fix(dl): route compression-off through ggml's own knob (3.1.1)

[lloyal-research]

Supersedes the 3.1.1 attempt, whose fix was silently ineffective: it passed --compress-mode=none as a raw base CUDA flag, but ggml-cuda appends its own -compress-mode=${GGML_CUDA_COMPRESSION_MODE} (ggml/src/ggml-cuda/CMakeLists.txt:210, knob default "size" at ggml/CMakeLists.txt:211) after base flags — and nvcc takes the last value. The v3.1.1 build emitted nvcc warning : incompatible redefinition for option 'compress-mode', the last value of this option was used, meaning the pack would have shipped driver-incompatible again. The run was cancelled before anything published or uploaded (npm 3.1.1 does not exist; the 3.1.1 binaries slot is empty; the tag points at a run that produced nothing).

This also sharpens the root cause one final step: the compression was never nvcc's implicit default — ggml explicitly opts into size compression via its knob. The knob accepts none, which is the setting this flavor needs: NVIDIA documents the new compress modes as incompatible with pre-CUDA-12.4 drivers (r550), and old-driver fleets (Cloud Run L4 = 535/12.2) are the DL pack's primary audience — proven fatal on the v3.1.0 rig run ("device kernel image is invalid" on every kernel load).

Stays on 3.1.1: the tag produced zero artifacts (cancelled pre-publish), so the ref is recreated on this merge commit instead of burning the number. Carries the version + optionalDependencies sync + lockfile reconcile. After merge: tag v3.1.2 → pipeline → the L4 gate proves the fix (a correct build shows no compress-mode redefinition warning) → publish-dl signs.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes the CUDA fatbin compression setting for the BACKEND_DL build flavor by routing “compression off” through ggml’s dedicated CMake knob (so it can’t be overridden by ggml-cuda’s later-appended nvcc flags), and bumps the release version to 3.1.2 with dependency metadata kept in sync.

Changes:

• Switch BACKEND_DL builds from using base CMAKE_CUDA_FLAGS to GGML_CUDA_COMPRESSION_MODE=none.
• Bump package version from 3.1.1 → 3.1.2.
• Sync optionalDependencies and package-lock.json to 3.1.2.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated no comments.

File	Description
scripts/build.js	Uses ggml’s compression-mode knob for BACKEND_DL to reliably disable fatbin compression.
package.json	Bumps root package version to 3.1.2 and updates optional dependency versions to match.
package-lock.json	Reconciles lockfile package version and optional dependency versions to 3.1.2.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.