Tuned H100/H200 triton fp8 block configs for fused_qkv_a_proj (#23939)

Signed-off-by: mgoin <mgoin64@gmail.com>
2025-08-29 13:28:35 -04:00
parent 4d7fe40fc0
commit b7adf94c4a
4 changed files with 271 additions and 3 deletions
--- a/benchmarks/kernels/bench_block_fp8_gemm.py
+++ b/benchmarks/kernels/bench_block_fp8_gemm.py
@@ -16,6 +16,7 @@ assert current_platform.is_cuda(), (
 # DeepSeek-V3 weight shapes
 DEEPSEEK_V3_SHAPES = [
    (512 + 64, 7168),
+    (2112, 7168),
    ((128 + 64) * 128, 7168),
    (128 * (128 + 128), 512),
    (7168, 16384),