[mxfp8 moe training] integrate new cuda kernel for blocked layout for groups along K

[danielvegamyhre]

Stacked PRs:

• [mxfp8 moe training] update readme with kernel microbenchmarks for dsv3 #3521
• [mxfp8 moe training] test floor vs rceil in 3d quantization benchmarks #3507
• [mxfp8 moe training] update benchmark plots to use new kernel #3506
• ->[mxfp8 moe training] integrate new cuda kernel for blocked layout for groups along K #3505
• [mxfp8 moe training] add CUDA kernel for per-group conversion of scale factors to blocked layout #3504

---

[mxfp8 moe training] integrate new cuda kernel for blocked layout for groups along K

Important note:

• Writing scales to blocked format is much more efficient when they are in row major format. Currently the CUDA kernel for mxfp8 dim1 quantization writes to column major to avoid uncoalesced writes to GMEM.
• However, I did repeated microbenchmarking on various devices and determined there is no performance regression if we write the scales to row major (see benchmarks below).
• In the quantization kernel, there are 64 threads per thread block, so we are only computing two scalar, one-byte E8M0 scale factors. So row major vs column major just means 2 transactions instead of 1 for writing the 2 bytes to GMEM. And this is negligible compared to the 64 bytes of quantized data that we'll be writing to global memory.

CUDA 2d tensor dim1 mxfp8 quantization kernel benchmarks for writing scales to col vs row major

FLOOR COL MAJOR
=== GPU 7 ===
M 16384 K 16384 BLOCK_SIZE 32
GPU: NVIDIA B200
torch version: 2.11.0.dev20251216+cu128
triton version: 3.6.0
mode: dim1_mxfp8_cuda_floor
time_us 150.4959985613823
mem_bw_gbps 5406.75488902199

FLOOR ROW MAJOR
=== GPU 7 ===
M 16384 K 16384 BLOCK_SIZE 32
GPU: NVIDIA B200
torch version: 2.11.0.dev20251216+cu128
triton version: 3.6.0
mode: dim1_mxfp8_cuda_floor
time_us 150.68800002336502
mem_bw_gbps 5399.8657880775645

RCEIL COL MAJOR
=== GPU 7 ===
M 16384 K 16384 BLOCK_SIZE 32
GPU: NVIDIA B200
torch version: 2.11.0.dev20251216+cu128
triton version: 3.6.0
mode: dim1_mxfp8_cuda_rceil
time_us 156.6080003976822
mem_bw_gbps 5195.743346021566

RCEIL ROW MAJOR
=== GPU 7 ===
M 16384 K 16384 BLOCK_SIZE 32
GPU: NVIDIA B200
torch version: 2.11.0.dev20251216+cu128
triton version: 3.6.0
mode: dim1_mxfp8_cuda_rceil
time_us 155.7759940624237
mem_bw_gbps 5223.4940364041595

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3505

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 5 New Failures

As of commit b09c4bd with merge base 7035fb7 ():

NEW FAILURES - The following jobs have failed:

• Run Regression Tests / test (CUDA 2.6, linux.g5.12xlarge.nvidia.gpu, torch==2.6.0, cuda, 12.6) / linux-job (gh)
  RuntimeError: Command docker exec -t 75998f85b471f5e19753abff779a49a6de6eb361f086deb4e994105cf33a1d1e /exec failed with exit code 2
• Run Regression Tests / test (CUDA 2.7, linux.g5.12xlarge.nvidia.gpu, torch==2.7.1, cuda, 12.6) / linux-job (gh)
  RuntimeError: Command docker exec -t c58ab338972dc92a25141eb47cc0d6a158ee92f24682145d9480da5bccd4d256 /exec failed with exit code 2
• Run Regression Tests / test (CUDA 2.8, linux.g5.12xlarge.nvidia.gpu, torch==2.8.0, cuda, 12.6) / linux-job (gh)
  RuntimeError: Command docker exec -t 32f1df797d7c1bf4a669996bece938d7c374608d6a83f418fcb5a9e7d1459b21 /exec failed with exit code 2
• Run Regression Tests / test (CUDA 2.9, linux.g5.12xlarge.nvidia.gpu, torch==2.9.1, cuda, 12.6) / linux-job (gh)
  RuntimeError: Command docker exec -t f68de68f2f4b12106fe629ec5dcd46c1c4d415c9d1dd2213f51a75268bb2d686 /exec failed with exit code 2
• Run Regression Tests / test-nightly (CUDA Nightly, linux.g5.12xlarge.nvidia.gpu, --pre torch --index-url https://downloa... / linux-job (gh)
  RuntimeError: Command docker exec -t b783952c157bac2ce51beed887c935276d04b496b7f1c9ebfbfad18531b613ba /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.