[Refactor] Consolidate SupportsEagle (#36063)

Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
This commit is contained in:
Benjamin Chislett
2026-03-13 23:22:40 +00:00
committed by GitHub
parent 54a6db827f
commit 8b346309a5
24 changed files with 229 additions and 235 deletions
+4 -9
View File
@@ -48,7 +48,7 @@ from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import set_default_rope_theta
from vllm.v1.attention.backend import AttentionType
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .interfaces import SupportsEagle, SupportsEagle3, SupportsLoRA, SupportsPP
from .qwen2 import Qwen2MLP as Qwen3MLP
from .qwen2 import Qwen2Model
from .utils import AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix
@@ -258,7 +258,9 @@ class Qwen3Model(Qwen2Model):
)
class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
class Qwen3ForCausalLM(
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -307,13 +309,6 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
self.model.make_empty_intermediate_tensors
)
def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
self.model.aux_hidden_state_layers = layers
def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
num_layers = len(self.model.layers)
return (2, num_layers // 2, num_layers - 3)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.embed_input_ids(input_ids)