Skip to content

[OpenVINO] Add cache_position input inside prepare_inputs method for Mamba #1517

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

Merged
IlyasMoutawwakil merged 3 commits into from
Nov 13, 2025
Merged

[OpenVINO] Add cache_position input inside prepare_inputs method for Mamba #1517

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
2c7de13
Add cache_position for Mamba prepare_inputs
nikita-savelyevv Nov 11, 2025
c44e641
Override prepare_inputs method for OVModelWithMambaForCausalLM
nikita-savelyevv Nov 12, 2025
ab805dc
Update optimum/intel/openvino/modeling_decoder.py
nikita-savelyevv Nov 12, 2025
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
24 changes: 21 additions & 3 deletions optimum/intel/openvino/modeling_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1295,15 +1295,20 @@ def _has_cache_inputs(model: openvino.Model) -> bool:
"past_key_values" in key.get_any_name() or "cache_params" in key.get_any_name() for key in model.inputs
)

def forward(
def prepare_inputs(
self,
input_ids: Optional[torch.LongTensor] = None,
input_ids: torch.LongTensor,
attention_mask: Optional[torch.LongTensor] = None,
cache_params=None,
use_cache: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
**kwargs,
):
) -> Dict:
if kwargs.get("past_key_values") is not None:
raise ValueError("`past_key_values` input is not supported for `OVModelWithMambaForCausalLM`")
if kwargs.get("position_ids") is not None:
raise ValueError("`position_ids` input is not supported for `OVModelWithMambaForCausalLM`")
Comment on lines +1307 to +1310
Copy link
Member

@IlyasMoutawwakil IlyasMoutawwakil Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this really needed ?

Copy link
Collaborator Author

@nikita-savelyevv nikita-savelyevv Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My motivation was to avoid possible confusion if anyone passes these arguments because parent's definition of prepare_inputs() has these arguments in its signature.

Copy link
Member

@IlyasMoutawwakil IlyasMoutawwakil Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

hmm i'm honestly not sure why this function is public, imo it should be private (it serves one purpose, an internal one)

Copy link
Collaborator Author

@nikita-savelyevv nikita-savelyevv Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is used externally during quantization https://github.com/huggingface/optimum-intel/blob/main/optimum/intel/openvino/quantization.py#L675

Copy link
Member

@IlyasMoutawwakil IlyasMoutawwakil Nov 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

that's still internal though 😅.. what i mean is that it shouldn't be used by a user
the downside of a public method is that we need to maintain its behaviour and only change it through deprecation process (over multiple versions).


inputs = {"input_ids": input_ids}
if "cache_position" in self.input_names:
if cache_position is None:
Expand Down Expand Up @@ -1340,6 +1345,19 @@ def forward(
batch_size = input_ids.shape[0]
inputs["beam_idx"] = np.arange(batch_size, dtype=int)

return inputs

def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
cache_params=None,
use_cache: Optional[bool] = None,
cache_position: Optional[torch.Tensor] = None,
**kwargs,
):
inputs = self.prepare_inputs(input_ids, attention_mask, cache_params, use_cache, cache_position, **kwargs)

self.request.start_async(inputs, share_inputs=True)
self.request.wait()
logits = torch.from_numpy(self.request.get_tensor("logits").data).to(self.device)
Expand Down
Loading