JiaLuo-CAN 5ff7497fa7 [rocm-libraries] ROCm/rocm-libraries#7537 (commit 07123f4) ...

[CK Tile] Fix Grouped Gemm quant mixed precision (#7537)

<Migrate from Internal repo PR>
test_ck_tile_grouped_gemm_quant_tensor would fail for mixed FP8/BF8
cases:
std::tuple<Row, Col, Row, FP8, F32, BF8, F32, F32, F16, TensorQuant,
False, True, False>,
std::tuple<Row, Col, Row, BF8, F32, FP8, F32, F32, F16, TensorQuant,
False, True, False>

GFX1250 would fail with incorrect results, GFX950 would fail when
compiling BF8+FP8 and give incorrect results for FP8+BF8.
The issue is due to the wrong ComputeDataType selection.
The fix is to consider original ADataType and BDataType even when
ComputeDataType is not void. For compiling error on gfx950, the bf8,
fp8, 16x16x32 warp Gemm is added.

2026-05-21 08:36:23 -07:00

..

ck

[rocm-libraries] ROCm/rocm-libraries#5652 (commit 7dc7d1d)

2026-05-18 17:46:01 +02:00

ck_tile

[rocm-libraries] ROCm/rocm-libraries#7537 (commit 07123f4)

2026-05-21 08:36:23 -07:00

rapidjson

Update pre-commit to fixed versions, run remod for ck_tile (#2895)

2025-10-16 15:29:17 -07:00