pytorch/pytorch

Why this case study matters

pytorch is a multi-language ML mega-monorepo (~80k+ files; C++/CUDA core, Python frontend, Bazel + setup.py + CMake build, generated _C stubs, JIT/FX graph machinery, distributed/cuda/xpu/mps/rocm backend matrix). Its structural-validation surface is dominated by one artefact: .lintrunner.toml — a 1876-line TOML manifest declaring 57 distinct linter “adapters” orchestrated by lintrunner, pytorch’s bespoke per-file lint runner (Rust binary that spawns Python adapter scripts).

lintrunner exists because at the time pytorch needed it, no existing tool handled their orchestration needs (multi-language scopes, init-then-lint two-phase adapters, S3-vendored binary fetch, partial-file lint via @PATHSFILE fanout). Anyone running a custom orchestrator (pytorch’s lintrunner; tensorflow’s buildifier/yapf-driven CI; bazel’s own buildifier) is in the same situation.

Headline catch

alint isn’t trying to replace lintrunner — but the structural subset of .lintrunner.toml is exactly what alint specialises in.

Of the 57 lintrunner adapters:

• ~28 are pure structural/regex (24 grep_linter.py shims + EXEC + NEWLINE + SPACES + TABS) — every one expressible as a 5-10 line alint rule today
• ~21 are command shellouts to mature external tools (clang- format, clang-tidy, ruff, codespell, pyfmt, mypy, pyrefly, actionlint, shellcheck, cmake) — alint shells out the same way via command: rules
• ~8 are AST-aware/semantic/git-diff-aware (test_has_main libcst, scoped_library AST, set_linter, gb_registry, header_only, import_linter, pyproject_linter, no_workflows_on_fork, stable_shim_*) — these stay on lintrunner

So ~49 of 57 adapters (≈86 %) are within alint’s grammar today; the AST-aware tail (~8/57 ≈ 14 %) stays on lintrunner. alint is what you would have built instead of lintrunner if lintrunner had existed and you’d realised you only needed 86 % of its expressivity.

Where alint earns its keep here

• Inventory legibility — A new pytorch contributor staring at the structural-validation surface today has to read a 1876-line .lintrunner.toml, 30 Python adapter files in tools/linter/adapters/, the .editorconfig, .clang-format (3.4 KB), .clang-tidy (3 KB), .cmakelintrc, pyrefly.toml, mypy.ini (×2), and pytest.ini to understand what rules apply where. The alint config is one file, declarative, with each rule’s scope, severity, and rationale visible in 5-10 lines.
• Fail-fast latency — alint has zero adapter-spawn cost: it walks the tree once and runs every rule in parallel against the in-memory file bytes. lintrunner spawns one Python process per code per file batch. For the 28 structural-only adapters, alint should be 10-100× faster.
• Strongest cross-cutting v0.10 demand signal — pytorch’s WORKFLOWSYNC adapter (cross-workflow sync-tag consistency across 144 workflow files) is the cleanest example of the cross_file_value_equals pattern, the highest-priority new primitive across P2a + P2b. pytorch’s torch/header_only_apis.txt registry is the cleanest example of registry_paths_resolve (a flat text file lists symbols, each must appear in a .cpp test file). pytorch is also one of the four sources confirming import_gate (k8s + airflow + golang/go + pytorch).
• The orchestration-replacement narrative — for repos that already have a custom orchestrator, alint is the “don’t build your own orchestrator next time” pitch — adopt alint for the structural floor, keep the AST tail on whatever AST-aware tool you needed in the first place.

Future story angles

• Custom-orchestrator launch tile — pair pytorch with bazel’s buildifier and tensorflow’s CI for the “structural subset of YOUR custom orchestration layer” story.
• cross_file_value_equals v0.10 launch post — pytorch WORKFLOWSYNC is the canonical example of the most-demanded primitive in the entire inventory.
• registry_paths_resolve v0.10 launch post — pytorch header_only_apis.txt is the cleanest pattern instance.
• CLAUDE.md early-adopter narrative — pytorch was an early adopter of agent-friendly documentation; the .ci/docker/ content-hash trap flagged in CLAUDE.md is a foot-gun pattern worth its own follow-up post.