Commit 8fc633a
committed
[GPU] GQA i8/i4: address review - kernel unit tests + readable static cases
Review feedback on the quantized-KV GroupQueryAttention tests:
- Add GPU kernel unit tests for the newly enabled i8/u8 data types (ref:
scatter_update_gpu_test.cpp / slice_gpu_test.cpp):
* scatter_update_gpu_{i8,u8}.d8111_axisB - templated int scatter (byte movement).
* slice_gpu_i8.bfyx - dedicated small-value i8 case (the shared TYPED_TEST data
exceeds the i8 range, so a separate case rather than extending DataTypes).
- Rewrite the *_static GQA test cases to declare named query/past_key/past_value/
expected_* variables with tensor-shape comments (matching the dynamic cases'
style) instead of positional literals passed to the helper, and note that the
static cases complement the existing dynamic (Slice+Concat) onnx_model_gqa_*
cases.
All green: scatter i8/u8 + slice i8 unit tests pass; GQA quant suite 27/27
(i8/i4 x per_channel/per_tensor x dynamic/static across IE_GPU/IE_CPU/INTERPRETER).1 parent 8b1c424 commit 8fc633a
4 files changed
Lines changed: 291 additions & 161 deletions
File tree
- src
- frontends/onnx/tests
- plugins/intel_gpu
- src/graph/impls/ocl
- tests/unit/test_cases
0 commit comments