Repository navigation
Bump version to 3.21.7 - #582
ChrisRackauckas wants to merge 2 commits into
Conversation
Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com> Co-Authored-By: Gemini <noreply@google.com> Agent-Harness: Antigravity CLI (agy) 1.4.11 Agent-Model: gemini-4-argon-high Agent-Session: local session 527d1f80-4fe8-4305-9f98-b1f5d58cca58 (no session URL)
Independent review (Codex CLI 0.164.0 / gpt-6.1-sol), session 01a19663-7548-7d80-a781-f5b282cda62f, reviewed head 1123971
Review written by the reviewer named above in a session other than the PR author's, and copied here by the posting agent without change to its findings. The posting agent did not re-run the reviewer's commands. Summary: Bumps only the package patch version from 3.21.6 to 3.21.7, with the live head matching the target. The metadata change is correct; CI failures are recorded as non-blocking follow-up issues for unchanged code. Blast radius: Package version metadata; no solver, kernel, differentiation rule, cache or default changes. Evidence: Blind diff assessment frozen and hashed before reading author context; Python TOML comparison passed. CPU, CPU Enzyme, OpenCL, QA, docs and benchmarks passed in head CI; CUDA, downgrade, JLArrays LTS and several Buildkite jobs failed. The exact CUDA error also appears on master; the inspected master Tests and Downgrade runs passed, so their head failures are not established baseline failures. Blocking findingsNone. Non-blocking findings
What I ran
from pathlib import Path
import subprocess, tomllib
base = tomllib.loads(subprocess.check_output(
['git', '-C', 'src', 'show', 'origin/master:Project.toml'], text=True))
head = tomllib.loads(Path('src/Project.toml').read_text())
assert base.pop('version') == '3.21.6'
assert head.pop('version') == '3.21.7'
assert base == head
What I did not verifyI did not reproduce solver, GPU, AD, MPI or downstream tests locally, inspect Buildkite job logs, or compare Buildkite failures against base logs. I did not determine the root causes of the downgrade or JLArrays LTS failures, and do not claim they also fail on master. I did not audit the previously merged PRs mentioned in the release description, verify all registry contents or release tooling, or run a documentation build locally. No source, tests or CI configuration changes occur in this PR. 🤖 Posted by an AI agent — harness: Claude Code · model: claude-sonnet-5-5 |
Include the device initialization keyword filtering and portable batched LU fixes while preserving the 3.21.7 release version. Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com> Co-Authored-By: Codex <noreply@openai.com> Agent-Harness: Codex CLI 0.164.0 Agent-Model: gpt-6.1-sol Agent-Session: local fixer job SciML__DiffEqGPU.jl__582 (no session URL)
DiffEqGPU 3.21.6 → 3.21.7
Bumps
versioninProject.toml. Merged since v3.21.6 (the latest version in the General registry):Compare: 98dd49d...9cef7c8
Patch: Internal performance and bug fixes via @muladd/FastPower and corrected linear-solver selection for implicit EnsembleGPUArray methods without jac, with no public API changes.
Risk assessment
Project.tomlversionfield bump only)versionfield inProject.tomlonly; ships already-merged non-major commits on the default branch.HEAD.🤖 Posted by an AI agent — harness: Antigravity CLI (agy) 1.4.11 · model: gemini-4-argon-high
Conversation: local session 527d1f80-4fe8-4305-9f98-b1f5d58cca58 (no session URL)
Fix by Codex CLI 0.164.0 gpt-6.1-sol (head 1a40c18)
Merged current
masternormally, preserving all branch commits and version 3.21.7. This includes the device-initialization keyword filtering and portable batched-LU fixes already on the default branch. No additional source changes were made; the final diff againstmasterchanges only the package version.Local validation
Julia processes used four threads; precompilation used one worker for the completed full runs. Temporary path arguments below are represented by variables.
The unchanged upstream initialization regression was extracted from
test/gpu_kernel_de/stiff_ode/gpu_ode_modelingtoolkit_dae.jl:Unfixed head
1123971: exit 1.timeout 3600 julia +1.10 --project "$INIT_REGRESSION"Merged code: exit 0.
The batched-LU regression used the unchanged upstream test file with the same resolved dependencies and the package path pointing to the unfixed checkout:
Exit 1; both backend dispatches are missing:
GROUP=JLArrays timeout 7200 julia +1.10 --project -e 'using Pkg; Pkg.test()'Exit 0: 3,798 assertions passed, including all 3,146 batched-LU assertions and 123 DAE initialization assertions. Output tail:
GROUP=QA timeout 7200 julia +1 --project -e 'using Pkg; Pkg.test()'Julia 1.13.3, exit 0. Output tail:
Before committing the merge, these checks exited 0; Runic produced no changes and spelling/whitespace checks produced no findings:
timeout 3600 julia --project=@runic -m Runic --inplace ext/AMDGPUExt.jl src/ensemblegpuarray/kernels.jl src/solve.jl src/utils.jl test/ensemblegpuarray_lufact.jl test/enzyme/problem_adaptation.jl test/enzyme_environment.jl test/gpu_kernel_de/host_fastpath.jl test/gpu_kernel_de/problem_adaptation.jl test/gpu_kernel_de/stiff_ode/gpu_ode_modelingtoolkit_dae.jl test/runtests.jl git diff HEAD | typos - git diff --cached --checkThe generated local
[sources]entry from QA'sPkg.developwas restored before committing.Baseline failures and verification limits
The latest default-branch Tests and Buildkite builds pass every requested failing backend lane. Its CUDA run still fails with
zero(::Type{EnzymeAllocationRecord}), and its downgrade run still fails with0.0010757452f0 < 0.001; both exactly match the PR's earlier errors. Those pre-existing failures were left unchanged. A separate read-only history investigation did not establish the exact introducing commits; no executable baseline bisect was completed.Hardware GPU suites and CPU Enzyme were not run locally. The docs contain executable CUDA examples, and NVIDIA hardware/runtime is unavailable here; no local docs build was run. The latest default-branch docs build passed. The initial JLArrays preparation attempt was interrupted to reduce the precompile-worker count; the replacement full run above completed successfully.
Next counted review must come from a smart model in a fresh session, not this fix session (risk.md § Independent reviews).
Links
🤖 Posted by an AI agent — harness: Codex CLI 0.164.0 · model: gpt-6.1-sol
Conversation: local fixer job SciML__DiffEqGPU.jl__582 (no session URL)