[irvineoy]

Several tasks accepted incorrect candidates through fixed inputs, incomplete semantic coverage, or references computed from mutated inputs. Other tasks exposed host validation code through broad edit boundaries. This change repairs 114 task packages and adds regressions for those failure modes.

Runtime dependency: every task in this PR was validated only on the ROCm 10 image lmsysorg/sglang@sha256:e20849665c105d389ef91d23c0dc73931aaa6f02056dd10e7b43e4f16c79df69 (PyTorch 2.11, Triton 3.8, FlyDSL 0.3.2). #123 has since made this image the gfx950 default on main.

Changes

• Restrict editable regions to device implementations while protecting host launch, reference, and comparison logic.
• Add independent timed input draws and replay checks where fixed inputs allowed shortcuts; preserve input snapshots and restore state around candidate calls.
• Extend correctness coverage for attention layouts and grouped heads, GDN modes, FLA/KDA geometry, convolution widths, quantization edge cases, and optional GEMM arguments.
• Enforce candidate compute boundaries and import closure, retain compiled module owners, and fix grouped-attention head indexing.
• Add a numerical GPU startup smoke check and focused CPU regressions, including compiler-portable fixtures and strict source-preservation checks.

Review fixes

• Editable decorators: src/harness_guard.py rejects host callbacks (pre_hook, post_hook, prune_configs_by, perf_model, do_bench, launch_metadata), non-lambda heuristics values, and lambdas that do more than launch-metadata arithmetic. Every shipped baseline still passes the guard.

• FlyDSL evaluator state (37 closure tasks and gemm_a8w8): candidates can no longer import the materialized timing helper _aka_benchmark, mutate imported modules (for example torch.cuda.Event.elapsed_time), read frame or function globals, reach sys through os.sys or builtins.__import__, or disable the PyTorch dispatch audit.

• FlyDSL import closure: resolution follows Python's precedence (regular package, then same-name module, then namespace directory), and reassigned sys aliases converge in linear time.

• gemm_a8w8: actual performance calls go through the candidate dispatch audit, and torch._scaled_mm is rejected.

• SSD replay: the replay keeps the timed ×32 cb amplification instead of compounding it to ×1024.

• New correctness-only controls (unscored):

  • triton_bmm: bf16/fp32, irregular dimensions, and strided/transposed/broadcast views.
  • triton_decode_attn_stage1 and triton_decode_attn_grouped_stage1: additional head widths, page sizes, group sizes, split counts and lengths.
  • triton_paged_prefix_prefill: head dims 16–128, page sizes, empty/unaligned lengths, windows.
  • chunk_local_cumsum: head_first and cu_seqlens/chunk_indices variants.
  • These task READMEs now state the supported domain derived from the protected wrappers.

• Documentation:

  • 19 Triton READMEs note that the protected wrapper fixes the launch configuration.
  • fused_clamp_act_mul documents that N need not be a power of two.
  • dsv4_fp4_indexer documents that task-local helper modules are allowed.
  • Stale grid and case-count text is corrected.

• Tests: the signal-distribution test reads originals from the base revision instead of HEAD, task import isolation restores sys.path, and the reviewed-history records cover the new bytes.

• Release notes: docs/reference/release-notes.md states that some repaired tasks change their timed input distributions or invocation preparation, so their baselines must be recomputed and earlier scores are not directly comparable.

Scored cases, tolerances, warmups, sample counts and timing methods are unchanged by the review fixes.

Some repaired tasks intentionally change timed input distributions or invocation preparation. Recompute their baselines; old and new scores are not directly comparable. No speedup or model-training gain is claimed.

Validation

GPU (MI355X/gfx950, ROCm 10 image above):

• All 114 modified task packages have source-matched, framework-finalized task_validator PASS reports.
  • The 58 packages changed by the review fixes were revalidated from scratch: 51 in the first round, and 7 more after their controls or text were added.
  • The other 56 packages are byte-identical to the originally validated sources.
  • Those 56 were validated before the stricter decorator guard. On CPU, the new guard still accepts every shipped baseline (437 tasks) and every prepared positive that the old guard accepted.
• Prepared target-language positives pass compile, correctness and performance through the framework TaskSession for every changed task that has one. These runs exercise the new FlyDSL performance-mode audit and the gemm_a8w8 audit.
  • The chunk_local_cumsum positive was extended for the new variable-length controls.
  • triton_decode_attn_grouped_stage1 uses its implemented initial candidate.

CPU (macOS):

• After rebasing onto main (Upgrade the MI355X default runtime to ROCm 10 #123 and the other merged changes; the only conflict was an append-only test file): full suite 18,563 passed. Every failure also fails on main on the same host, including one process-deadline test that fails intermittently there. make check-perf-helpers passes, and the task packages are unchanged from the validated sources.
• Before the rebase: full suite 18,309 passed. The remaining failures are Linux-only tests that also fail on main on the same host, plus one known process-timing flake. The full Linux suite has not been rerun on the final tree; CI covers it.
• tests/test_candidate_launch_and_state_boundaries.py reproduces each closed bypass. 84 of its 124 cases fail on the previous head.
• make check-perf-helpers passes.

Known limitations and follow-ups

• prepare_fn timing: it forces one graph replay per sample in gemm_a8w8 and four torch2hip tasks. Launch latency can dominate their smallest shapes.
• hip2hip/matrix_multiplication inputs: they remain exactly representable in BF16, so a reduced-precision kernel can pass. This predates this PR.
• Byte pins in tests/reviewed_contract_history.py: they still cover 45 Triton READMEs and configs, so documentation edits require regenerated records.
• Fixed launch configuration: the protected wrappers now fix the Triton launch configuration (tiles, num_warps, num_stages) for the 19 symbol-scoped tasks. This narrows the optimization space. It is accepted for this PR so that the existing GPU reports stay valid; a follow-up PR will expose validated tile, warp and stage settings for these tasks and revalidate them.
• FlyDSL tasks outside this PR: they still lack the timing-helper and evaluator-state protections added here; Harden task validation and repair across the remaining FlyDSL tasks #129 ports them.
• Structural checks: the static and dispatch checks are task-contract enforcement, not a Python sandbox.