build: cache native compilation and incremental image builds

[DuncanBetts]

Upstream image rebuilds discard the entire CMake/Ninja build tree when source files change. This PR retains that tree and adds ccache: editing one C++ file and one CUDA file then runs seven affected build steps instead of all 278. Native builds can enable the same compiler cache through standard CMake launchers.

• Document optional native CMake compiler launchers, incremental rebuilds, cache configuration, statistics, and how to disable the launchers again. The ordinary native build needs neither ccache nor container tooling.
• Retain the CMake/Ninja tree in a locked build cache mount. Synchronize sources by content, remove deleted files, and use a different build directory when installed toolchain/library versions change. Copy completed binaries out of the cache for the runtime image.
• Install ccache and rsync only in the build stage. Keep ccache in a separate mount with a 20 GiB limit, so compiler outputs can also be reused when Ninja needs to compile again.
• Use the existing Docker workflow. No Compose file, Podman-specific workflow, extra build stage, or Python dependency is introduced.

No project-specific CMake option is needed: CMake already supports compiler launchers for all three languages. The runtime image and inference behavior are unchanged.

Rebuild comparison: upstream versus this PR

Both recordings follow the same workload: build the image, add the same harmless static_assert to src/artifact/typed_binding.cpp and src/ops/launcher/residual_add.cu, rebuild, then restore the sources. The baseline is upstream without this PR; the after result is the final implementation with persistent Ninja and compiler caches.

Rebuild after the two-file edit	Upstream without this PR	With this PR
Image rebuild wall time	127.750 s	20.090 s
Ninja steps	278	7
Compiler invocations	263	2

The recorded rebuild is approximately 6.4× faster. These are single-run measurements recorded separately, not a contemporaneous controlled benchmark. The baseline is the original upstream recording; it has not been rerun alongside the final revision. The times cover image rebuilding, not native-build or inference performance.

The after run started with empty, isolated build and compiler caches. Its initial build made 263 compiler calls; the edited rebuild made only two more. Ninja skipped all 261 unchanged compiler inputs, so the cumulative ccache statistics are 265 calls, zero hits, and 265 misses. CUDA device linking still took about 13.4 seconds on the rebuild.

• Upstream recording, without this PR
• Final recording with this PR, including five chapter markers

Validation

Using Ubuntu 24.04, CMake 3.28.3, CUDA 13.1.115, and ccache 4.9.1:

• The revised Dockerfile built both product binaries and the runtime image successfully (278 Ninja steps; all 263 compiler calls cacheable).
• Restoring the two source files with timestamps from 2000 correctly rebuilt both objects (two ccache hits).
• Changing only non-code build context reported ninja: no work to do, with no new compiler-cache calls.
• After clearing all three launchers and moving ccache out of PATH in a disposable build container, the C++ artifact-binding and CUDA residual-add objects compiled successfully.
• Both runtime binaries passed --help with NVIDIA driver libraries exposed; the runtime image contains neither ccache nor rsync.
• git diff --check passed.

Validation used Podman's Dockerfile support because Docker and a native host CMake/CUDA toolchain are not installed here. Docker BuildKit and a bare-host build were not executed; the native instructions use standard CMake settings and introduce no container dependency.

References

• CMake compiler launchers
• ccache manual
• Docker build cache mounts

The original PR said "with careful steering and review by myself". This updated version I have not reviewed properly yet, so I am reverting its status to "draft" until I do.

Changes made using AI. Thanks for creating Ninfer Neroued, I hope you find these build time savings useful in future work, you can take it as a token of my appreciation :).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e64041e07a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".