Pin imageDiscoverMemoryEstimate tests to Mac host#24
Merged
cryptopoly merged 1 commit intomainfrom May 1, 2026
Merged
Conversation
The imageDiscoverMemoryEstimate describe block called the function with no explicit device, which falls through to inferDeviceFromHostPlatform in src/utils/videos.ts: - macOS host -> 'mps' device path - Linux/Win -> 'cuda' device path The expected values in the three failing tests were calibrated against the MPS path (FLUX text encoders, MPS runtime overhead, MPS-specific runtimeFootprint overrides). On a Mac developer box they pass; on the ubuntu-latest CI runner the host-inference returns 'cuda' and the function picks the CUDA branch, producing different numbers: modelFootprintGb 7.5 -> 4.5 (loses MPS multiplier) modelFootprintGb 41.3 -> 18.3 (no MPS GGUF overhead) modelFootprintGb 20 -> 16 (skips runtimeFootprintMpsGb override) This has been broken on CI since PR #20 (video-gen merge) introduced the host-platform fallback. Every push to main since 2026-04-29 has failed Build Desktop App > test on these three lines. Fix - Stub navigator + navigator.userAgentData.platform to 'macOS' inside the imageDiscoverMemoryEstimate describe block via beforeEach + vi.stubGlobal. - Restore globals after each test via afterEach + vi.unstubAllGlobals. The assessImageGenerationSafety describe block above is untouched because it always passes an explicit device, so host-inference never fires there. Verified locally: all 207 tests pass on macOS.
cryptopoly
added a commit
that referenced
this pull request
May 1, 2026
PR #22's no-system-RAM-fallback path returns {'vram_total_gb': None} on Linux CI runners (no torch.cuda, no nvidia-smi). The pre-existing test_snapshot_vram_values_are_numeric required (int, float) which breaks on those runners. This fix originally landed in branch fix/test-host-platform-mock (commit 3b147a9) but was pushed after PR #24 had already merged, so only the imageDiscoverMemoryEstimate Mac pin (commit 7bbeeef) made it into main. The orphan commit went unnoticed until run 25223969487 on this PR's first CI ride re-surfaced the same failure. Loosen the type check to (int, float, type(None)) and rename the test to ..._numeric_or_none so the intent is loud at the call site.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three tests in `src/utils/tests/images.test.ts` have been failing on CI since 2026-04-29 (PR #20 video-gen merge). They pass locally on Mac developer boxes but fail on `ubuntu-latest` because:
CI failure summary on commit 92e6ca9:
```
modelFootprintGb 7.5 -> 4.5 (no MPS multiplier)
modelFootprintGb 41.3 -> 18.3 (no MPS GGUF overhead)
modelFootprintGb 20 -> 16 (skips runtimeFootprintMpsGb override)
```
Fix
Stub `navigator` + `navigator.userAgentData.platform` to `"macOS"` inside the `imageDiscoverMemoryEstimate` describe block via `beforeEach` + `vi.stubGlobal`. Restore globals after each test via `afterEach` + `vi.unstubAllGlobals`.
The `assessImageGenerationSafety` describe block above is untouched — it always passes an explicit `device`, so host-inference never fires there.
Test plan
Why now
Surfaced on the v0.7.2 smoke-test follow-up merges (#21, #22, #23). The bug pre-dates all three of those PRs but was on the list of CI failures inherited from PR #20.