[V1] Refactor model executable interface for all text-only language models (#10374)

Signed-off-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
Roger Wang
2024-11-17 05:18:46 +00:00
committed by GitHub
parent 4fd9375028
commit 643ecf7b11
43 changed files with 483 additions and 90 deletions
+11 -8
View File
@@ -324,11 +324,8 @@ class Phi3SmallModel(nn.Module):
make_empty_intermediate_tensors_factory(["hidden_states"],
config.hidden_size))
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
def forward(
self,
@@ -337,9 +334,13 @@ class Phi3SmallModel(nn.Module):
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors],
inputs_embeds: Optional[torch.Tensor],
) -> Union[torch.Tensor, IntermediateTensors]:
if get_pp_group().is_first_rank:
hidden_states = self.embed_tokens(input_ids)
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.get_input_embeddings(input_ids)
if (self.mup_embedding_multiplier is not None
and self.mup_embedding_multiplier > 0.0):
hidden_states = hidden_states * self.mup_embedding_multiplier
@@ -397,8 +398,8 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
else:
self.dummy_token_indices = None
def get_input_embeddings(self):
return self.model.embed_tokens
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
def set_input_embeddings(self, value):
self.model.embed_tokens = value
@@ -433,6 +434,7 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
output_hidden_states = self.model(
input_ids=input_ids,
@@ -440,6 +442,7 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
kv_caches=kv_caches,
attn_metadata=attn_metadata,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
)
output_hidden_states = output_hidden_states
return output_hidden_states