Pull requests / #3423

#3423 [Other] Set the policy baseline for the standalone GGML overlay test

open · @hulkbig · 0 comments · View on GitHub

AMD / HIPNVIDIA / CUDALinux

Description

Fixes #3412.

The documented `cmake -P cmake/ggml_tq_overlay_test.cmake` invocation fails on CMake 3.x because its script scope has no policy baseline. In particular, the `IN_LIST` check uses the OLD behavior of CMP0057 even though the project requires CMake 3.24 or newer. Running the main CMake project first does not initialize policies for a separate script invocation.

Set `cmake_minimum_required(VERSION 3.24)` in the standalone test, matching the project baseline. The overlay generator and GPU source are unchanged.

Validation (Linux x86_64, CPU-only script execution):
- Used Strata source at `9ae93f7b1fe989d04a8e404d32cddaeeff10505f` and a complete llama.cpp checkout at the pinned `3cf03257f219afbe7334045ff7c6a06ac68c627d`.
- Before the change, the unmodified overlay test fails with CMP0057 / Unknown arguments on CMake 3.24.3 and 3.31.10; it passes on 4.4.4.
- After the change, the same standalone test passes on all three versions.
- All 121 generated/copied files are byte-identical across the three fixed runs and the pre-change CMake 4.4.4 run.
- Command: `cmake -DLLAMA=<pinned llama.cpp checkout> -DOUT=<dedicated scratch directory> -P cmake/ggml_tq_overlay_test.cmake`.

Not run: CUDA/HIP/SYCL builds, GPU inference, model benchmarks, or online CI. This change only initializes the standalone test's CMake policies and makes no performance claim.

AI-assisted implementation and validation.

Related on this site

Quick links to related pages.