Pull requests / #1148
#1148 tests: hip_prefill_wmma_gemm_parity skips on a card without matrix cores instead of failing
closed · @xjc10 · 0 comentários · No GitHub
Descrição
On gfx1030 (RX 6900 XT) `hip_prefill_wmma_gemm_parity` fails on 0.1.40.1 (`82f46a8`): the WMMA GEMM kernel declines every one of the 440 shapes, the suite counts zero passes and ends with "FAILED: zero test cases passed (all were declined or skipped)". The HIP ctest run therefore fails this one test on every RDNA2 card.
The test now checks the architecture first with the kernels' own `gfx_arch_is_gfx11_wmma` and exits 77 where there are no matrix cores; CMake marks 77 as `SKIP_RETURN_CODE`, as `hip_prompt_attn_wmma` already does. The zero-passes failure stays for cards that have WMMA.
Measured on the RX 6900 XT (gfx1030, ROCm 10.0) at `82f46a8`: before, `hip_prefill_wmma_gemm_parity` is the one failure among the 19 HIP tests ("zero test cases passed"); after, 19 tests: 13 pass, 6 skip (this one, and the five that already skip there: `hip_prefill_hcd_exact_parity`, `hip_prompt_attn_wmma`, `hip_prefill_hipblaslt_gemm`, `hip_prefill_fused_moe`, `hip_prefill_fused_iq`), 0 fail.
Developed with an AI coding assistant; every number above was measured on 2x RX 6900 XT (gfx1030, PCIe 4.0 x8 each) / Ryzen 5 5600X, ROCm 10.0.
No site
Links install, modelos, releases.