Fully generalise input embedding handling in Transformers modelling backend (#51247)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-08-06 16:21:24 +00:00
committed by GitHub
parent 566c80edf9
commit e7b8d59460
6 changed files with 300 additions and 47 deletions
@@ -46,7 +46,6 @@ from vllm.model_executor.layers.attention import (
from vllm.model_executor.layers.attention.mla_attention import get_mla_dims
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.fused_moe import MoERunner
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
from vllm.model_executor.models.interfaces import (
SupportsEagle,
SupportsEagle3,
@@ -58,12 +57,14 @@ from vllm.model_executor.models.interfaces_base import VllmModel
from vllm.model_executor.models.transformers.fuser import BaseFuser, Fusers
from vllm.model_executor.models.transformers.fusers import MLAFuser
from vllm.model_executor.models.transformers.utils import (
attrsetter,
can_enable_torch_compile,
get_feature_request_tip,
init_on_device_without_buffers,
log_replacement,
named_state,
replace_conv_class,
replace_embedding_class,
replace_linear_class,
)
from vllm.model_executor.models.utils import (
@@ -84,17 +85,6 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
class ScaledVocabParallelEmbedding(VocabParallelEmbedding):
"""`VocabParallelEmbedding` that scales its output."""
def __init__(self, *args, embed_scale: float, **kwargs):
super().__init__(*args, **kwargs)
self.embed_scale = embed_scale
def forward(self, input_: torch.Tensor) -> torch.Tensor:
return super().forward(input_) * self.embed_scale
class Base(
nn.Module,
VllmModel,
@@ -179,24 +169,9 @@ class Base(
# Input embeddings
input_embeddings = self.model.get_input_embeddings()
if not isinstance(input_embeddings, PPMissingLayer):
names = ("embedding_size", "hidden_size")
embedding_dim = getattr_iter(self.text_config, names, None)
assert embedding_dim is not None
embedding_kwargs = dict(
num_embeddings=self.text_config.vocab_size,
embedding_dim=embedding_dim,
org_num_embeddings=self.text_config.vocab_size,
quant_config=self.quant_config,
self.model.set_input_embeddings(
replace_embedding_class(input_embeddings, self.quant_config)
)
embed_scale = getattr(input_embeddings, "embed_scale", None)
if embed_scale is not None:
# Some models scale embeddings inside the input embedding layer
new_input_embeddings = ScaledVocabParallelEmbedding(
**embedding_kwargs, embed_scale=float(embed_scale)
)
else:
new_input_embeddings = VocabParallelEmbedding(**embedding_kwargs)
self.model.set_input_embeddings(new_input_embeddings)
# Initialize any parameters that have not had their modules replaced
self.init_parameters(self.model)
@@ -380,16 +355,6 @@ class Base(
tip,
)
def attrsetter(attr: str) -> Callable[[object, object], None]:
"""Set a possibly nested attribute, like the inverse of attrgetter."""
parent, _, name = attr.rpartition(".")
def setter(obj: object, value: object):
attr_parent = attrgetter(parent)(obj) if parent else obj
setattr(attr_parent, name, value)
return setter
module_lists = []
module_list_idx = None
for i, name in enumerate(names):