Support causal masking in FlashAttention #1947
Draft
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
Signed-off-by: Julian Oppermann <[email protected]>
jopperm
requested review from
whitneywhtsang,
etiotto,
Dewei-Wang-sh and
a team
etiotto
reviewed
Aug 26, 2024
This was referenced Aug 27, 2024
whitneywhtsang
pushed a commit
that referenced
this pull request
Aug 30, 2024
Add support for distributing `tt.make_range` ops according to the desired warp size. This is a PoC which assumes that multiple warps are only needed along one dimension. See #1947 for more context / the complete PoC. --------- Signed-off-by: Julian Oppermann <[email protected]>
whitneywhtsang
pushed a commit
that referenced
this pull request
Sep 4, 2024
…#2013) Detect and handle flash attention with causal masking in `-convert-triton-to-tritongpu-warp` by supporting `tt.make_range` and *two* dependent attention-`for`-loops. See #1947 for more context / the complete PoC. --------- Signed-off-by: Julian Oppermann <[email protected]>
whitneywhtsang
pushed a commit
that referenced
this pull request
Sep 4, 2024
…2043) Splits `make_range` into SG-sized subranges, and handles row-vector broadcasts (e.g. `1x64 -> 16x64`) in `MatchTargetSize`. See #1947 for more context / the complete PoC. --------- Signed-off-by: Julian Oppermann <[email protected]>
Dewei-Wang-sh
pushed a commit
that referenced
this pull request
Sep 9, 2024
Support canonicalization of dependent `scf.for` loops by re-gluing individual results after the loop. See #1947 for more context / the complete PoC. --------- Signed-off-by: Julian Oppermann <[email protected]>
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Outdated suggestions cannot be applied.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.
PoC extension of the advance path to handle causal masking in FlashAttention-2.
Summary of changes:
tt.make_rangethroughout all passes of the advance path.TritonToTritonGPUWarpandMatchTargetSizeto support two dependent attention-for-loops.tt.broadcastof row vectors: Here, we need to select and splat a single value per thread. To that end, I needed to introduce an op for querying the lane ID in TritonGEN.The generated code passes result verification against PyTorch (both
causal=Trueandcausal=False).Remaining issues: