Open Godlovecui opened 2 months ago
To support SM89's FP8 GEMM in CUTE, you will need to warp its PTX as shown here and implement the corresponding trait as demonstrated here.
The PTX has already been warped in CUTLASS, which serves as a good reference. Given that its layout is similar toSM80_16x8x32_S32S8S8S32_TN, you can replicate the trait of SM80_16x8x32_S32S8S8S32_TN, with the exception of the ValType.
which copy operation and copy traits should I follow? When I use:
using g2s_copy_op = SM80_CP_ASYNC_CACHEGLOBAL<
you just have to add the MMA atoms right? that should not require anything else. Its really hard to help you if you are not posting your entire diff. Can you please make an MR so that we can see the code?
This is my github commit, gemm-multi-stage_e4m3_v2.cu is my test code for e4m3. I add MMA atoms in 3rd/cutlass/include/cute/arch/mma_sm80.hpp, https://github.com/Godlovecui/cute-gemm/commit/c0168852a3a62e76dd0de33e05b410510aa02837#diff-2e927406d89afc8be5b4c85da736860c23e0eac8e871bbe8c0ebbe7ebf67fe7c Thank you very much~ @thakkarV
This diff is nearly 100k lines large. Please make a small self contained MR
This is my new commit, which only contain my modified code, Thank you for your help~ @thakkarV
https://github.com/Godlovecui/cute-gemm/commit/d7fdb5dc0910fc0a788f466195886233c473331b
gemm-multi-stage_e4m3_v2.cu is my modified code for support FP8. The step to reproduction:
I left a review on your code
I left a review on your code
ok, I try and modify it according to your review, thank you very much. @thakkarV
This issue has been labeled inactive-30d
due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d
if there is no activity in the next 60 days.
Hi: I'd like to test FP8 in RTX 4090. I can find some BF16 functions like SM80_16x8x8_F32BF16BF16F32_TN in cutlass/include/cute/arch/mma_sm80.hpp, however, I can't find some FP8 functions like SM80_16x8x8_F32E4M3E4M3FP32_TN. So, how to implement SM80_16x8x8_F32E4M3E4M3FP32_TN? Thank you~