Conversation
… fp8 type with HIP
Binyang2014
reviewed
Oct 13, 2025
Binyang2014
requested changes
Oct 13, 2025
Binyang2014
reviewed
Oct 14, 2025
Binyang2014
reviewed
Oct 14, 2025
Binyang2014
reviewed
Oct 14, 2025
Contributor
Binyang2014
left a comment
There was a problem hiding this comment.
Pls refer this for AMD fp8: https://github.com/ROCm/rccl/blob/e738c03e39f681b5f38409788c781a7ffd3cc73e/src/device/reduce_kernel.h#L440-L443
… fp8 elements; skip special FP8 cal_vectors_helper
…_fp8x2_e5m2 and __fp8x4_e5m2
…p8_e4m3 and __fp8_e5m2. Fix the issue of data validation for __fp8_e5m2 with HIP. Add missing handling of sizeof(T) == 1 for FP8 in allreduceAllPairs and allreduce7.
Binyang2014
reviewed
Oct 23, 2025
Contributor
|
We need to put our reduction function to a unified place. We can change this in another PR |
|
There was an error handling pipeline event 79c0b7b6-6c02-4298-8105-6a2b4742057f. |
…CUDA_ARCH__ < 1000
chhwang
approved these changes
Oct 25, 2025
Binyang2014
approved these changes
Oct 25, 2025
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add FP8 support for Allreduce on both NVIDIA and AMD platform.
Add new data type: fp8_e4m3 and fp8_e5m2
Allreduce performance
1. NVIDIA H100
Nccl-tests with MSCCLPP
# out-of-place in-place
# size count type redop root time algbw busbw #wrong time algbw busbw #wrong
# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
MSCCLPP:
134217728 67108864 half sum -1 496.6 270.25 472.94 0 499.2 268.85 470.49 0
134217728 67108864 bfloat16 sum -1 497.2 269.94 472.39 0 497.4 269.83 472.20 0
NCCL:
(Using MSCCLPP_FORCE_NCCL_FALLBACK_OPERATION="allreduce")
134217728 134217728 f8e4m3 sum -1 735.0 182.60 319.55 0 709.0 189.32 331.31 0
134217728 134217728 f8e5m2 sum -1 734.6 182.70 319.73 0 709.2 189.26 331.21 0
2. AMD MI300
Rccl-tests with MSCCLPP
# out-of-place in-place
# size count type redop root time algbw busbw #wrong time algbw busbw #wrong
# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
MSCCLPP:
134217728 134217728 fp8_e4m3 sum -1 763.0 175.91 307.83 0 761.9 176.16 308.28 0
134217728 134217728 fp8_e5m2 sum -1 762.7 175.97 307.94 0 762.2 176.09 308.15 0
134217728 67108864 half sum -1 752.6 178.33 312.09 0 752.9 178.27 311.97 0
RCCL:
(Using MSCCLPP_FORCE_NCCL_FALLBACK_OPERATION="allreduce")
134217728 134217728 fp8_e4m3 sum -1 838.3 160.11 280.20 0 839.3 159.92 279.86 0
134217728 134217728 fp8_e5m2 sum -1 837.4 160.28 280.49 0 838.9 159.99 279.98 0