jagadish-amd

diff --git a/benchmarks/float8/bench_matmul.py b/benchmarks/float8/bench_matmul.py
index 01d4577b4..e0f058a6a 100644
--- a/benchmarks/float8/bench_matmul.py
+++ b/benchmarks/float8/bench_matmul.py
@@ -138,13 +138,9 @@ def run(
                 scale_a = to_blocked(scale_a)
                 scale_b = to_blocked(scale_b)
         elif recipe == "mxfp4_cutlass":
-            # Use the blockwise scales from to_mx
-            if is_ROCM():

jagadish-amd

diff --git a/aten/src/ATen/BlasBackend.h b/aten/src/ATen/BlasBackend.h
index 03b00cc2156..ebcb591eddc 100644
--- a/aten/src/ATen/BlasBackend.h
+++ b/aten/src/ATen/BlasBackend.h
@@ -37,6 +37,8 @@ enum class ScalingType : std::uint8_t {
   BlockWise1x32, // fp8_e8m0fnu scales
   BlockWise1x128, // fp32 scales
   BlockWise128x128, // fp32 scales
+  // ROCm gfx950: hipBLASLt Block_32_UE8M0_32_8_EXT pre-swizzled E8M0 scales (see HIPBLASLT_MATMUL_MATRIX_SCALE_BLK32_UE8M0_32_8_EXT)
+  BlockWiseBlk32Ue8m0_32_8_EXT,

jagadish-amd

    fast_accum  name      M      K      N  ref_pct_top_peak  pct_top_peak  ref_time_s    time_s  fp8_speedup
0         True     0   1024   1024   1024          0.015986      0.010914    0.000058  0.000043     1.365439
1         True     1   1536   1536   1536          0.047957      0.018315    0.000066  0.000086     0.763805
2         True     2   2048   2048   2048          0.075174      0.036426    0.000099  0.000103     0.969102
3         True     3   3072   3072   3072          0.098859      0.073937    0.000255  0.000170     1.495804
4         True     4   4096   4096   4096          0.127724      0.104420    0.000468  0.000286     1.635096
5         True     5   6144   6144   6144          0.118588      0.087502    0.001701  0.001152     1.475736
6         True     6   8192   8192   8192          0.187609      0.071746    0.002548  0.003332     0.764848
7         True     7  12288  12288  12288          0.153743      0.056482    0.010494  0.014283     0.734754
8         True     8  16384  16384  16384