[Refactor] Consolidate SupportsEagle (#36063)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
This commit is contained in:
@@ -60,7 +60,13 @@ from vllm.model_executor.model_loader.weight_utils import (
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.v1.attention.backend import AttentionType
|
||||
|
||||
from .interfaces import SupportsLoRA, SupportsPP
|
||||
from .interfaces import (
|
||||
EagleModelMixin,
|
||||
SupportsEagle,
|
||||
SupportsEagle3,
|
||||
SupportsLoRA,
|
||||
SupportsPP,
|
||||
)
|
||||
from .utils import (
|
||||
AutoWeightsLoader,
|
||||
PPMissingLayer,
|
||||
@@ -313,7 +319,7 @@ class ApertusDecoderLayer(nn.Module):
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class ApertusModel(nn.Module):
|
||||
class ApertusModel(nn.Module, EagleModelMixin):
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
@@ -357,8 +363,6 @@ class ApertusModel(nn.Module):
|
||||
else:
|
||||
self.norm = PPMissingLayer()
|
||||
|
||||
self.aux_hidden_state_layers = tuple[int, ...]()
|
||||
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
@@ -384,13 +388,14 @@ class ApertusModel(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)
|
||||
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(
|
||||
@@ -472,7 +477,9 @@ class ApertusModel(nn.Module):
|
||||
return loaded_params
|
||||
|
||||
|
||||
class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
class ApertusForCausalLM(
|
||||
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
|
||||
):
|
||||
packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}
|
||||
|
||||
# LoRA specific attributes
|
||||
@@ -520,13 +527,6 @@ class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
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 _init_model(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
|
||||
Reference in New Issue
Block a user