Repository navigation
fix(cuda-oxide): use valid sparse FP8 metadata - #1487
Merged
Merged
Conversation
Signed-off-by: nihalp <nihalp@nvidia.com>
nihalpasham
added
the
cuda-oxide
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
2 tasks done
nihalpasham
added a commit
that referenced
this pull request
Oct 11, 2026
Update CUDAoxide from `nightly-2026-08-28` to `nightly-2026-10-11`. This brings in rust-lang/rust#163242, needed for the separate kernel parameter-attribute work in #1483 (related to #1088 and #1098). ## Changes - Adapt to rustc API changes, generic floating-point intrinsics, and integer min/max. - Follow the new `core::intrinsics::reflection::type_id` path in the kernel-name helper. - Update toolchain/Nix pins, diagnostic snapshots, and intrinsic evidence. Intrinsic declarations are unchanged. ## Validation - CI: 55 passed, 7 expected skips; all 234 examples compile. - Compiler tests, formatting, clippy, and IR comparison pass. Targeted checks pass on both RTX 5090s with no memcheck errors. - GPU sweep: 232/233 passed. The existing FP8 example failure is addressed separately in #1487; hardware-limited cases retain compile-only coverage. The existing scaffold issue—mixing registry and Git runtime types—also remains a separate follow-up. ## Checklist - [x] All commits signed off - [x] SPDX headers on new source files, where applicable --------- Signed-off-by: nihalp <nihalp@nvidia.com>
nihalpasham
marked this pull request as ready for review
October 11, 2026 11:58
nihalpasham
added
bug
2 tasks done
Signed-off-by: nihalp <nihalp@nvidia.com>
nihalpasham
requested review from
binarybana,
cryos,
elibol,
kkraus14,
roivanov and
rparolin
as code owners
October 11, 2026 13:23
nihalpasham
added a commit
that referenced
this pull request
Oct 11, 2026
Use rustc's `noalias` and proven `dereferenceable(N)` facts on `cuda-oxide` kernel references. This lets LLVM reuse input reads across writes to separate outputs. Raw pointers and `DisjointSlice` storage receive no new attributes. Rust's interior-mutability, pinning, and lifetime rules still apply. Explicit `readonly` is deferred because current NVPTX backends can lose volatile reads. Includes regression tests, corrected atomic example pointer contracts, and load-width checks that accept cached loads. Validation: compiler/backend suites and strict checks pass; reference launches pass memory, race, and synchronization checks on both RTX 5090s. LLVM and both NVVM paths preserve two volatile reads while combining ordinary reads. Final hosted CI is pending. The existing FP8 failure is handled by #1487. The existing AtomicF16 sanitizer failure remains a separate follow-up. Partially addresses #1088 and #1098. ## Checklist - [x] All commits signed off - [x] SPDX headers on new source files --------- Signed-off-by: nihalp <nihalp@nvidia.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
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.
Fix
sparse_mma_fp8, which tested twelve metadata encodings even though plain FP8mma.sppermits only six. The invalid encodings caused 5,104 mismatches on RTX 5090.The example and layout probe now use the six valid encodings. Inline comments and generated safety docs explain the restriction and link to the PTX specification. Instruction lowering and ABI are unchanged.
Also set
rust-analyzer.cargo.featuresto[]so VS Code stops enabling every optional feature.Validation
nightly-2026-10-11after cuda-oxide: update to nightly-2026-10-11 #1482; generated outputs and all 1,029 ABI entries pass.Follow-up to #1295. Running examples on GPUs in CI remains separate work.