Add gfx950 (MI355X) preload tuning table for preshuffle GEMM#411
Open
Add gfx950 (MI355X) preload tuning table for preshuffle GEMM#411
Conversation
gfx950 has different LDS/VMEM latencies compared to gfx942 (MI300X). Without tuned preload values, preshuffle GEMM regresses -14.6% E2E on DeepSeek-R1 MoE workloads. With gfx950-specific preload table, the regression is eliminated (-0.7%, within noise). Changes: - Add _TILE_PRELOAD_TABLE_GFX950 with tuned (dsrd_preload, dvmem_preload) values for common tile sizes on MI355X - Update _get_preload() to accept gpu_arch parameter and select the appropriate table - Use gfx950 table for both FP8/INT8 and BF16 paths on MI355X Tuned on Tensorwave MI355X cluster with DeepSeek-R1 MoE expert shapes (N=7168, K=18432).
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
gfx950 (MI355X) has different LDS/VMEM latencies compared to gfx942 (MI300X). Without tuned preload values, preshuffle GEMM regresses -14.6% E2E on DeepSeek-R1 MoE workloads. With gfx950-specific preload table, the regression is eliminated (-0.7%, within noise).
E2E Results (DeepSeek-R1, TP=8)
Changes
_TILE_PRELOAD_TABLE_GFX950with tuned(dsrd_preload, dvmem_preload)values for common tile sizes on MI355X_get_preload()to acceptgpu_archparameter and select the appropriate tableTesting
Tuned on Tensorwave MI355X cluster (mia1-p02-g32) with DeepSeek-R1 MoE expert shapes (N=7168, K=18432).