Enable Eagle3 speculative decoding for GPT-OSS model (#25246)
Signed-off-by: Eldar Kurtic <8884008+eldarkurtic@users.noreply.github.com>
This commit is contained in:
@@ -27,7 +27,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.utils import cdiv
|
||||
|
||||
from .interfaces import SupportsPP
|
||||
from .interfaces import SupportsEagle3, SupportsPP
|
||||
from .utils import (AutoWeightsLoader, WeightsMapper, extract_layer_index,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory, make_layers,
|
||||
@@ -238,6 +238,7 @@ class GptOssModel(nn.Module):
|
||||
self.make_empty_intermediate_tensors = (
|
||||
make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], self.config.hidden_size))
|
||||
self.aux_hidden_state_layers = tuple[int, ...]()
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.embedding(input_ids)
|
||||
@@ -261,8 +262,12 @@ class GptOssModel(nn.Module):
|
||||
x = intermediate_tensors["hidden_states"]
|
||||
residual = intermediate_tensors["residual"]
|
||||
|
||||
aux_hidden_states = []
|
||||
for i in range(self.start_layer, self.end_layer):
|
||||
layer = self.layers[i]
|
||||
if i in self.aux_hidden_state_layers:
|
||||
aux_hidden_states.append(x if residual is None else x +
|
||||
residual)
|
||||
x, residual = layer(x, positions, residual)
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors({
|
||||
@@ -270,6 +275,9 @@ class GptOssModel(nn.Module):
|
||||
"residual": residual
|
||||
})
|
||||
x, _ = self.norm(x, residual)
|
||||
|
||||
if len(aux_hidden_states) > 0:
|
||||
return x, aux_hidden_states
|
||||
return x
|
||||
|
||||
def _load_weights_mxfp4(
|
||||
@@ -610,7 +618,7 @@ class GptOssModel(nn.Module):
|
||||
weights, stacked_params_mapping)
|
||||
|
||||
|
||||
class GptOssForCausalLM(nn.Module, SupportsPP):
|
||||
class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
|
||||
packed_modules_mapping = {"qkv": ["q_proj", "k_proj", "v_proj"]}
|
||||
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
@@ -658,6 +666,13 @@ class GptOssForCausalLM(nn.Module, SupportsPP):
|
||||
self.make_empty_intermediate_tensors = (
|
||||
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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.get_input_embeddings(input_ids)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user