fix(ggml-cuda): require nonzero graph ID for property bypass

[cheese-cakee]

Summary

Prevent a direct GGML graph from replaying an older graph's CUDA/HIP executable when its metadata arena is reused and forced property-check bypass is requested.

Direct graphs have an unset generation ID (uid = 0). The old bypass condition accepted that unset ID as a match. Rebuilding a different operation at the same first-node address could therefore replay the old operation: the regression changes NEG to SQR, but the old libraries return -3 instead of 9 for input 3.

The bypass now requires a matching, nonzero generation ID. Zero-ID graphs retain ordinary property inspection, capture, and replay. This is a correctness fix, not a throughput claim.

Changes

• Tighten the bypass predicate in ggml-cuda.cu and document the requirement in ggml-cuda.h.
• Add a model-free direct-backend regression with persistent device buffers and a reused metadata arena.
• Register the regression with CTest for CUDA/HIP graph-enabled builds, with a 60-second timeout and exit-77 handling for an unavailable GPU or disabled runtime graph support.
• Check all 16 output elements on every submission and assert the existing graph telemetry. Eager-only execution cannot count as a pass.

Verification

Base: 5c775e31dc700ed2d5fd9dabbe1cb20bd083d949 (current main when prepared).

Runtime fix: standalone hardware validation

The pre-publication validation report records the following results for the same runtime fix shipped in this PR. Adding the registered regression test and CTest wiring did not change the production predicate or invalidate these earlier hardware results.

Hardware	Before fix, forced bypass	With fix	Subsequent capture/replay
RTX 4050 Laptop, CUDA sm_89	Stale NEG: -3 instead of 9	Correct SQR: 9	Correct SQR with changed input: 400
Lucebox6 R9700, HIP gfx1201	Stale NEG: -3 instead of 9	Correct SQR: 9	Correct SQR with changed input: 400
Lucebox6 Strix Halo, HIP gfx1151	Stale NEG: -3 instead of 9	Correct SQR: 9	Correct SQR with changed input: 400

Normal property inspection also produced the correct result on all three targets. The HIP runs used ROCm 7.2.4 / HIP 7.2.53211. These are the recorded standalone-validation results, not a claim that the later CTest packaging was rerun on HIP.

Registered regression: additional CUDA validation

Fresh local validation: RTX 4050 Laptop (sm_89), WSL Ubuntu, CUDA 12.6, Release build.

Check	Result
Registered focused CTest	PASS, 1/1
Normal inspection and forced-bypass request	PASS, 192 exact output-element comparisons across 12 submissions
Rebuild fallback, recapture, and replay with changed inputs	PASS, checked through graph telemetry in both cases
Same new regression against preserved pre-fix libraries	Normal-inspection control passes; forced bypass fails with -3 instead of 9, exit 134
git diff --check	PASS

Reproduce with a graph-enabled server build:

cmake --build server/build --target test_direct_zero_uid_graph_aba -j4
ctest --test-dir server/build -R '^test_direct_zero_uid_graph_aba$' --output-on-failure --no-tests=error
# Select a different visible device when needed:
server/build/test_direct_zero_uid_graph_aba --device 1

For HIP, configure with -DDFLASH27B_GPU_BACKEND=hip -DGGML_HIP_GRAPHS=ON.

[cubic-dev-ai]

All reported issues were addressed across 4 files

_{You’re at about 90% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>