Repository navigation
fix(tests): repair three mlxcel-core tests after the MLX pin bump - #644
Merged
Merged
Conversation
Follow-up to #625 (part of #623). The MLX pin bump to e9463bb and the CUDA build surfaced three mlxcel-core test failures, all test-side issues rather than product regressions. test_fused_paged_decode_gqa_and_batched exercised the Metal-only fused paged-decode kernel, which throws "No Metal back-end" and aborts the process on a CUDA build (CUDA uses the gather-then-SDPA fallback; native CUDA paged attention is #634); it now returns early on non-Metal backends via metal_is_available(). test_from_bytes_fp16_native_dtype asserted the array stays float16, but the CUDA loader widens float16 to float32 at load time (since 2026-03), so the dtype and itemsize assertions are now backend-aware (float16/2 on Metal, float32/4 on CUDA). steel_gemm_edge_tile_safe_load_matches_reference compared a JIT steel GEMM against a composite reference with a 1e-3 bound; under the new pin the JIT kernel reduces in a different tile order, giving ~0.02 max abs divergence (about 0.1% relative on magnitude-16 sums), so the fp32 reduction-order bound is relaxed to 5e-2, still tight enough to catch a gross edge-tile safe-load bug.
inureyes
added
type:chore
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Follow-up to #625 (part of #623). The pin bump to e9463bb plus the CUDA build surfaced three mlxcel-core test failures, all test-side (not product regressions). See the commit body for the per-test analysis: the Metal-only fused paged-decode test now skips on non-Metal backends, the fp16 native-dtype test is backend-aware for the CUDA float16-to-float32 load-time widening, and the steel GEMM edge-tile tolerance is relaxed from 1e-3 to 5e-2 for fp32 reduction-order differences under the new JIT kernel. All three now pass under
cargo test --release --features cuda -p mlxcel-core -- --test-threads=1; the rest of the suite is unaffected.