make fp8 blockwise linear differentiable; use new kernels

[danielvegamyhre]

Stacked PRs:

• ->make fp8 blockwise linear differentiable; use new kernels #2602
• Add Float8BlockwiseLinear with Triton kernels for quantization and GEMMs #2592

---

make fp8 blockwise linear differentiable; use new kernels

Summary

• Acknowledgement: this PR is based on the PoC by @vkuzo in [not for land] float8 blockwise scaling training prototype using deep_gemm #2386, using some different kernels for quantization and with some additional tests + benchmarks

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2602

• 📄 Preview Python docs built from this PR

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

⏳ 13 Pending, 1 Unrelated Failure

As of commit 5a4f0fa with merge base 0e00df3 ():

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• PR Label Check / Check PR Labels (gh) (detected as infra flaky with no log or failing log classifier)

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