-
Notifications
You must be signed in to change notification settings - Fork 2.4k
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.
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
Fix support of "inputs_embeds" input of VLMs in SDPAToPA transformation #28674
base: master
Are you sure you want to change the base?
Conversation
// for "input_ids" | ||
processed_input_ids = input_ids_node; | ||
} else { | ||
OPENVINO_ASSERT(processed_input_ids, "Counln't process neither input_ids, nor inputs_embeds."); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
OPENVINO_ASSERT(processed_input_ids, "Counln't process neither input_ids, nor inputs_embeds."); | |
OPENVINO_ASSERT(processed_input_ids, "Couldn't process neither input_ids, nor inputs_embeds."); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Fixed
for (const auto& target : input_ids_target_inputs) { | ||
target.replace_source_output(processed_input_ids); | ||
} | ||
} else if (input_ids_node->get_friendly_name() == "inputs_embeds") { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
could you add a new unit test for this scenario?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, of course, I'm working on it rn
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Created a test for the "inputs_embeds" input based on the nanoLLaVA pattern
VLMs have tokens and their generation already served as the "inputs_embeds" input, so there's no need to transform the input in the way this is done for "input_ids".
Tickets:
Signed-off-by: Andrii Staikov andrii.staikov@intel.com