Repository navigation
docs: accept SEP 48, a Julia execution backend - #609
Conversation
An optional backend="julia" at Tier 3: one Julia engine for CPUs and NVIDIA, AMD and Apple GPUs, in binary64 and binary32 (SEP 33), owning SEP 47's whole execution session. pip install solprop is unchanged; the julia extra adds only juliacall, and `solprop julia setup` installs a pinned Julia and committed environments, refused until run. Includes the merge scope, the Julia code rules and tools (Runic, Aqua, JET), CI, verification, phases and open questions Q1-Q7. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Solprop-managed environments with committed manifests; a per-run SimulationSettings.device, with GPU packages loaded on demand through package extensions from stacked environments of one joint resolution, which native checks against its build; the Julia CI job on every non-docs PR; `solprop julia setup`/`status` over Python functions; Linux x86-64 and macOS arm64 qualified; the full section 5.3 scope at merge; Julia 1.13 pinned by patch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7428570932
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - **Adaptive runs** take the same accepted steps, with the same control trace on equal | ||
| readings (SEP 47 §4.3). | ||
| - **Fields and every statistic** agree within bounds derived from the operations' | ||
| summation counts ([testing.md](../testing.md)). The experiment's windows agreed to | ||
| $10^{-13}$ or better. |
There was a problem hiding this comment.
Separate adaptive accuracy checks from identical-input control tests
For independently adaptive Julia and NumPy runs, rounding differences in reductions can legitimately change accepted steps and orders, so requiring identical accepted steps and summation-based field bounds would reject correct implementations. SEP 47 §4.3 explicitly permits different accepted step counts and requires identical control decisions only when fed identical recorded readings. Specify per-case accuracy envelopes against a tighter or converged reference for adaptive runs, retain rounding budgets for matched discrete histories, and test identical-input control traces separately. This also follows the numerical-comparison requirement in AGENTS.md.
AGENTS.md reference: AGENTS.md:L343-L346
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in ab619ef.
- §4 now defines three comparison classes:
- M, matched discrete histories, with SEP 25 §§4.3–4.4 rounding budgets;
- A, independent adaptive runs, against a converged reference within an
rtol-set envelope, each backend counting its own work exactly; - C, control identity on recorded readings, per SEP 47 §4.3.
- §7.1 assigns every comparison to one of these classes, including groups, repeat runs and each kind of resume.
- The requirement for equal accepted steps in adaptive runs is removed.
The GPT-6 review raised the same point as R1. Part III records the response.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1578994e99
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - the shared helpers (Chebyshev rows, Adams basis) bit for bit against recorded | ||
| values; |
There was a problem hiding this comment.
Compare numerical helpers using rounding bounds rather than bits
For independently computed Chebyshev rows and Adams bases, this requirement would reject valid Julia implementations whose rounding differs from the recorded values. The existing tests/test_helpers.py::test_points_agree_with_the_plain_form asserts an operation-derived bound and records bitwise agreement only as information. Specify similarly justified bounds for Julia, retaining exact comparisons only for structural results; otherwise the proposed blocking suite contradicts the repository's numerical-comparison contract.
AGENTS.md reference: AGENTS.md:L337-L342
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Accepted, fixed in cf7e0c1. §6.6's Core group now compares the Chebyshev rows and the Adams basis with the plain Python forms within operation-derived bounds, as tests/test_helpers.py::test_points_agree_with_the_plain_form does. Bitwise agreement is recorded as information, not asserted, and exact comparison is kept for structural results only: refusals, the refused position and NaN propagation.
| - CPUs at 1 thread and the default budget: Apple M5 Max, Intel Xeon Gold 6240 and AMD | ||
| EPYC 7763; | ||
| - each GPU of §7.4, binary64 and binary32. | ||
| - **Whole runs** on two cases per machine. |
There was a problem hiding this comment.
Remove full-length runs from the qualification requirements
Requiring two whole runs per machine makes full-length propagation part of this change's qualification, because §7.6 explicitly requires the §7.5 records before merge. This conflicts with the repository's short-window qualification policy, which permits complete release tracking only as a separate activity, never as change qualification. Replace this requirement with short windows or move whole-run measurements to separately invoked release tracking outside the merge checklist.
AGENTS.md reference: AGENTS.md:L125-L137
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Already fixed in ab619ef, under the maintainer's decision on R10: whole runs became optional and are listed as not required for merge in §7.6. cf7e0c1 goes further: §7.5 now says whole runs are not part of this SEP's qualification at all. Like python -m benchmarks track, they are run separately by hand, never as a change's qualification.
| - At run time solprop points `juliacall` at that Julia and environment through | ||
| `PYTHON_JULIACALL_EXE` and `PYTHON_JULIACALL_PROJECT`, which bypasses juliapkg's | ||
| resolution. |
There was a problem hiding this comment.
Pin compatible JuliaCall and PythonCall versions together
When the Python extra selects a different juliacall release within its tested range (§5.1) from the PythonCall.jl version pinned in the committed manifest, this override bypasses the resolver that would reconcile them. The JuliaCall existing-environment documentation explicitly requires PythonCall.jl to match the JuliaCall version and states that loading Julia otherwise fails. Thus a successful setup can still leave the backend unusable, including across virtual environments sharing the depot. Specify matching bridge-version pairs and validate them before startup rather than independently allowing a Python version range and one fixed Julia manifest.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in ab619ef, the same point as R4 of the GPT-6 review. The julia extra pins juliacall exactly to the release paired with the manifests' PythonCall.jl (§§5.1, 6.10). The pair is part of the setup's identity and is checked before the runtime starts; a mismatch is refused by name. That check also covers several virtual environments sharing one setup, which §6.5's tests exercise.
| - The setup can run on an HPC login node and be used offline on compute nodes that | ||
| share the file system. |
There was a problem hiding this comment.
