[ET-VK] Implement select_at_dim_as_symint #15617
Merged
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
## Context The SDPA custom op accepts the `input_pos` (i.e. cache position) argument as a symbolic integer. The value of the symbolic integer is obtained by selecting the first element of a cache position input tensor and converting it to symint via local_scalar_dense. Currently, ET-VK handles this in a hacky manner. 1. the select + local_scalar_dense op pattern is removed, and the cache pos tensor is passed directly into the custom sdpa ops 2. Single element tensors that have users that are all select + local_scalar_dense will be interpreted as symints instead of tensors Unfortunately, this technique will not work for the huggingface implementation of transformer models, since the cache pos input tensor has not just a single element but is expected to be a vector of integer cache positions corresponding to all cache positions that will be updated. ## Changes Introduce a custom op to capture the select + local_scalar_dense op pattern, which is the proper way to handle the op pattern. Note that a custom op is needed because this op needs to access the staging buffer data of the input tensor, whereas `select` would typically be executed via a compute shader. The reason for this is because the `input_pos` value is needed to configure the sizes of attention weight tensors participating in the custom SDPA op, so the value must be set before any command buffers are dispatched. As a consequence of this change, the previous handling of select + local scalar dense can also be removed. Differential Revision: [D86340340](https://our.internmc.facebook.com/intern/diff/D86340340/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15617
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 Cancelled Job, 5 Unrelated FailuresAs of commit 2f9378a with merge base 2b02316 ( CANCELLED JOB - The following job was cancelled. Please retry:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
This was referenced Nov 5, 2025
Merged
This PR needs a
|
manuelcandales
approved these changes
Nov 5, 2025
## Context The SDPA custom op accepts the `input_pos` (i.e. cache position) argument as a symbolic integer. The value of the symbolic integer is obtained by selecting the first element of a cache position input tensor and converting it to symint via local_scalar_dense. Currently, ET-VK handles this in a hacky manner. 1. the select + local_scalar_dense op pattern is removed, and the cache pos tensor is passed directly into the custom sdpa ops 2. Single element tensors that have users that are all select + local_scalar_dense will be interpreted as symints instead of tensors Unfortunately, this technique will not work for the huggingface implementation of transformer models, since the cache pos input tensor has not just a single element but is expected to be a vector of integer cache positions corresponding to all cache positions that will be updated. ## Changes Introduce a custom op to capture the select + local_scalar_dense op pattern, which is the proper way to handle the op pattern. Note that a custom op is needed because this op needs to access the staging buffer data of the input tensor, whereas `select` would typically be executed via a compute shader. The reason for this is because the `input_pos` value is needed to configure the sizes of attention weight tensors participating in the custom SDPA op, so the value must be set before any command buffers are dispatched. As a consequence of this change, the previous handling of select + local scalar dense can also be removed. Differential Revision: [D86340340](https://our.internmc.facebook.com/intern/diff/D86340340/) [ghstack-poisoned]
meta-codesync
bot
merged commit 0726751
into
gh/SS-JIA/353/base
154 of 162 checks passed
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.
Stack from ghstack (oldest at bottom):
Context
The SDPA custom op accepts the
input_pos(i.e. cache position) argument as a symbolic integer. The value of the symbolic integer is obtained by selecting the first element of a cache position input tensor and converting it to symint via local_scalar_dense.Currently, ET-VK handles this in a hacky manner.
Unfortunately, this technique will not work for the huggingface implementation of transformer models, since the cache pos input tensor has not just a single element but is expected to be a vector of integer cache positions corresponding to all cache positions that will be updated.
Changes
Introduce a custom op to capture the select + local_scalar_dense op pattern, which is the proper way to handle the op pattern.
Note that a custom op is needed because this op needs to access the staging buffer data of the input tensor, whereas
selectwould typically be executed via a compute shader. The reason for this is because theinput_posvalue is needed to configure the sizes of attention weight tensors participating in the custom SDPA op, so the value must be set before any command buffers are dispatched.As a consequence of this change, the previous handling of select + local scalar dense can also be removed.
Differential Revision: D86340340