[mxfp8 moe training] update roofline and bench script to use new cuda kernel

[danielvegamyhre]

Stacked PRs:

• [mxfp8 moe training] readme updates: update table of contents, organization; remove outdated content #3567
• [mxfp8 moe training] add kernel microbenchmarks table to readme #3566
• ->[mxfp8 moe training] update roofline and bench script to use new cuda kernel #3565

---

[mxfp8 moe training] update roofline and bench script to use new cuda kernel

• Remove to_mx from benchmark plots (these are for RCEIL, and we aren't using it for RCEIL due to poor perf Use cvt.rp.satfinite.ue8m0x2.f32 PTX instruction in Inductor codegen for mxfp8 quantization pytorch#170635)
• Add new wgrad_with_hp arg added to _to_mxfp8_then_scaled_grouped_mm (set to false).
• Update gemm_kernel_choice keyword arg to kernel_preference (to match refactor).
• Replace Triton kernel with new CUDA kernel added in [mxfp8 moe training] add CUDA kernel for per group blocked layout with groups along M #3546 for per-group scale factor blocked layout transform with groups along M

Plot update preview

Before (see green bar for "per group along M conversion to blocked layout" kernel is disproportionally large):

After (see green bar is super tiny now):

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3565

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• B200 runners are down due to network issues

✅ No Failures

As of commit 32bccb2 with merge base 8bb433e ():
💚 Looks good so far! There are no failures yet. 💚

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