[Bug] Fix reverse-mode ndarray gradients on Vulkan by binding the grad buffer directly.#801
Closed
duburcqa wants to merge 2 commits into
Closed
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 160f30c446
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…d buffer directly. Reverse-mode autodiff over an `ndarray` argument produced all-zero gradients on the Vulkan/MoltenVK backend: the forward pass and the gradient value were computed correctly, but the accumulate never reached the host `.grad` buffer. Root cause: on backends with `spirv_has_physical_storage_buffer` the ndarray data and gradient pointers are passed as buffer_device_address `u64`s embedded in the args struct, and MoltenVK misreads the SECOND such pointer member (the gradient slot) - it resolves to a garbage address, so the reverse accumulate no-ops. The data slot (first pointer) reads fine, and quadrants' own Metal backend reads the identical SPIR-V correctly; only MoltenVK's SPIRV-Cross mishandles it. This is independent of the member's offset or alignment (verified: a 16-aligned grad slot still fails) and of the buffer type (uniform vs storage). Route the gradient ndarray through the bound-storage-buffer descriptor path instead of the physical-address path - the same path already used on backends without physical_storage_buffer, and the reason field gradients already work on Vulkan. The data pointer keeps the physical-address path. This sidesteps the MoltenVK struct-member defect with no change to the args layout or the data path, and no regression on Metal/CUDA/CPU. With this fix every backend supports ndarray reverse-mode autodiff, so the ndarray autodiff tests no longer need the `archs_support_ndarray_ad` arch restriction: drop the list and let the tests run on the full expected-arch matrix (the tests fail on Vulkan without this fix).
duburcqa
force-pushed
the
duburcqa/fix_vulkan_ndarray_grad
branch
from
July 20, 2026 19:21
160f30c to
e9ba11a
Compare
…nd skip tests accordingly.
duburcqa
force-pushed
the
duburcqa/fix_vulkan_ndarray_grad
branch
from
July 20, 2026 20:21
5c3237d to
6d2021a
Compare
Contributor
Author
|
Closing: the bound-buffer route this PR takes is capped by Metal's 31 buffer binding slots and cannot compile a real reverse-mode kernel (Genesis's rigid backward already overflows it). Root cause and both rejected attempts documented in #802. A real fix requires MoltenVK to read the gradient device address correctly on the physical-storage-buffer path. |
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
1 participant
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.
Issue: #
Stacked on #800 -> #799 -> #796 -> #795. Review/merge those first; this PR is based on #800 (which adds Metal to the ndarray-autodiff test matrix) and additionally enables Vulkan once the gradient-buffer bug below is fixed. Base retargets down the stack as each lands.
Brief Summary
copilot:summary
Walkthrough
copilot:walkthrough