[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
+15 -17
View File
@@ -63,7 +63,13 @@ from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .interfaces import (
EagleModelMixin,
SupportsEagle,
SupportsEagle3,
SupportsLoRA,
SupportsPP,
)
from .utils import (
AutoWeightsLoader,
is_pp_missing_parameter,
@@ -391,7 +397,7 @@ class MiniCPMDecoderLayer(nn.Module):
@support_torch_compile
class MiniCPMModel(nn.Module):
class MiniCPMModel(nn.Module, EagleModelMixin):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -413,8 +419,6 @@ class MiniCPMModel(nn.Module):
self._init_layers(prefix, config, cache_config, quant_config)
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.aux_hidden_state_layers = tuple[int, ...]()
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], self.config.hidden_size
)
@@ -455,19 +459,18 @@ class MiniCPMModel(nn.Module):
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
aux_hidden_states = []
aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual)
for idx, layer in enumerate(
islice(self.layers, self.start_layer, self.end_layer)
):
if idx in self.aux_hidden_state_layers:
aux_hidden_states.append(
hidden_states + residual if residual is not None else hidden_states
)
hidden_states, residual = layer(
positions,
hidden_states,
residual,
)
self._maybe_add_hidden_state(
aux_hidden_states, idx + 1, hidden_states, residual
)
if not get_pp_group().is_last_rank:
return IntermediateTensors(
@@ -550,7 +553,9 @@ class MiniCPMModel(nn.Module):
return loaded_params
class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
class MiniCPMForCausalLM(
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -611,13 +616,6 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.embed_input_ids(input_ids)
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 forward(
self,
input_ids: torch.Tensor | None,