From 158c372fcad257a2d011c3cd3a710ffd82c26c8e Mon Sep 17 00:00:00 2001 From: John Travers Date: Sun, 15 Nov 2026 20:38:01 +0000 Subject: [PATCH] ci: compile the CUDA lane for AMPERE86 too; docs: API fixes need no SEP The device-compile lane now builds the cuda preset for HOPPER90 and for AMPERE86, and runs the lambda-tag check over each build. Only an AMPERE86 build compiles the device GEMM that caused #815. AGENTS.md, contributing.md and SEP 0 section 2 no longer list a public interface redesign as needing a SEP. API fixes and cleanups that change no model or physics follow the ordinary issue and pull-request path (maintainer's decision, recorded in SEP 0's log). Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01XGENPwwkUUG9KSAwneHjgH --- .github/workflows/ci.yml | 10 +++++++++- AGENTS.md | 5 +++-- docs/dev/contributing.md | 5 +++-- docs/dev/seps/sep-0000-process.md | 12 ++++++++---- docs/dev/tooling.md | 8 +++++--- native/ci/check_lambda_tags.py | 6 +++--- 6 files changed, 31 insertions(+), 15 deletions(-) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 2337d9365..9f430f0d1 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -604,9 +604,16 @@ jobs: device-compile-cuda: needs: changes if: needs.changes.outputs.lanes == 'full' && needs.changes.outputs.device == 'true' - name: device-compile (CUDA 12.8) + name: device-compile (CUDA 12.8, ${{ matrix.arch }}) runs-on: ubuntu-24.04 timeout-minutes: 150 + strategy: + fail-fast: false + matrix: + # HOPPER90 is the preset's production target. AMPERE86 (the NVIDIA A40) + # compiles kernels that HOPPER90 does not, such as the device GEMM behind + # #815, which only that build's objects can show. + arch: [HOPPER90, AMPERE86] steps: - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1 with: @@ -637,6 +644,7 @@ jobs: cmake --preset cuda \ -DCMAKE_C_COMPILER=gcc-14 -DCMAKE_CXX_COMPILER=g++-14 -DCMAKE_CUDA_HOST_COMPILER=g++-14 \ -DCUDA_ROOT=/usr/local/cuda-12.8 \ + -DSOLPROP_NATIVE_CUDA_ARCH=${{ matrix.arch }} \ -DPython_EXECUTABLE="$GITHUB_WORKSPACE/.venv/bin/python" - name: Compile, tests included working-directory: native diff --git a/AGENTS.md b/AGENTS.md index 3b5349d13..5c9ebe5aa 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -259,8 +259,9 @@ The long form is [SEP 0](docs/dev/seps/sep-0000-process.md); the index is [`docs/dev/seps/index.md`](docs/dev/seps/index.md). - **When:** new physics models, routes, geometries, mode families or numerical schemes; - changed defaults or conventions; public interface redesigns; substantive - kernel-contract changes; process changes. Not for fixes, documentation, tests, + changed physics defaults or conventions; substantive kernel-contract changes; process + changes. Not for fixes, API fixes and cleanups that change no model or physics (even + when they rename, merge or refuse public parameters), documentation, tests, refactoring, tooling or optimisations that leave results unchanged. - **Writing:** from `template.md`, standing alone (SEP 0 ยง7), with its expected cost in code, docstring and test lines; merged early as `Draft`. diff --git a/docs/dev/contributing.md b/docs/dev/contributing.md index eec8514f9..80d5d1c87 100644 --- a/docs/dev/contributing.md +++ b/docs/dev/contributing.md @@ -76,8 +76,9 @@ for an editable checkout, and a wheel has no sources beside it to compare. ## Agree and implement a change Work on a topic branch and keep changes focused. New physics, numerical schemes, -public-interface redesigns, and substantive contract changes require an accepted -[enhancement proposal](seps/sep-0000-process.md). Other changes larger than a +changed physics defaults or conventions, and substantive contract changes require an +accepted [enhancement proposal](seps/sep-0000-process.md); an API fix or cleanup that +changes no model or physics does not. Other changes larger than a focused fix start with an agreed issue or written plan. Keep the manual, docstrings, tests, and release-note fragments consistent with the diff --git a/docs/dev/seps/sep-0000-process.md b/docs/dev/seps/sep-0000-process.md index 538c33d41..0df00e9e0 100644 --- a/docs/dev/seps/sep-0000-process.md +++ b/docs/dev/seps/sep-0000-process.md @@ -51,14 +51,15 @@ A SEP is required for: - a new