[OVEP GPU] add GQA in support list for GPU backend #830
Conversation
ankitm3k
force-pushed
the
support_GQA_GPU
branch
from
5708bda to
c4f7cdd
Compare
| "beam_idx", | ||
| "past_key_values", | ||
| "present", | ||
| "total_seq_len", |
There was a problem hiding this comment.
@Kotomi-Du Does the stateful model post translation into OVIR comprise of total_seq_len input always? Is this a general case for all LLMs now (since which OV toolkit version this was added)?
There was a problem hiding this comment.
it is the input name from Msft generic model (specifically Phisilica model), not the Epctx OVIR model OV toolkit generated
| {"Atanh", V_2020_4, {"CPU"}}, | ||
| {"Atanh", V_2022_1, {"GPU"}}, | ||
| {"Attention", V_2023_0, {"CPU", "GPU"}}, | ||
| {"GroupQueryAttention", V_2025_1, {"CPU", "GPU"}}, |
There was a problem hiding this comment.
Please add the JIRA in the PR description that enables GQA Op for CPU & GPU plugins in the ONNX OV frontend. Please make sure this change doesn't conflict with GQA support for NPU in your validation process (FYI we are not targeting to support GQA for NPU currently)
There was a problem hiding this comment.
If GQA is enabled for CPU it will also be marked for NPU. Can you revert the change
There was a problem hiding this comment.
GQA should NOT be marked as supported on CPU. NPU uses CPU's capability. As @preetha-intel mentioned, if it is marked as supported on CPU it would result in the op targeting to run on NPU and in case of compilation failure would run on OV CPU instead of MLAS which is currently a CP+ production requirement.
4 participants
GQA is originally supported by OV starting from 2025.1. This PR is to align with OV support.