tests : drop SYCL special-casing in test-backend-ops.cpp - #28688
Conversation
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 1, 384)); // m=12 (N=384) | ||
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 32, 384)); // m=12 (batch) | ||
| test_cases.emplace_back( | ||
| new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 4, 384, { 2, 3 })); // m=12 (multi-dim) | ||
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 768, 1, 768)); // m=12 (N=768) | ||
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 1, 640)); // m=20 (N=640) | ||
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 32, 640)); // m=20 (batch) | ||
| test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 1280, 1, 1280)); // m=20 (N=1280) |
There was a problem hiding this comment.
These UT cases are designed for SYCL backend code.
It should be kept.
How do you think?
There was a problem hiding this comment.
When a new case is added for a backend, other backends UT will be impacted.
We have met the more cases before: other backend add new UT cases which SYCL backend doesn't support.
It break the SYCL backend UT.
It will push SYCL backend to support the new cases as soon.
Why we remove them now?
There was a problem hiding this comment.
These UT cases are designed for SYCL backend code.
The test-backend-ops only has generic tests that work with all backends. We should not add tests that work only with a single backend. See #28254 (comment) for more info.
There was a problem hiding this comment.
I originally used a special-case test because the operator was only implemented for SYCL at the time. PR #28490 adds the CPU implementation, now these tests can pass CPU tests. But other GPU versions are not implemented, should I remove the tests from the PR?
Conflict resolutions: - llama-context: keep LLAMA_GRAPH_TIMING, add upstream gf_res_prev_active reuse check - llama-graph: take upstream input tensor names; self_kq_mask_swa to match - server-context: combine spec_prefill_active and stateless task checks - test-backend-ops: take upstream (SYCL special-casing dropped in ggml-org#28688) - vulkan: take upstream split, reapply coopmat shmem pad (also for the new MULMAT_QUANT spec), UMA readback gate (now in ggml-vulkan-buffers.cpp), argsort nrows and topk comment Assisted-by: Claude Opus 5.5
4 participants
Overview
ref #28254 (comment)
Partially revert the
test-backend-opschanges from #28254 as they are specific to the SYCL backend.Requirements