propagation route, geometry, spatial decomposition or mode family; - a new numerical scheme: a stepper, a transform, a quadrature, a coupled flow; - a change to a physics default or to a convention of [conventions.md](../conventions.md); -- a redesign of a public interface, or a change that breaks many users' scripts; - a substantive change to the kernel contract ([kernel-contract.md](../kernel-contract.md)) or a new backend; - a change to how solprop is developed (a *Process* SEP, like this one). -A SEP is not needed for bug fixes, documentation, tests, refactoring, tooling and -continuous integration, dependency updates, small additions to the public interface (a -keyword, a helper, a statistic), or optimisations that leave results unchanged (they +A SEP is not needed for bug fixes, API fixes and cleanups that change no model or +physics (even when they rename, merge or refuse public parameters, or change which +parameters a route takes), documentation, tests, refactoring, tooling and continuous +integration, dependency updates, small additions to the public interface (a keyword, a +helper, a statistic), or optimisations that leave results unchanged (they still need a benchmark ledger entry, [benchmarks.md](../benchmarks.md)). Those follow the working agreement of `AGENTS.md`: an issue or a short agreed plan, then a pull request. @@ -389,3 +390,6 @@ that edits this document and adds a line to its history. - 2026-11-11: The backend tiers are gone (ADR-077): a SEP is `Final` on `numpy_ref` and the native CPU build. A proposal estimates its cost in lines of code, docstrings and tests, and its phase reports record the actual numbers. +- 2026-11-15: API fixes and cleanups that change no model or physics no longer need a SEP, + even when they change public parameters; a public-interface redesign is no longer on the + list of changes that require one (maintainer's decision). diff --git a/docs/dev/tooling.md b/docs/dev/tooling.md index 70de89fb5..f10f9aac0 100644 --- a/docs/dev/tooling.md +++ b/docs/dev/tooling.md @@ -319,9 +319,11 @@ jaxtyping hook) run in the coverage lane only; the other test lanes set its lint and strict types, clang-format and its pytest suite (`lint pytest`); clang-tidy and the native tests (`tidy`); and ASan and UBSan (`sanitize`). The distribution wheel's stage is native-wheel's Linux job. -- **device-compile** (Linux, two jobs): on a change to `native/`, compiles the `cuda` - preset under CUDA 12.8 and the `hip` preset under ROCm 7.2, tests included, with warnings - as errors and no device. About 25 and 11 minutes, with no cache. A fault only a card +- **device-compile** (Linux, three jobs): on a change to `native/`, compiles the `cuda` + preset under CUDA 12.8 for `HOPPER90` and for `AMPERE86`, and the `hip` preset under + ROCm 7.2, tests included, with warnings as errors and no device. Each CUDA build then + runs `native/ci/check_lambda_tags.py` over its objects. About 20 to 28 minutes for each + CUDA job and 13 to 19 for HIP, with no cache. A fault only a card shows, or a test that fails on one, still needs a run on a device. - **gui** (Linux x86-64, macOS arm64, Windows x64): `uv run poe check-gui`, the front end's gate (its types against `gui/pyproject.toml`, its docstrings and its tests, with diff --git a/native/ci/check_lambda_tags.py b/native/ci/check_lambda_tags.py index bc7e93795..d488e583c 100644 --- a/native/ci/check_lambda_tags.py +++ b/native/ci/check_lambda_tags.py @@ -24,9 +24,9 @@ Standard library only; it needs GNU ``nm`` (binutils) and object files that a CUDA build compiled, at any optimisation level: the members exist whether or not a link happens to keep the wrong copy. It sees only the kernels the build's device target -instantiates: the CUDA device-compile lane builds for compute capability 9.0, which -compiles no ``device_dense_gemm``, so the lambda behind #815 shows only in a build for -8.6 (``AMPERE86``), checked by hand over that build's objects. +instantiates: a build for compute capability 9.0 compiles no ``device_dense_gemm``, so +the lambda behind #815 shows only in a build for 8.6 (``AMPERE86``). The CUDA +device-compile lane builds both and runs this check over each. """ from __future__ import annotations