[Model Runner V2] Support Eagle3 (no CUDA graph) (#35029)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
This commit is contained in:
Woosuk Kwon
2026-02-21 12:55:24 -08:00
committed by GitHub
parent 965fe45935
commit a4047d4ea9
7 changed files with 169 additions and 49 deletions
+1 -1
View File
@@ -9,7 +9,7 @@ def init_speculator(vllm_config: VllmConfig, device: torch.device):
speculative_config = vllm_config.speculative_config
assert speculative_config is not None
if speculative_config.use_eagle():
from vllm.v1.worker.gpu.spec_decode.eagle import EagleSpeculator
from vllm.v1.worker.gpu.spec_decode.eagle.speculator import EagleSpeculator
return EagleSpeculator(vllm_config, device)
raise NotImplementedError(f"{speculative_config.method} is not supported yet.")
@@ -0,0 +1,46 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import cast
import torch.nn as nn
from vllm.config import SpeculativeConfig
from vllm.logger import init_logger
from vllm.model_executor.models.interfaces import SupportsEagle3, supports_eagle3
logger = init_logger(__name__)
def set_eagle3_aux_hidden_state_layers(
model: nn.Module,
spec_config: SpeculativeConfig,
) -> None:
if not supports_eagle3(model):
raise RuntimeError("Model does not support EAGLE3 interface")
# mypy may infer the class-level overload for supports_eagle3.
# Narrow explicitly to the runtime protocol instance.
if isinstance(model, type):
raise RuntimeError("Expected model instance for EAGLE3 configuration")
eagle3_model = cast(SupportsEagle3, model)
aux_layers = get_eagle3_aux_layers_from_config(spec_config)
if aux_layers:
logger.info("Using Eagle3 auxiliary layers from config: %s", aux_layers)
else:
aux_layers = eagle3_model.get_eagle3_aux_hidden_state_layers()
logger.info("Using Eagle3 auxiliary layers from model: %s", aux_layers)
eagle3_model.set_aux_hidden_state_layers(aux_layers)
def get_eagle3_aux_layers_from_config(
spec_config: SpeculativeConfig,
) -> tuple[int, ...] | None:
if not (spec_config and spec_config.draft_model_config):
return None
hf_config = spec_config.draft_model_config.hf_config
if not hasattr(hf_config, "eagle_aux_hidden_state_layer_ids"):
return None
layer_ids = hf_config.eagle_aux_hidden_state_layer_ids
if layer_ids and isinstance(layer_ids, (list, tuple)):
return tuple(layer_ids)
return None
@@ -9,7 +9,6 @@ from vllm.config import VllmConfig
from vllm.config.compilation import CUDAGraphMode
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.logger import init_logger
from vllm.model_executor.model_loader import get_model
from vllm.triton_utils import tl, triton
from vllm.v1.attention.backend import AttentionMetadataBuilder
from vllm.v1.kv_cache_interface import KVCacheConfig
@@ -20,7 +19,8 @@ from vllm.v1.worker.gpu.attn_utils import (
from vllm.v1.worker.gpu.block_table import BlockTables
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
from vllm.v1.worker.gpu.sample.gumbel import gumbel_sample
from vllm.v1.worker.gpu.spec_decode.eagle_cudagraph import EagleCudaGraphManager
from vllm.v1.worker.gpu.spec_decode.eagle.cudagraph import EagleCudaGraphManager
from vllm.v1.worker.gpu.spec_decode.eagle.utils import load_eagle_model
logger = init_logger(__name__)
@@ -73,18 +73,7 @@ class EagleSpeculator:
self.cudagraph_manager = EagleCudaGraphManager(vllm_config, device)
def load_model(self, target_model: nn.Module) -> None:
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=self.draft_model_config
)
share_lm_head = True
if share_lm_head and hasattr(target_model, "lm_head"):
if hasattr(self.model, "lm_head"):
del self.model.lm_head
self.model.lm_head = target_model.lm_head
self.model = load_eagle_model(target_model, self.vllm_config)
def set_attn(
self,
@@ -0,0 +1,52 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch.nn as nn
from vllm.config import VllmConfig
from vllm.model_executor.model_loader import get_model
def load_eagle_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Module:
from vllm.compilation.backends import set_model_tag
speculative_config = vllm_config.speculative_config
assert speculative_config is not None
draft_model_config = speculative_config.draft_model_config
with set_model_tag("eagle_head"):
eagle_model = get_model(
vllm_config=vllm_config, model_config=draft_model_config
)
# Share target embeddings when the draft checkpoint does not include
# its own vocab embedding table.
share_embeddings = True
if hasattr(eagle_model, "has_own_embed_tokens"):
share_embeddings = not eagle_model.has_own_embed_tokens
if share_embeddings:
target_language_model = (
target_model.get_language_model()
if hasattr(target_model, "get_language_model")
else target_model
)
inner_model = getattr(target_language_model, "model", None)
target_embed_tokens = None
if inner_model is not None:
if hasattr(inner_model, "embed_tokens"):
target_embed_tokens = inner_model.embed_tokens
elif hasattr(inner_model, "embedding"):
target_embed_tokens = inner_model.embedding
if target_embed_tokens is not None and hasattr(eagle_model, "model"):
if hasattr(eagle_model.model, "embed_tokens"):
del eagle_model.model.embed_tokens
eagle_model.model.embed_tokens = target_embed_tokens
# Only share target lm_head when the draft model does not own one.
share_lm_head = True
if hasattr(eagle_model, "has_own_lm_head"):
share_lm_head = not eagle_model.has_own_lm_head
if share_lm_head and hasattr(target_model, "lm_head"):
if hasattr(eagle_model, "lm_head"):
del eagle_model.lm_head
eagle_model.lm_head = target_model.lm_head
return eagle_model