working mxfp8 quantization cpp extension

[lessw2020]

on blackwell.

=== Numerical Accuracy Testing ===
Input shape: torch.Size([4, 32])
Output shape: torch.Size([4, 32])
Scales shape: torch.Size([4, 1])
Row 0: max abs value = 4.000
Row 1: max abs value = 3.500
Row 2: max abs value = 4.000
Row 3: max abs value = 3.500

=== Performance Testing ===

Size: 2048x2048
Average time: 0.040 ms
Throughput: 422.94 GB/s

Size: 4096x4096
Average time: 0.106 ms
Throughput: 634.14 GB/s

Size: 8192x8192
Average time: 0.360 ms
Throughput: 746.15 GB/s

=== All tests completed ===

[danielvegamyhre]

Confirmed cpp extension build and tests run w/o issue on my B200 devgpu:

(ao) [danvm@devgpu031.atn1 ~/private-torchao/torchao/experimental/mxfp8_cpp (lessw2020/cpp_mxfp8_main)]$ python test_cpp_extension.py 
GPU: NVIDIA B200
Compute Capability: 10.0
=== Testing Basic MXFP8 Quantization ===

1. Row-wise quantization only:
Output (row-wise):
  Shape: torch.Size([512, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 262144
Scales (row-wise):
  Shape: torch.Size([512, 16])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 8192

2. Column-wise quantization only:
Output (column-wise):
  Shape: torch.Size([512, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 262144
Scales (column-wise):
  Shape: torch.Size([16, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 8192

3. Both row-wise and column-wise quantization:
Output (row-wise):
  Shape: torch.Size([512, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 262144
Output (column-wise):
  Shape: torch.Size([512, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 262144
Scales (row-wise):
  Shape: torch.Size([512, 16])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 8192
Scales (column-wise):
  Shape: torch.Size([16, 512])
  Dtype: torch.uint8
  Device: cuda:0
  Size in bytes: 8192

=== Testing Different Block Sizes ===

1x32 blocks:
Scales shape for 1x32 blocks: torch.Size([128, 4])

32x1 blocks:
Scales shape for 32x1 blocks: torch.Size([4, 128])

=== Numerical Accuracy Testing ===
Input shape: torch.Size([4, 32])
Output shape: torch.Size([4, 32])
Scales shape: torch.Size([4, 1])
Row 0: max abs value = 4.000
Row 1: max abs value = 3.500
Row 2: max abs value = 4.000
Row 3: max abs value = 3.500

=== Performance Testing ===

Size: 256x256
Average time: 0.021 ms
Throughput: 12.71 GB/s

Size: 512x512
Average time: 0.021 ms
Throughput: 50.84 GB/s

Size: 1024x1024
Average time: 0.021 ms
Throughput: 203.38 GB/s

Size: 2048x2048
Average time: 0.025 ms
Throughput: 677.47 GB/s

Size: 4096x4096
Average time: 0.088 ms
Throughput: 764.40 GB/s

Size: 8192x8192
Average time: 0.320 ms
Throughput: 838.93 GB/s

=== All tests completed ===