ci: align pre-commit hooks and GitHub templates

[mc-nv]

What does the PR do?

Aligns this repository with the org-wide setup consolidated in triton-inference-server/.github:

• .pre-commit-config.yaml: shared baseline hooks, conventional-commit message validation (commit-msg stage), and the centralized add-license hook from the org .github repository (rev: v1.4.3 — excludes .github/ templates, never rewrites LICENSE files).
• Pre-commit CI runs only on files modified by the PR.
• Caller stub for the org-wide reusable conventional-pr workflow (@v1.4.3): validates the PR title against Conventional Commits (hard gate — it becomes the squash-merge commit), derives one human-readable label per distinct type found in the title and all conforming commit subjects (e.g. ci: → CI/CD, feat: → feature, fix: → fix), enforces org-wide label colors/descriptions, detects cherry-picks, and fails if no type is derivable and no type label is assigned.
• No per-repo templates: PR templates and issue routing are inherited from the org-wide defaults in triton-inference-server/.github (issues route to the server repository).

Repo-specific: keeps this repo's deliberate Ruff formatter/linter + pyupgrade toolchain (per review feedback); the centralized add-license hook is NOT enabled because this repository has no LICENSE file — follow-up: add a LICENSE, then enable it.

Depends on triton-inference-server/.github#5 (pinned tags current: v1.4.3 — already exist, CI is green).

Pros / Cons

Pros

• Single home (org .github repo) for hooks, templates, issue routing, and the PR-title workflow — one change propagates everywhere.
• Repos carry only a config file and two small workflow stubs; no template copies to drift.
• Commit/PR title format enforced both locally (commit-msg hook) and in CI (reusable workflow), with automatic type labels.
• License hooks never modify LICENSE files.

Cons / risks

• Version-pinned dependency on the org .github repository (tags are write-once; changes ship as a new tag + rev bump).
• Template inheritance requires the org .github repository to remain public.

Related Issues / PRs

• Resolves: TRI-1100

Related PRs:

• ci: migrate the centralized add-license hook to the org .github repository developer_tools#61
• ci: align pre-commit hooks and GitHub templates core#517
• ci: align pre-commit hooks and GitHub templates common#161
• ci: align pre-commit hooks and GitHub templates backend#127
• ci: align pre-commit hooks and GitHub templates python_backend#449
• ci: align pre-commit hooks and GitHub templates tensorrt_backend#124
• ci: align pre-commit hooks and GitHub templates onnxruntime_backend#351
• ci: align pre-commit hooks and GitHub templates pytorch_backend#200
• ci: align pre-commit hooks and GitHub templates openvino_backend#119
• ci: align pre-commit hooks and GitHub templates identity_backend#36
• ci: align pre-commit hooks and GitHub templates checksum_repository_agent#14
• ci: align pre-commit hooks and GitHub templates local_cache#16
• ci: align pre-commit hooks and GitHub templates redis_cache#24
• ci: align pre-commit hooks and GitHub templates repeat_backend#19
• ci: align pre-commit hooks and GitHub templates square_backend#22
• ci: align pre-commit hooks and GitHub templates third_party#80
• ci: align pre-commit hooks and GitHub templates tutorials#158
• ci: align pre-commit hooks and GitHub templates server#8890
• ci: align pre-commit hooks and GitHub templates client#909
• ci: align pre-commit hooks and GitHub templates tensorrtllm_backend#861
• ci: align pre-commit hooks and GitHub templates vllm_backend#133
• ci: align pre-commit hooks and GitHub templates perf_analyzer#467
• chore: never modify the COPYRIGHT template and guard .github templates model_analyzer#1032

Test plan

• pre-commit validate-config passes; pre-commit run --files <PR diff> passes locally with the centralized hooks pinned to the .github branch SHA.
• After .github#5 merges and v1.4.3 exists: the conventional-pr check validates this PR's own title and applies the ci label.
• No LICENSE file content is changed by this PR.

Caveats

• None beyond the merge-order note above.

Checklist

• PR title follows <commit_type>: <Title> (conventional commit)
• I ran pre-commit locally on all files changed by this PR and it passes
• Copyright header is correct on all changed files
• External contributors: I have read the Contribution guidelines and signed the Contributor License Agreement

[Vinya567]

Was the Ruff → black/isort/flake8 swap intentional for TRI-1100? Has pre-commit run --all-files been run to confirm no surprise formatting diffs?

[mc-nv]

Good catch — the Ruff → black/isort/flake8 swap was not intentional. This repo's Ruff formatter/linter + pyupgrade setup is a deliberate toolchain choice (with the trt_llm excludes), and the alignment should have preserved it the same way it preserved perf_analyzer's newer revs and client's mypy. Fixed in ab760e4: the config is now main's Ruff/pyupgrade toolchain verbatim, plus only the shared additions (commit-msg conventional-commit validation, default_install_hook_types).

On --all-files: it was not run repo-wide, by design — CI checks only the files a PR modifies, and the fleet-wide policy avoids mass reformatting. For the record I did measure the impact of the (now-reverted) swap before answering: black/isort would have wanted changes in roughly half a dozen files, i.e. real churn — which the revert makes moot. With Ruff kept, the PR diff passes all hooks cleanly and no repo-wide formatting change is introduced.

