Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Remove intermediate states copying in Mllama #617

Closed
wants to merge 7 commits into from
Closed

Remove intermediate states copying in Mllama #617

Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
974f939
Assign intermediate states to pre-located tensor to avoid memcopying
jkaniecki Dec 11, 2024
007bc06
Yapf formatting
jkaniecki Dec 11, 2024
dbb35dd
Merge branch 'habana_main' into Intermediate_states_accumulation
kdamaszk Dec 13, 2024
f4f9322
Use tensor indexing to avoid possible recompilations
jkaniecki Dec 17, 2024
8252fe6
Merge branch 'HabanaAI:habana_main' into Intermediate_states_accumula…
jkaniecki Dec 17, 2024
f6dff07
Update mllama.py
jkaniecki Dec 17, 2024
b99a3fe
Yapf formatting
jkaniecki Dec 17, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 10 additions & 6 deletions vllm/model_executor/models/mllama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -509,20 +509,26 @@ def forward(
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> Union[Tuple, BaseModelOutput]:
encoder_states = ()
encoder_states = torch.empty(
(len(self.output_hidden_states), hidden_states.size(0),
hidden_states.size(1), hidden_states.size(2)),
dtype=hidden_states.dtype,
device=hidden_states.device)
hidden_states_idx = 0

for i, encoder_layer in enumerate(self.layers):
if i in self.output_hidden_states:
encoder_states = encoder_states + (hidden_states, )
encoder_states[hidden_states_idx] = hidden_states
hidden_states_idx += 1
hidden_states = encoder_layer(
hidden_states,
attention_mask,
)

if len(self.layers) - 1 in self.output_hidden_states:
encoder_states = encoder_states + (hidden_states, )
encoder_states[hidden_states_idx] = hidden_states
jkaniecki marked this conversation as resolved.
Show resolved Hide resolved

return hidden_states, encoder_states
return hidden_states, encoder_states.permute(1, 2, 3, 0)


class MllamaVisionModel(nn.Module):
Expand Down Expand Up @@ -658,8 +664,6 @@ def forward(self, pixel_values: torch.Tensor,
attention_mask=attention_mask,
)
hidden_state, intermediate_hidden_states = output[0], output[1]
intermediate_hidden_states = torch.stack(intermediate_hidden_states,
dim=-1)

# apply global encoder
hidden_state = self.layernorm_post(hidden_state)
Expand Down
Loading