Fix a performance issue with Helion-emitted Flash Attention #181
Conversation
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
facebook-github-bot
added
the
CLA Signed
jansel
requested changes
Jun 17, 2025
Comment on lines
851
to
895
| lhsSize = node.args[1].meta['val'].size() | ||
| rhsSize = node.args[2].meta['val'].size() | ||
| # check to see if it is 3D | ||
| reduceDim = False | ||
| if len(lhsSize) == 3: | ||
| env = CompileEnvironment.current() | ||
| lhsDimIdx = env.get_block_id(lhsSize[0]) | ||
| rhsDimIdx = env.get_block_id(rhsSize[0]) | ||
| if lhsDimIdx is not None and rhsDimIdx is not None: | ||
| lhsDimVal = env.block_sizes[lhsDimIdx] | ||
| rhsDimVal = env.block_sizes[rhsDimIdx] | ||
| if (lhsDimVal.from_config(ctx.cg.device_function.config) == 1 and | ||
| rhsDimVal.from_config(ctx.cg.device_function.config) == 1): | ||
| reduceDim = True | ||
|
|
||
| if not reduceDim: | ||
| return expr_from_string( | ||
| f"tl.dot(lhs, rhs, acc=acc, input_precision={tf32!r})", | ||
| lhs=lhs, | ||
| rhs=rhs, | ||
| acc=acc, | ||
| ) | ||
| # create reshape, dot, then reshape | ||
| lhs_shape_str = ctx.cg.device_function.tile_strategy.shape_str( | ||
| [*node.args[1].meta["val"].size()[1:]] | ||
| ) | ||
| rhs_shape_str = ctx.cg.device_function.tile_strategy.shape_str( | ||
| [*node.args[2].meta["val"].size()[1:]] | ||
| ) | ||
| acc_shape_str = ctx.cg.device_function.tile_strategy.shape_str( | ||
| [*node.args[0].meta["val"].size()[1:]] | ||
| ) | ||
| out_shape_str = ctx.cg.device_function.tile_strategy.shape_str( | ||
| [*node.meta['val'].size()] | ||
| ) | ||
| lhs_reshape = expr_from_string(f"tl.reshape(lhs, {lhs_shape_str})", lhs=lhs) | ||
| rhs_reshape = expr_from_string(f"tl.reshape(rhs, {rhs_shape_str})", rhs=rhs) | ||
| acc_reshape = expr_from_string(f"tl.reshape(rhs, {acc_shape_str})", rhs=acc) | ||
| comp = expr_from_string( | ||
| f"tl.dot(lhs, rhs, acc=acc, input_precision={tf32!r})", | ||
| lhs=lhs_reshape, | ||
| rhs=rhs_reshape, | ||
| acc=acc_reshape, | ||
| ) | ||
| return expr_from_string(f"tl.reshape(lhs, {out_shape_str})", lhs=comp) |
There was a problem hiding this comment.
Duplicate code with the above? Please refactor into a helper function.
jansel
reviewed
Jun 21, 2025
jansel
approved these changes
Jun 23, 2025
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
|
@manman-ren has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
oulgen
reviewed
Jun 23, 2025
| @@ -12,9 +12,9 @@ | |||
| @helion.kernel( | |||
| config=helion.Config( | |||
| # This config was autotuned on a 3090, it won't be fast for other architectures | |||
There was a problem hiding this comment.
is this comment still valid? maybe delete it?
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.
Triton doesn't handle 3D dots well in terms of performance. Perform the following operation when emitting dots from Helion when the highest dimension is 1 for 3D dots:
Perf Results on H100: