Fix embed scaling + CUDA graphs in Transformers modelling backend (#48010)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -78,6 +78,17 @@ if TYPE_CHECKING:
|
|||||||
logger = init_logger(__name__)
|
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(
|
class Base(
|
||||||
nn.Module,
|
nn.Module,
|
||||||
VllmModel,
|
VllmModel,
|
||||||
@@ -156,22 +167,26 @@ class Base(
|
|||||||
self.attention_instances = self.create_attention_instances()
|
self.attention_instances = self.create_attention_instances()
|
||||||
|
|
||||||
# Input embeddings
|
# Input embeddings
|
||||||
self.embed_scale = None
|
|
||||||
input_embeddings = self.model.get_input_embeddings()
|
input_embeddings = self.model.get_input_embeddings()
|
||||||
if not isinstance(input_embeddings, PPMissingLayer):
|
if not isinstance(input_embeddings, PPMissingLayer):
|
||||||
# Some models scale embeddings inside the input embedding layer
|
|
||||||
self.embed_scale = getattr(input_embeddings, "embed_scale", None)
|
|
||||||
names = ("embedding_size", "hidden_size")
|
names = ("embedding_size", "hidden_size")
|
||||||
embedding_dim = getattr_iter(self.text_config, names, None)
|
embedding_dim = getattr_iter(self.text_config, names, None)
|
||||||
assert embedding_dim is not None
|
assert embedding_dim is not None
|
||||||
self.model.set_input_embeddings(
|
embedding_kwargs = dict(
|
||||||
VocabParallelEmbedding(
|
num_embeddings=self.text_config.vocab_size,
|
||||||
self.text_config.vocab_size,
|
embedding_dim=embedding_dim,
|
||||||
embedding_dim=embedding_dim,
|
org_num_embeddings=self.text_config.vocab_size,
|
||||||
org_num_embeddings=self.text_config.vocab_size,
|
quant_config=self.quant_config,
|
||||||
quant_config=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
|
# Initialize any parameters that have not had their modules replaced
|
||||||
self.init_parameters(self.model)
|
self.init_parameters(self.model)
|
||||||
@@ -590,10 +605,7 @@ class Base(
|
|||||||
_init_parameters(module, dtype)
|
_init_parameters(module, dtype)
|
||||||
|
|
||||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||||
inputs_embeds = self.model.get_input_embeddings()(input_ids)
|
return self.model.get_input_embeddings()(input_ids)
|
||||||
if self.embed_scale is not None:
|
|
||||||
inputs_embeds *= self.embed_scale
|
|
||||||
return inputs_embeds
|
|
||||||
|
|
||||||
def forward(
|
def forward(
|
||||||
self,
|
self,
|
||||||
@@ -608,16 +620,6 @@ class Base(
|
|||||||
input_ids = None
|
input_ids = None
|
||||||
inputs_embeds = intermediate_tensors["hidden_states"]
|
inputs_embeds = intermediate_tensors["hidden_states"]
|
||||||
|
|
||||||
# If the model scales embeddings inside the input embedding layer we must
|
|
||||||
# ensure they are scaled here since VocabParallelEmbedding will not do it
|
|
||||||
if (
|
|
||||||
self.embed_scale is not None
|
|
||||||
and input_ids is not None
|
|
||||||
and inputs_embeds is None
|
|
||||||
):
|
|
||||||
inputs_embeds = self.embed_input_ids(input_ids)
|
|
||||||
input_ids = None
|
|
||||||
|
|
||||||
# Add batch dimension before entering Transformers model
|
# Add batch dimension before entering Transformers model
|
||||||
if input_ids is not None and input_ids.ndim == 1:
|
if input_ids is not None and input_ids.ndim == 1:
|
||||||
# [seq_len] -> [1, seq_len]
|
# [seq_len] -> [1, seq_len]
|
||||||
|
|||||||
Reference in New Issue
Block a user