate_batch_across_dp( num_tokens_unpadded=num_tokens_unpadded, parallel_config=self.vllm_config.parallel_config, allow_microbatching=False, allow_dp_padding=self.cudagraph_dispatcher.cudagraph_mode != CUDAGraphMode.NONE, num_tokens_padded=num_tokens_padded, uniform_decode=None, num_scheduled_tokens_per_request=None, ) assert not should_ubatch, "DBO ubatching not implemented for EAGLE" num_tokens_dp_padded = num_tokens_padded if num_toks_across_dp is not None: num_tokens_dp_padded = int(num_toks_across_dp[self.dp_rank].item()) return num_tokens_dp_padded, num_toks_across_dp class EagleProposer(SpecDecodeBaseProposer): def __init__( self, vllm_config: VllmConfig, device: torch.device, runner=None, ): super().__init__( vllm_config, device, pass_hidden_states_to_model=True, runner=runner, ) # 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 assert sampling_metadata.temperature is not None # Use epsilon comparison to detect greedy sampling (temperature ~ 0.0) # consistent with sampler.py's _SAMPLING_EPS threshold temperature = sampling_metadata.temperature # Avoid division by zero if there are greedy requests. if not sampling_metadata.all_random: is_greedy = temperature < _SAMPLING_EPS temperature = torch.where(is_greedy, 1.0, 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