[JAX] triangle attention#122
Merged
Merged
Conversation
mariogeiger
commented
Jun 25, 2025
mariogeiger
commented
Collaborator
- VJP rule
- Call cuda code only if we are running on a GPU
mariogeiger
changed the title
mariogeiger
changed the title
mariogeiger
changed the title
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR introduces a new triangle_attention operator in the JAX backend, complete with custom VJP support and optional CUDA acceleration, and updates documentation and tests accordingly.
- Added the
triangle_attentionimplementation with a custom forward/backward rule and GPU fallback. - Updated the API reference to include a new Triangle section for
triangle_attention. - Introduced unit tests for gradient correctness and basic functionality on CPU and CUDA.
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| docs/api/cuequivariance_jax.rst | Added a Triangle section in the API docs and autogeneration entry. |
| cuequivariance_jax/tests/triangle_attention_test.py | New tests covering gradient checks and output shapes for triangle_attention. |
| cuequivariance_jax/cuequivariance_jax/triangle/_triangle_attention.py | Implemented triangle_attention with custom VJP, JAX reference, and CUDA fallback. |
| cuequivariance_jax/cuequivariance_jax/triangle/init.py | Added license header in new triangle package directory. |
| cuequivariance_jax/cuequivariance_jax/init.py | Exported triangle_attention in the root module’s public API. |
Comments suppressed due to low confidence (2)
cuequivariance_jax/cuequivariance_jax/triangle/_triangle_attention.py:42
- The return type annotation '-> jax.Array' does not reflect that this function returns a tuple of three arrays; consider annotating it as '-> tuple[jax.Array, jax.Array, jax.Array]' for clarity.
) -> jax.Array: # [B, N, H, S_qo, D]
cuequivariance_jax/cuequivariance_jax/triangle/_triangle_attention.py:69
- [nitpick] The variable 'a' is ambiguous; consider renaming it to 'attn_logits' or 'attention_scores' to clarify its purpose.
a = jnp.einsum("...ai,...bi->...ab", q, k, precision=precision)
hsadasivan
reviewed
Jul 9, 2025
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. |
Collaborator
There was a problem hiding this comment.
Empty file. Do we need this?
I see imports are already taken care of in cuequivariance_jax/cuequivariance_jax/init.py
hsadasivan
reviewed
Jul 9, 2025
| def triangle_attention_fwd(q, k, v, mask, bias, scale, precision=None): | ||
| a, lse, amax = fwd_p.bind(q, k, v, mask, bias, scale=scale, precision=precision) | ||
| residuals = (a, lse, q, k, v, mask, bias) | ||
| return (a, lse, amax), residuals |
Collaborator
There was a problem hiding this comment.
why are we returning residuals? Is it a JAX requirement?
hsadasivan
reviewed
Jul 9, 2025
| da, dlse, damax = cotangents | ||
|
|
||
| dq, dk, dv, dbias = bwd_p.bind( | ||
| da, a, lse, q, k, v, mask, bias, scale=scale, precision=precision |
Collaborator
There was a problem hiding this comment.
For consistency with torch, isn't it better to do bias,mask rather than mask,bias ?
Collaborator
Author
There was a problem hiding this comment.
good catch, done
hsadasivan
reviewed
Jul 9, 2025
| def fn(q, k, v, mask, bias): | ||
| return cuex.triangle_attention(q, k, v, mask, bias, scale) | ||
|
|
||
| output, lse, amax = fn(q, k, v, mask, bias) |
Collaborator
There was a problem hiding this comment.
why dont we have more elaborate parametrized tests for triattn in jax like in trimul (shapes, precision etc.)
borisfom
approved these changes
Jul 16, 2025
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
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.