=== therock#5314 NT/NN/TN layout sweep on gfx1151 (Ubuntu 24.04, iommu=pt, ROCm7.13), 2026-06-08T06:38:35-06:00 ===
### TORCH_BLAS_PREFER_HIPBLASLT=0 (rocBLAS path)
HIPBLASLT= 0 | torch 2.11.0+rocm7.13.0a20260506
  [NT] fwd  y=x@W.T
      2048^3  fp16  36.1  bf16  36.0 TFLOPS
      4096^3  fp16  21.8  bf16  21.8 TFLOPS
      8192^3  fp16  24.9  bf16  25.2 TFLOPS
  [NN] dgrad dX=dY@W
      2048^3  fp16  27.5  bf16  31.0 TFLOPS
      4096^3  fp16  32.8  bf16  29.2 TFLOPS
      8192^3  fp16  23.7  bf16  24.2 TFLOPS
  [TN] wgrad dW=dY.T@x
      2048^3  fp16  35.6  bf16  34.7 TFLOPS
      4096^3  fp16  39.9  bf16  40.5 TFLOPS
      8192^3  fp16  30.9  bf16  31.2 TFLOPS
### TORCH_BLAS_PREFER_HIPBLASLT=1 (hipBLASLt path)
HIPBLASLT= 1 | torch 2.11.0+rocm7.13.0a20260506
  [NT] fwd  y=x@W.T
      2048^3  fp16  36.9  bf16  35.5 TFLOPS
      4096^3  fp16  32.4  bf16  34.0 TFLOPS
      8192^3  fp16  34.5  bf16  34.1 TFLOPS
  [NN] dgrad dX=dY@W
      2048^3  fp16  34.3  bf16  35.1 TFLOPS
      4096^3  fp16  35.5  bf16  35.2 TFLOPS
      8192^3  fp16  30.6  bf16  31.0 TFLOPS
  [TN] wgrad dW=dY.T@x
      2048^3  fp16  23.6  bf16  23.8 TFLOPS
      4096^3  fp16  16.3  bf16  16.3 TFLOPS
      8192^3  fp16   8.4  bf16   8.4 TFLOPS
