QVAC-24061 chore: release @qvac/inference 0.20.0

[opaninakuffo]

🎯 What problem does this PR solve?

Cuts @qvac/inference 0.20.0 from current main so the engine can publish before @qvac/sdk 0.20.0. After #4338, the SDK published lane installs ^0.20.0 from npm; this package has to be on the registry first.

Metadata only: version, changelog, NOTICE. No SDK bump. No publish.

📝 How does it solve it?

• Stamp packages/inference to 0.20.0.
• Changelog from the lockstep 0.19.1 backmerge (4dbb66d8e / --base-version=0.19.1), plus the main-only PRs that sat between the previous .0 and that patch (#4010, #4234, #4342, #4384).
• Regenerated NOTICE (TranslatePsy AfriSLM, Nemotron 3.5 ASR, JS dep versions).

Companion SDK PR #4543 still lockstep-bumps inference on the SDK changelog branch. That bump belongs here; do not merge #4543 until this is published and the SDK PR drops the in-tree inference version change.

Do not add skip-sdk-pod-checks.

🧪 How was it tested?

• Changelog generator: --package=inference --base-commit=4dbb66d8e --base-version=0.19.1
• prettier --check on inference changelog markdown
• generate-notice.js inference (models + JS)

💥 Breaking Changes

Documented in packages/inference/changelog/0.20.0/breaking.md:

• Diffusion: clip_on_cpu / vae_on_cpu / control_net_cpu → params_backend / backend
• System prompts apply on every completion path, including kvCache
• splitMode: 'row' rejected; use 'layer'
• Diffusion VAE constant names include the type tag

🔌 API Changes

TurboVec createVectorIndex / loadVectorIndex, MiniMax-H3 video(), Parakeet Nemotron, AudioGen understand/remake/audioEdit, TTS sampleRate/cancel, tool grammar tool_choice / toolErrors. See changelog/0.20.0/api.md.

📦 Models

16 added, 4 removed. Full lists in changelog/0.20.0/models.md.

Added models

ABOT_WORLD_0_5B_LF_TAEHV_VAE
ABOT_WORLD_0_5B_LF_WAN_VAE
LTX_2_3_AUDIO_VAE
LTX_2_3_VIDEO_VAE
PARAKEET_0_6B_F16
PARAKEET_0_6B_Q4_0
PARAKEET_0_6B_Q8_0
PARAKEET_NEMOTRON_0_6B_F16
PARAKEET_NEMOTRON_0_6B_Q4_0
PARAKEET_NEMOTRON_0_6B_Q8_0
TRANSLATEPSY_AFRISLM_0_8B_TRANSLATION_Q4_K_M
TRANSLATEPSY_AFRISLM_0_8B_TRANSLATION_Q8_0
TRANSLATEPSY_AFRISLM_2B_TRANSLATION_Q4_K_M
TRANSLATEPSY_AFRISLM_2B_TRANSLATION_Q8_0
TRANSLATEPSY_AFRISLM_4B_TRANSLATION_Q4_K_M
TRANSLATEPSY_AFRISLM_4B_TRANSLATION_Q8_0

Removed models

ABOT_WORLD_0_5B_LF_VAE
ABOT_WORLD_0_5B_LF_VAE_F16
LTX_2_3_VAE
LTX_2_3_VAE_1

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

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):

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

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.

[github-actions]

QVAC E2E — base run recorded

Base: run 35407621343 · suite full · sha 5ff84e8

Platform	Passed	Total	Failed
android/android	382	526	0
desktop/linux	520	526	0
desktop/macos	520	526	0
ios/ios	370	526	0

All recorded platforms are green — nothing for test-e2e-rerun-failed to re-run.

[lauripiisang]

Four changelog issues, all in packages/inference/changelog/0.20.0/.

