Hunt correctness for extend attention + RPE #504
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: Stanley Winata <[email protected]>
Signed-off-by: Stanley Winata <[email protected]>
Avoid flex_attention for RPE as it is unclear there is a correct implementation possible given the limitations of the create_block_mask re conditionals. Instead we use the manual torch implementation that is known to be correct. As we update the test and the extend_attention_rpe to use a static max_rpe_context_length, a new error appears that suggests some issue with the indexing in extend_attention_rpe. Repro: ``` pytest tests/kernel/wave/attention/extend_attention_test.py --run-e2e -v -k "rpe" ``` Errors out with: ``` E Diagnostics: E <stdin>:282:18: error: 'vector.gather' op operand iree-org#2 must be vector of integer or index values, but got 'index' E %468 = "vector.gather"(%109, %39, %39, %467, %44) : (memref<?xf32, strided<[1], offset: ?>>, index, index, vector<4xi1>, vector<4xf32>) -> vector<4xf32> E ^ ``` Signed-off-by: Nicolas Vasilache <[email protected]>
Signed-off-by: Alex Zinenko <[email protected]>
This reverts commit 1a12817.
nicolasvasilache
force-pushed
the
users/stan/extendRPE
branch
from
2b97ac1 to
c08692a
Compare
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Outdated suggestions cannot be applied.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.
Avoid flex_attention for RPE as it is unclear there is a correct implementation possible given the
limitations of the create_block_mask re conditionals.
Instead we use the manual torch implementation that is known to be correct.
As we update the test and the extend_attention_rpe to use a static max_rpe_context_length,
a new error appears that suggests some issue with the indexing in extend_attention_rpe.
Repro:
Errors out with:
Signed-off-by: Nicolas Vasilache [email protected]