ci: add minimal-version resolution matrix to cpu and gpu jobs

[janickm]

Summary

Add a resolution: [pinned, minimal] matrix dimension to both CI jobs, validating that declared minimum dependency bounds actually work.

How it works

Matrix cell	Resolution strategy	What it validates
pinned	uv sync (lock file)	Current behavior -- tests against exact locked versions
minimal	uv sync --resolution lowest-direct	Declared minimum bounds are installable and pass tests

The minimal cell re-resolves the workspace from scratch using the lowest allowed version for each direct dependency (e.g. torch>=2.9 resolves to torch==2.9.0 instead of 2.12.0). This catches cases where code uses APIs not available at the declared minimum.

Details

• Both cells run linting (pre-commit) and unit tests
• Wheel build only runs on pinned (we publish against locked versions)
• publish-testpypi requires all matrix cells (needs: [cpu, gpu]) -- both pinned and minimal must pass before publishing
• fail-fast: false -- both cells run to completion even if one fails
• Cache keys are separated: cpu-pinned / cpu-minimal, gpu-sm{arch}-pinned / gpu-sm{arch}-minimal

Prerequisites

Depends on #211 which pins lower bounds on all workspace dependencies (enabling --resolution lowest-direct to resolve without failures).

Local validation

Tested locally with torch 2.9.0 (lowest-direct resolution):

• CPU tests: 50 passed
• GPU tests: 21 passed (RTX A6000)

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

This PR adds a resolution: [pinned, minimal] strategy matrix to both the cpu and gpu CI jobs, with the minimal cell re-resolving dependencies via uv sync --resolution lowest-direct to validate declared minimum bounds. Lint and wheel build steps are correctly gated to the pinned cell; cache keys are cleanly separated; and publish-testpypi gates on all matrix cells passing.

• The lock file itself was regenerated with --resolution lowest-direct, writing resolution-mode = \"lowest-direct\" into uv.lock [options]. This makes the mode sticky for all future uv lock invocations, so both the pinned and minimal cells now install the same version set — the CONTRIBUTING.md promise of "pinned versions from uv.lock" no longer refers to latest-compatible locked versions.
• flashdreams/pyproject.toml also drops Windows-specific overrides (triton-windows, the pytorch-cu130 index, and win32 PyTorch source routing), and the lock file drops all Windows platform markers.
• Several package lower bounds are bumped workspace-wide (opencv-python-headless 4.5→4.6, pytest-asyncio 0.23→0.24, torchvision 0.26→0.24, torch 2.11→2.9 in two integrations).

Confidence Score: 4/5

Safe to merge after confirming whether the lock file being permanently lowest-direct is intentional.

The CI plumbing is sound — matrix expansion, cache key separation, conditional lint/wheel steps, and publish gating all work as described. The uv.lock gaining resolution-mode = lowest-direct in its options block makes every future uv lock call continue resolving to lowest-direct, meaning both CI cells install identical package versions with no coverage of latest compatible versions. The ambiguity around intent is what keeps this from a clean score.

uv.lock — specifically the new [options] resolution-mode entry and the loss of all Windows platform markers.

Important Files Changed

Filename	Overview
.github/workflows/ci.yml	Adds resolution matrix to both cpu and gpu jobs; skips lint and wheel build for minimal; cache keys correctly separated; one minor inefficiency in minimal CPU install
uv.lock	Regenerated with --resolution lowest-direct, adding [options] resolution-mode = "lowest-direct" which makes the lock file sticky to lowest-direct; also drops all Windows platform markers; "pinned" and "minimal" CI cells now resolve the same dependency versions
flashdreams/pyproject.toml	Removes Windows-specific triton-windows dep and pytorch-cu130 index override; bumps opencv lower bound from 4.5 to 4.6
pyproject.toml	Bumps pytest-asyncio minimum from 0.23 to 0.24 in test and lint dependency groups
CONTRIBUTING.md	Updates CI documentation to reflect both pinned and minimal resolution strategies; accurate if lock file keeps non-lowest-direct semantics, slightly misleading if lock file is sticky to lowest-direct
integrations/omnidreams/pyproject.toml	Lowers torch minimum from 2.11 to 2.9 and torchvision from 0.26 to 0.24, consistent with workspace-wide torch floor; bumps opencv lower bound to 4.6

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    push[Push / merge_group] --> cpu & gpu

    subgraph cpu["cpu job (matrix: pinned | minimal)"]
        C1[Install system deps] --> C2[Checkout + setup uv]
        C2 --> C3{matrix.resolution}
        C3 -- pinned --> C4["uv sync --extra dev --group lint"]
        C3 -- minimal --> C5["uv sync --resolution lowest-direct\n--extra dev --group lint"]
        C4 --> C6["Run linter checks (pinned only)"]
        C4 --> C7["Build wheel (pinned only)"]
        C5 -. skipped .-> C6
        C5 -. skipped .-> C7
        C6 --> C8[pytest -m ci_cpu]
        C7 --> C8
        C5 --> C8
    end

    subgraph gpu["gpu job (matrix: pinned | minimal)"]
        G1[Detect GPU arch] --> G2[Checkout + setup uv]
        G2 --> G3{matrix.resolution}
        G3 -- pinned --> G4["uv sync --extra dev"]
        G3 -- minimal --> G5["uv sync --resolution lowest-direct --extra dev"]
        G4 --> G6[pytest -m ci_gpu]
        G5 --> G6
        G6 --> G7[Trim uv cache]
    end

    cpu & gpu --> pub{all 4 cells pass?}
    pub -- yes, main branch --> PYPI[publish-testpypi]
    pub -- no --> FAIL[Block publish]

Loading

Comments Outside Diff (1)

1. uv.lock, line 373-374 (link)

   Lock file pinned to lowest-direct — "pinned" and "minimal" cells test identical versions

   The resolution-mode = "lowest-direct" entry in the lock file's [options] block is sticky: every future uv lock invocation (dependency updates, bots) will continue resolving to lowest-direct. This means the "pinned" cell (uv sync, which installs what's in the lock file) and the "minimal" cell (uv sync --resolution lowest-direct, which re-resolves) resolve to the same package versions. Concretely, the lock file now pins aiohttp to 3.9.0 instead of the previously locked 3.13.5. There is no longer any CI coverage against the latest compatible versions.

   If the intent is for the lock file to represent the minimum viable baseline, the CONTRIBUTING.md wording ("pinned versions from uv.lock") could be clarified to "pinned minimum versions". If the intent is for pinned to continue tracking latest compatible versions, the lock file should be regenerated without --resolution lowest-direct (removing this [options] block).

_{Reviews (1): Last reviewed commit: "ci: add minimal-version resolution matri..." | Re-trigger Greptile}

[greptile-apps]

The minimal CPU cell installs --group lint (which pulls in pre-commit and ty) even though the "Run linter checks" step is skipped for that cell. Only --group test is needed to satisfy pytest-asyncio and imageio-ffmpeg for the unit-test step, saving two unnecessary package installs per minimal run.

Suggested change

	if [ "${{ matrix.resolution }}" = "minimal" ]; then
	uv sync --resolution lowest-direct --extra dev --group lint \
	--no-install-package transformer-engine-torch
	else
	if [ "${{ matrix.resolution }}" = "minimal" ]; then
	uv sync --resolution lowest-direct --extra dev --group test \
	--no-install-package transformer-engine-torch
	else

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!