1. splitMode: 'row' entry names the wrong key and enum for completion models.
breaking.md and CHANGELOG_LLM.md say "modelConfig.splitMode is 'none' or 'layer'".
There are two keys with two enums in src/schemas/llamacpp-config.ts:

• llmConfigBaseSchema → 'split-mode', now ['none', 'layer', 'tensor']
• embedConfigBaseSchema → splitMode, now ['none', 'layer']

As written, someone running 'split-mode': 'tensor' reads this as tensor parallelism
being dropped, and someone hitting the validation error can't find the key they set.
Neither spelling appears in the SDK source or the generated API page, so this entry is
the only public documentation of it. Suggested replacement:

Fabric 10549.1.0 dropped llama.cpp's unused row split.

• Completion models: modelConfig['split-mode'] is 'none', 'layer' or 'tensor'. 'row' is rejected.
• Embedding models: modelConfig.splitMode is 'none' or 'layer'. 'row' is rejected.

'tensor' on completion models is unaffected.

2. The embed removal needs one clause on what was lost.
Before this cut, the completion description called 'row' "legacy; degrades to 'layer'"
while the embed description called it "tensor parallelism". Please confirm against the
Fabric change which is accurate and say so — embedding models have no 'tensor' to move
to, so "use 'layer'" is either a rename or a downgrade.

3. Parakeet language validation tightening is undocumented.
src/schemas/transcription-config.ts added
/^(|auto|[a-zA-Z]{2,3}(-[a-zA-Z]{2,4})?)$/ to a field that was plain z.string().
Configs that used to load now fail at loadModel — e.g. zh-Hans-CN, or a language
name rather than a code. This belongs in breaking.md.

4. No ABot-World section, though the package exports the API.
src/index.ts exports worldCreateScene and worldStep, and the SDK 0.20.0 notes carry
"ABot-World Sessions" under New APIs. The two changelogs mirror each other section for
section otherwise.

Note (1), (3) and (4) also apply to packages/sdk/changelog/0.20.0/ in #4543, plus
docs/website/content/docs/reference/release-notes/v0.20.x.mdx, and both backmerges.
packages/inference/CHANGELOG.md is currently byte-identical to CHANGELOG_LLM.md —
worth keeping that true.

Separately: CHANGELOG.md still goes 0.20.0 → 0.17.0, while npm has 0.17.1, 0.18.0,
0.18.1, 0.18.2, 0.19.0 and 0.19.1. Since this PR backfills changelog/0.17.0/, adding
one-line stub sections for those six would close the gap for good.

[github-actions]

QVAC E2E — android — ✅ all tests passed (382/526, 4734s)

Config: suite=(none) · filter=(none) · exclude=(none)
Inference: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/inference@0.20.0 · Test-suite: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/test-suite@0.11.3
Device pool: Samsung S25 Ultra - Android 15 — ☀️ highly available
View run · Artifacts: reports · Device Farm logs

[github-actions]

QVAC E2E — windows — ⚠️ no results

Config: suite=(none) · filter=(none) · exclude=(none)
Inference: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/inference@0.20.0 · Test-suite: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/test-suite@0.11.3
View run

The test job did not produce a results artifact (e.g. no device started within the start-timeout, or a job-level failure). Check the run and the device pool status above.

[github-actions]

QVAC E2E — ios — ✅ all tests passed (370/526, 2497s)

Config: suite=(none) · filter=(none) · exclude=(none)
Inference: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/inference@0.20.0 · Test-suite: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/test-suite@0.11.3
Device pool: iPhone 16 Pro - iOS 18 — ☀️ highly available
View run · Artifacts: reports · Device Farm logs

[github-actions]

QVAC E2E — linux — ✅ all tests passed (520/526, 2717s)

Config: suite=(none) · filter=(none) · exclude=(none)
Inference: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/inference@0.20.0 · Test-suite: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/test-suite@0.11.3
View run · Artifacts: reports

[github-actions]

QVAC E2E — macos — ✅ all tests passed (520/526, 1931s)

Config: suite=(none) · filter=(none) · exclude=(none)
Inference: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/inference@0.20.0 · Test-suite: branch:5ff84e82b705899012b2affc78069dd8ded10d3b:@qvac/test-suite@0.11.3
View run · Artifacts: reports