[Vinya567]

Good catch — the Ruff → black/isort/flake8 swap was not intentional. This repo's Ruff formatter/linter + pyupgrade setup is a deliberate toolchain choice (with the trt_llm excludes), and the alignment should have preserved it the same way it preserved perf_analyzer's newer revs and client's mypy. Fixed in ab760e4: the config is now main's Ruff/pyupgrade toolchain verbatim, plus only the shared additions (commit-msg conventional-commit validation, default_install_hook_types).

On --all-files: it was not run repo-wide, by design — CI checks only the files a PR modifies, and the fleet-wide policy avoids mass reformatting. For the record I did measure the impact of the (now-reverted) swap before answering: black/isort would have wanted changes in roughly half a dozen files, i.e. real churn — which the revert makes moot. With Ruff kept, the PR diff passes all hooks cleanly and no repo-wide formatting change is introduced.

Thanks for the fix. Verified ab760e4 keeps Ruff/pyupgrade verbatim and only adds the shared alignment pieces (commit-msg hook, conventional-pr stub, modified-files CI). LGTM.

[greptile-apps]

Greptile Summary

This PR aligns the repository with the org-wide CI/tooling baseline defined in triton-inference-server/.github: a new thin caller workflow for conventional commit/PR-title validation, a reworked pre-commit workflow that runs only on PR-modified files, and a commit-msg stage hook for local conventional-commit enforcement.

• conventional-pr.yml: New caller stub delegates to the org reusable workflow (@v1.4.3); uses the dual-event (pull_request / pull_request_target) conditional pattern to give fork PRs write access for labeling while keeping same-repo PRs on the safer event.
• pre-commit.yml: Replaces the blanket pre-commit/action run with a scoped git diff | xargs pre-commit run --files invocation; --diff-filter=d skips deleted paths and --no-run-if-empty handles deletion-only PRs cleanly.
• .pre-commit-config.yaml: Activates the commit-msg hook stage via default_install_hook_types and adds the conventional-pre-commit hook to enforce Conventional Commits locally.

Confidence Score: 5/5

Safe to merge — all changes are CI configuration; no application code is touched and the dual-event workflow guard is correctly constructed.

The three changed files are CI/tooling config only. The conventional-pr caller correctly isolates same-repo vs fork triggers. The pre-commit workflow scoping logic (diff-filter, no-run-if-empty, null-delimited xargs) is sound. The one finding — an unpinned major-version tag on actions/cache — is a consistency nit that does not affect correctness.

.github/workflows/pre-commit.yml — the actions/cache@v4 pin is the only item worth a second look for consistency with the fully-pinned neighbouring actions.

Important Files Changed

Filename	Overview
.github/workflows/conventional-pr.yml	New caller stub for the org-wide conventional-pr reusable workflow; uses the correct dual-event / conditional pattern to gate same-repo vs fork PRs and grants only the minimum permissions needed for label management.
.github/workflows/pre-commit.yml	Switched from the pre-commit action to a manual install + scoped run; correctly handles deleted files (--diff-filter=d) and empty diffs (--no-run-if-empty). actions/cache is pinned to a mutable @v4 tag while the other actions use full semver pins.
.pre-commit-config.yaml	Adds default_install_hook_types to activate commit-msg stage and installs the conventional-pre-commit hook at that stage; existing hooks are unchanged.

Sequence Diagram

sequenceDiagram
    participant Dev as Developer
    participant GH as GitHub Actions
    participant OrgWF as org .github reusable workflow (v1.4.3)
    participant PC as pre-commit

    Dev->>GH: Opens / updates PR
    GH->>GH: conventional-pr.yml triggers (pull_request or pull_request_target)
    GH->>GH: if-condition selects correct event for same-repo vs fork
    GH->>OrgWF: "uses: conventional-pr.yml@v1.4.3"
    OrgWF->>GH: Validates PR title (Conventional Commits)
    OrgWF->>GH: "Derives & applies type label (ci, feat, fix ...)"

    GH->>GH: pre-commit.yml triggers (pull_request)
    GH->>GH: "git diff --name-only -z --diff-filter=d HEAD^1 HEAD"
    GH->>PC: xargs -0 --no-run-if-empty pre-commit run --files changed-files
    PC-->>GH: Hook results (ruff, pyupgrade, codespell, pre-commit-hooks)

    Dev->>Dev: git commit (local)
    Dev->>PC: commit-msg hook fires (conventional-pre-commit v4.4.0)
    PC-->>Dev: Pass / fail on commit message format

Loading

_{Reviews (3): Last reviewed commit: "ci: harden CI workflows and configs per ..." | Re-trigger Greptile}

[Vinya567]

LGTM.

[mc-nv]

Closing: the team is moving away from centralized org-level configuration in favor of per-repository self-contained setups (see triton-inference-server/server#8897 for the first decentralized implementation). Branch retained for reference. TRI-1100