Add benchmarking and testing for blockwise fp8 GEMM using Triton and … #2740
Conversation
…torch._scaled_mm - Introduced a new benchmarking script to compare performance between Triton kernels and torch._scaled_mm for blockwise fp8 GEMM operations. - Added tests for correctness and performance of scaled_mm implementations, including various matrix sizes commonly used in LLMs. - Implemented scaled_mm kernels to support blockwise fp8 GEMM operations, preserving scaling precision. - Enhanced the Float8BlockwiseLinear class to support both Triton and scaled_mm backends. - Included error handling and edge case tests for the new implementations.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2740
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
Hi @romanywu! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
meta-cla
bot
added
the
CLA Signed
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
There was a problem hiding this comment.
Thanks for taking a crack at this! I left a couple comments, there are some serious issues with this approach I think, we can't use other scaling types and other GEMMs as an approximation, we need to perform these benchmarks against torch._scaled_mm + CUDA 12.9 that dispatches to the actual cutlass kernel in PT core for groupwise+blockwise scaling.
| a_fp8, a_scale, b_fp8, b_scale | ||
| ) | ||
|
|
||
| # For now, use tensorwise scaling as a baseline comparison |
There was a problem hiding this comment.
I'm confused, why are we comparing against tensorwise here? Tensorwise will have substantially different perf than blockwise, we need to do a 1:1 comparison here
| Blockwise fp8 GEMM using torch._scaled_mm instead of Triton kernel. | ||
|
|
||
| This is a simplified implementation that approximates blockwise scaling | ||
| using row-wise and column-wise scaling supported by torch._scaled_mm. |
There was a problem hiding this comment.
torch._scaled_mm supports DSV3 style groupwise and blockwise scaling now, using CUDA 12.9. We need to use that, we can't approximate
issue: #2713