Repository navigation
Conversation
* Guard dim_t overflow in compute_size, reshape and grow The model loader validates shapes with a per-multiplication overflow check (since OpenNMT#2091), but the StorageView API paths still multiply dims unchecked: compute_size() and StorageView::reshape() can wrap int64_t (e.g. a shape of {2^62+1, 8} wraps to 8), yielding a view whose shape claims far more elements than _size, and grow() can wrap its addition. Apply the same per-multiplication guard used by the loader to compute_size() and reshape(), add an overflow check to grow(), and reject the overflowing cases with std::invalid_argument in two new unit tests, mirroring the OpenNMT#2094 test style. * Accept zero dimensions in the overflow guards A zero dimension yields an empty tensor, which cannot overflow, but the guards divided by it to compare with the running product: x86_64 raises SIGFPE, while AArch64 division by zero returns 0 and made every later check reject valid zero-dim shapes. Skip the division when the dimension is zero. * Check empty dimensions before multiplying the shape size A zero dimension makes the size 0 whatever its position, so the other dimensions are never multiplied and an overflowing prefix next to a zero dimension is not an error. Skipping the division is not enough: the prefix can overflow before the zero is reached, which made the result depend on where the zero appeared in the shape. The same ordering in reshape() also left the known size at 0 when the new shape combined a zero dimension with an unknown one, and that 0 is the divisor of the compatibility check below (crash on master today). The unknown dimension is now rejected in that case. * Cover the grow() guard with unit tests Split the negative increment out of the overflow condition so it gets its own message, and pin the three cases: a negative size (which used to silently shrink the dimension), an overflowing addition (tested on {INT64_MAX, 0}, where the zero dimension keeps the allocation empty), and normal growth. --------- Co-authored-by: Jordi Mas <jmas@softcatala.org>
* Fix CPU Conv1D dilation in oneDNN and GEMM paths * Keep INT8 quantization scales finite for tiny weight rows
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.
Master moved after you merged #1 (8c8cab2
Guard dim_t overflow in compute_size, reshape and growOpenNMT#2097, 0495cfaFix CPU Conv1D dilation and INT8 scale overflowOpenNMT#2117), so OpenNMT#2077 shows a conflict again.One textual conflict, in
tests/storage_view_test.cc: this branch's MPS instantiation ofStorageViewDeviceTestand master's new overflow tests, both kept.On an M4 (macOS 27), same flags as before (
-DWITH_MPS=ON): C++ suite 427 tests, 427 passed.Conv1DDilation, which was skipped before, now runs on CPU and MPS and passes after OpenNMT#2117. CI for this head runs in my fork on push.Thanks for the 8/16-slot TopK split, that was the cost I had left unmeasured.
🤖 Generated with Claude Code