biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
Roger Wang
2024-11-16 21:18:46 -08:00
committed by GitHub
parent 4fd9375028
commit 643ecf7b11
43 changed files with 483 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/jais.py
View File

@@ -250,6 +250,9 @@ class JAISModel(nn.Module):
make_empty_intermediate_tensors_factory(["hidden_states"],
config.n_embd))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.wte(input_ids)
def forward(
self,
input_ids: torch.Tensor,
@@ -257,9 +260,11 @@ class JAISModel(nn.Module):
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[IntermediateTensors, torch.Tensor]:
if get_pp_group().is_first_rank:
inputs_embeds = self.wte(input_ids)
if inputs_embeds is None:
inputs_embeds = self.get_input_embeddings(input_ids)
if self.wpe is not None:
position_embeds = self.wpe(position_ids)
hidden_states = inputs_embeds + position_embeds
@@ -311,6 +316,9 @@ class JAISLMHeadModel(nn.Module, SupportsPP):
self.make_empty_intermediate_tensors = (
self.transformer.make_empty_intermediate_tensors)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.transformer.get_input_embeddings(input_ids)
def forward(
self,
input_ids: torch.Tensor,
@@ -318,9 +326,11 @@ class JAISLMHeadModel(nn.Module, SupportsPP):
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[IntermediateTensors, torch.Tensor]:
hidden_states = self.transformer(input_ids, positions, kv_caches,
attn_metadata, intermediate_tensors)
attn_metadata, intermediate_tensors,
inputs_embeds)
return hidden_states
def compute_logits(