Skip to content

Release notes: 2026-05-28 — gpu-decorator-mir

Status: Ready for review
Repo: li-langverse/lic
PR: branch cursor/gpu-decorator-mir-5b3a
PH / REQ: PH-7d, G-gpu, T-decorators-gpu-mir
Author: agent

Summary (one sentence)

@gpu and @gpu(devices=N) now survive MIR lowering and appear in lic verify telemetry, establishing compiler-visible Li GPU placement on def declarations before LKIR/vendor codegen.

Agent continuation (required)

  1. Read: compiler/mir/include/li/mir.hpp, compiler/mir/lower.cpp, compiler/mir/mir.cpp, compiler/lic/main.cpp, docs/game-dev/specs/lig-rfc.md, and docs/verification/provability-gaps.md (G-gpu).
  2. Run: cmake --build build --target lic -j 4, ./scripts/check-mir-gpu-decorator.sh, and LI_REPO_ROOT=$PWD ./li-tests/run_all.sh decorators.
  3. Then: connect MIR @gpu placement to a proof-carrying LKIR/catalog launch path in lig; add fail tests before any device-buffer/address-space lowering.
  4. Blocked on: address-space proofs, LKIR lowering, device-buffer contracts, and backend emitters for CUDA/HIP/Metal/SPIR-V are still open G-gpu work.

Changed (specific)

Area What Evidence
MIR MirDecorator now carries gpu and gpu_devices; copy_decorators() records @gpu and @gpu(devices=N). lic verify li-tests/decorators/gpu_multi_device_ok.li reports mir_gpu_def=1 mir_gpu_multi_device_def=1.
Verify telemetry count_mir_gpu_def, count_mir_gpu_multi_device_def, and lic verify output expose GPU placement counts. ./scripts/check-mir-gpu-decorator.sh exit 0.
Policy @gpu(devices=0), non-integer devices, and vendor/backend arguments are rejected rather than accepted as impossible or vendor-specific placement. lic check --format=json li-tests/decorators/gpu_devices_zero_fail.li returns E0322 with gpu_devices; decorators suite compile-fail rows cover gpu_devices_ident_fail.li and gpu_vendor_arg_fail.li.
Tests Added gpu_only_ok.li, gpu_multi_device_ok.li, gpu_devices_zero_fail.li, gpu_devices_ident_fail.li, gpu_vendor_arg_fail.li, and manifest rows. LI_REPO_ROOT=$PWD ./li-tests/run_all.sh decorators exit 0.
Docs Updated execution decorators, lig RFC, and provability gaps to mark G-gpu as partial. docs/superpowers/specs/2026-05-16-li-execution-decorators.md, docs/game-dev/specs/lig-rfc.md, docs/verification/provability-gaps.md.

Not changed (scope fence)

  • No CUDA, HIP/ROCm, Metal, Vulkan/wgpu, SPIR-V, or LKIR codegen was added.
  • No vendor driver calls, new trusted axioms, device allocation API, or device memory copy path was added.
  • No performance numbers or benchmark thresholds changed; lig still benchmarks backend/runtime paths separately from compiler placement metadata.
  • No multi-GPU scheduler, stream overlap, peer-to-peer copy, or distributed runtime was implemented.

Breaking changes

N/A — @gpu(devices=0) was previously inert/meaningless and is now rejected; valid @gpu source remains accepted.

Security

No new trusted runtime surface was added. Security improvement is static rejection of impossible GPU placement (devices=0), non-literal device counts, and source-level vendor strings, plus explicit telemetry so future agents can gate GPU lowering before vendor/runtime calls are introduced.

Performance

N/A — no GPU execution path or performance claim changed. This PR only adds MIR metadata and verification telemetry.

Downstream

Repo Action
lig Use mir_gpu_def / mir_gpu_multi_device_def as the compiler-side handoff for future LKIR launch work.
benchmarks N/A — no benchmark rows changed; future lig perf PRs should update lig-kernels.toml.
lip / lit / lis N/A — no package manager, test runner, or server API changed.

CHANGELOG entry (paste into Unreleased)

### Added
- **PH-7d / G-gpu decorator telemetry:** `@gpu` and `@gpu(devices=N)` now survive MIR lowering and `lic verify` reports `mir_gpu_def` / `mir_gpu_multi_device_def`; invalid device counts and source-level vendor args are rejected before LKIR/codegen — [2026-05-28-gpu-decorator-mir.md](docs/release-notes/2026-05-28-gpu-decorator-mir.md).