Materialize CUDA artifacts during GPU-less setup
On an HPC login node without a CUDA driver, the specified manifest instantiation and precompilation do not establish the promised offline GPU environment. The CUDA.jl precompilation documentation explains that without a driver, neither runtime nor compiler artifacts are selected or downloaded unless a toolkit version is explicitly configured. Consequently, setup can finish successfully while the first CUDA run on an offline compute node cannot compile kernels or load its runtime. Require setup to select the target toolkit and materialize its required artifacts before recording completion, and verify the workflow with networking disabled on the compute node.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in ab619ef, the same point as R5 of the GPT-6 review.
- Setup gains
--cuda-runtime, so it can select the CUDA runtime on a node with no GPU or driver. - It fetches and verifies every artifact, lazy ones included, before publishing the environment.
- Runs keep Julia's package manager offline, so a missing artifact is a setup refusal by name, never a download attempt.
- The CUDA qualification pass (§7.4) prepares the setup on a GPU-less node, then makes the first GPU run with networking disabled.
Answer every point of GPT-6's review and the automated review's matching finding, with the maintainer's decisions on R3, R8 and R10, and revise Part I: comparison classes M, A and C; the double-float algorithms, hypotheses and witnesses quoted exactly (the sloppy addition and the unanalysed division replaced); a phase domain with a three-word range reduction and an admission refusal above it; an exact juliacall pin; complete offline artifacts; per-vendor prerequisites (system ROCm); a setup identity with copied engine sources, locking and atomic publication; the signal policy and its phase 1 probes; taper and resume scope; short-window survival tests and mandatory records. Still Draft. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ab619ef630
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - **Class M, a matched discrete history:** fixed schedules, replays and fixed-order | ||
| methods. Both backends take the same steps by construction. Fields and every | ||
| statistic agree within SEP 25's rounding budget: the operation counts, with the | ||
| conditioning and propagated sensitivity of what is compared (SEP 25 §§4.3–4.4 and | ||
| [testing.md](../testing.md)). Each budget is derived in a comment beside its test. The |
There was a problem hiding this comment.
Classify ordinary variable-order replay as an accuracy comparison
For a Multistep replay, the revised Class M incorrectly assumes that sharing a step schedule matches the complete discrete computation. FixedSchedule records step lengths, not orders or rejected trials, and SEP 25 §4.4 explicitly explains that Adams replay still chooses orders and can therefore follow a different history. Requiring rounding-level agreement here would reject valid Julia replays even after the independent-adaptive checks were corrected. Classify ordinary variable-order replay as Class A, and reserve Class M for tests that explicitly match method, orders, startup and history.
AGENTS.md reference: AGENTS.md:L343-L346
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Accepted, fixed in cf7e0c1.
- §4 now reserves class M for a fully matched discrete computation: the same prescribed schedule, a method that makes no discrete choice on it (
Multistep(variable_order=False), the Runge–Kutta tableaux, RK4IP, the low-storage replay), and matched startup, restart history and save-plane handling. - It states, citing SEP 25 §4.4, that a
Multistepreplay is not class M, becauseFixedSchedulerecords step lengths only. - §7.1 places variable-order replay in class A, and so are groups, repeat runs and resumes whose method chooses orders. They are class M only when the method makes no discrete choice.
Variable-order Multistep replay is class A (SEP 25 section 4.4), class M needs a fully matched discrete computation; the helpers are compared within operation-derived bounds as tests/test_helpers.py does; whole runs are outside the qualification. Part III answers all six Codex findings. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Accept SEP 48, an optional Tier 3 Julia backend implementing SEP 47's complete execution-session contract on CPU, CUDA, ROCm and Metal. John Travers reviewed and accepted the proposal, approved the implementation-plan revisions, and delegated this proposal's acceptance and merge.
The acceptance revision makes setup identity and selection explicit, adds the execution-only
SimulationSettings.backend_optionsmapping, requires artifact-completeness checks before Julia loading, and defines main-thread initialization and atomic stop ownership. DF32 admission bounds actual replay and Adams-history spans, including restored history. Input noise is binary64 only; every binary32 path, including Metal, continues to refuse noise. Physics defaults are unchanged.Production adapter integration waits for SEP 47's merged descriptor/request split and generic preparation entry point. Package/setup work can start independently. Implementation will use one draft PR; that PR will be left ready for maintainer review rather than merged automatically.
The maintainer accepted a process-wide limitation in Julia 1.13.3: after Julia starts, SIGTERM may terminate the application without final output or a checkpoint, including with custom or ignored Python handlers. The proposal retains actual Python SIGINT handling and lifetime-owned atomic stop tokens. Issue #610 stays open to revisit signal preservation with stable Julia 1.14 and compatible device packages; it no longer blocks this SEP under the accepted policy.
Status: Accepted. Review: Reviewed (John Travers, human; GPT-6, AI). Implementation: package preparation only; production session work remains gated by phase 1 probes.
Tracks #608; requires #588 and the preparation changes in #607. Related: #559, #368, #557, #560, #590, #561, #333, #356 and #596. This documentation PR closes no implementation issue.
Validation on the acceptance tree:
git diff --checkand codespell on the three changed documents pass.pytest tests/test_seps.py: 349 passed, three expected process-SEP skips.poe docs: Sphinx builds cleanly with warnings as errors, using the lightweight environment.Separate macOS arm64 runtime feasibility probes confirmed foreign-thread atomic cancellation during threaded GC and restored custom SIGINT handling. They also independently reproduced fatal SIGTERM despite ignored/custom Python handlers; the maintainer explicitly accepted that limitation before merge. The amendment passed independent review. No propagation, GPU qualification or full code gate was run for this documentation-only change.