[Performance, Hardware] MoE weights padding to AMD MI300x GPUs #1836
Conversation
This file contains 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
HaiShaw
requested review from
merrymercy,
Ying1123,
zhyncs and
ispobock
as code owners
merrymercy
requested changes
Oct 29, 2024
|
Please fix the CI. |
|
@merrymercy Fixed the CI just now. Thanks! |
merrymercy
requested changes
Oct 30, 2024
| @@ -572,6 +588,18 @@ def process_weights_after_loading(self, layer: Module) -> None: | |||
| start += shard_size | |||
|
|
|||
| layer.w13_scale = torch.nn.Parameter(max_w13_scales, requires_grad=False) | |||
| # If ROCm, apply weight padding (min. Mem channel contention) only if set | |||
| if is_hip() and bool(int(os.getenv("MOE_PADDING", "0"))): | |||
There was a problem hiding this comment.
move all is_hip under a single branch.
e.g., L555
There was a problem hiding this comment.
@merrymercy understand, the order of data crunching makes me intend to keep dummy padding at very last to avoid error prone situation from intervening normalize_, _dequantize. _fp8_quant, etc., and easier to read.
merrymercy
approved these changes
Oct 30, 2024
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Outdated suggestions cannot be applied.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.
Motivation
Padding MoE weights (last dim) to minimize Memory Channel Contention (only to AMD Instinct GPUs)
Test shows approximate performance boost of prefill +2.2%, decode +3.0% for Grok-1 on setting: b32/i1024/o512
Modifications
As mentioned: fused_moe.py and layer.py
To enable this feature, set binary flag
MOE_PADDING=1at command line, orexport MOE_PADDING=1in console.Checklist