QVAC-23075 feat[api]: expose image_no_upscale in the SDK config schema

[yingying0906]

🎯 What problem does this PR solve?

• The VisionPsy Flash checkpoint and the base one are indistinguishable through the SDK. Their mmprojs declare identical vision hparams, so a Flash checkpoint loaded without this key silently runs base preprocessing. That changes the image token count, so it moves both accuracy and encode time.

📝 How does it solve it?

• Bumped LLM addon to 0.44.0
• image_no_upscale on the llamacpp completion config, an optional "on" | "off" that overrides the idefics3-style preprocessing rule the GGUF declares.
• "on" rounds the image's long side up to a whole number of slices and caps it, so an image below the cap keeps its own resolution and becomes far fewer slices.
• "off" always stretches the long side to the cap.
• Unset keeps the model's own value, so existing callers are unaffected. There is deliberately no default, so the addon keeps the GGUF value rather than being forced onto base preprocessing.
• The two sdk-python/_generated files are in the same commit on purpose. pr-checks-sdk-python.yml runs generate.py --check and fails with "Generated Python client is stale" if contract/schema.json moves without them, which then cascades into merge-guard / validate-pr.
• Replaces SmolVLM2 with VisionPsy Flash Q4 in the existing desktop, mobile, and Electron vision E2E tests. Bonus: This reduces the total download size by about 130 MB.

🧪 How was it tested?

CI run: https://github.com/tetherto/qvac/actions/runs/32269149688?pr=3854

• SDK and inference unit tests validate accepted, rejected, and unset image_no_upscale values.
• A new E2E test processes the same 64×64 image with "on" and "off" and verifies that "on" uses less than half as many prompt tokens.

🔌 API Changes

await sdk.loadModel({
  modelType: ModelType.llamacppCompletion,
  modelSrc: VISIONPSY_NANO_460M_MULTIMODAL_Q4_K_M,
  modelConfig: {
    projectionModelSrc: MMPROJ_VISIONPSY_NANO_460M_MULTIMODAL_Q8_0,
    image_no_upscale: 'on'
  }
})

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1217241821002488

[github-actions]

Review Status

Current Status: ✅ APPROVED
Approvals so far: Team Lead: 1, Member: 1

[github-actions]

License compliance — findings detected (warn-only)

Critical: 0 · High: 7 · Medium: 0

Dependency	License	Scope	Severity	Outcome
@qvac/classification-ggml@^0.19.1	(none detected)	runtime	High	blocks
@qvac/embed-llamacpp@^0.33.0	(none detected)	runtime	High	blocks
@qvac/llm-llamacpp@^0.44.0	(none detected)	runtime	High	blocks
@qvac/ocr-ggml@^0.17.0	(none detected)	runtime	High	blocks
@qvac/translation-nmtcpp@^0.9.0	(none detected)	runtime	High	blocks
@qvac/tts-ggml@^0.7.4	(none detected)	runtime	High	blocks
@qvac/vla-ggml@^0.20.0	(none detected)	runtime	High	blocks

How to resolve a blocking finding:

• Remove or replace the disallowed dependency, or
• If the license is genuinely acceptable, run the compliance SKILL and record the decision in .github/license-allowlist.yml (CODEOWNERS-reviewed), or
• For a one-off, a maintainer can apply the license-override label (High findings only; Critical cannot be overridden).

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./.github/actions/release-merge-guard
• ./docs/website
• ./packages/ggml-coload-smoke
• ./packages/fabric/test/integration
• ./packages/inference-addon-cpp/mobile
• ./packages/sdk/e2e
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/vla-ggml/sim/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/asr-ggml/benchmarks/server

[github-actions]

QVAC E2E — ios — ❌ failed

Totals: 90/106 passed · 2 failed · 97.8% · 859s
Config: suite=smoke · filter=(none) · exclude=(none)
Device pool: iPhone 16 Pro - iOS 18 — ☀️ highly available
View run · Artifacts: reports · Device Farm logs

Results by section

• vla: 0/4 ❌

Failed tests

• vla-hparams-shape: Setup failed: Failed to load model: Failed to load weights, error: failed to load SmolVLA model from: /var/mobile/Containers/Data/Application/C3AD2F51-18C7-469C-9041-78990EB252CE/Documents/.qvac/mode…
• vla-run-synthetic-shape: Setup failed: Failed to load model: Failed to load weights, error: failed to load SmolVLA model from: /var/mobile/Containers/Data/Application/C3AD2F51-18C7-469C-9041-78990EB252CE/Documents/.qvac/mode…

[github-actions]

QVAC E2E — android — ✅ all tests passed (92/106, 1987s)

Config: suite=smoke · filter=(none) · exclude=(none)
Device pool: Google Pixel 10 Pro - Android 16 — ☀️ highly available
View run · Artifacts: reports · Device Farm logs · Device Farm logs

[github-actions]

QVAC E2E — windows — ✅ all tests passed (106/106, 450s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports

[github-actions]

QVAC E2E — linux — ✅ all tests passed (106/106, 323s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports

[github-actions]

QVAC E2E — macos — ✅ all tests passed (106/106, 226s)

Config: suite=smoke · filter=(none) · exclude=(none)
View run · Artifacts: reports