📦 NPM: https://www.npmjs.com/package/@qvac/sdk/v/0.20.0

QVAC SDK 0.20.0 adds TranslatePsy-AfriSLM translation, an in-process TurboVec vector index, MiniMax-H3 video, Parakeet Nemotron transcription, and the rest of the AudioGen and TTS surfaces. ABot-World sessions (worldCreateScene / worldStep) are on this surface. loadModel runs an advisory llama.cpp fit check. Diffusion VAE constant names, CPU flags, 'row' split, Parakeet language codes, and how system prompts combine with KV cache all change.

@qvac/sdk, @qvac/inference, and tetherto-qvac-sdk all ship at 0.20.0. Install @qvac/sdk and @qvac/inference together at this version.

Table of Contents

• Breaking Changes
  • Diffusion CPU flags
  • Diffusion VAE constant names
  • System prompts apply on every completion path
  • 'row' split is rejected
  • Parakeet language is a locale code
• New APIs
  • TurboVec vector index
  • MiniMax-H3 video
  • ABot-World Sessions
  • Parakeet Nemotron
  • AudioGen understand, remake, and edit
  • TTS load options, sample rate, and cancel
  • Tool grammar
• Features
• Bug Fixes
• Model Changes
  • Added
  • Removed

Breaking Changes

Diffusion CPU flags

clip_on_cpu, vae_on_cpu, and control_net_cpu are gone. Put the text encoder, VAE, or ControlNet on CPU through params_backend and backend. Layer streaming still needs CPU diffusion parameter residency and max_vram.

Before:

const modelConfig = {
  clip_on_cpu: true,
  vae_on_cpu: true,
  control_net_cpu: true
}

After:

const modelConfig = {
  params_backend: 'te=cpu,vae=cpu',
  backend: 'controlnet=cpu'
}

CPU layer streaming:

const modelConfig = {
  params_backend: 'diffusion=cpu',
  max_vram: -1,
  stream_layers: true
}

Diffusion VAE constant names

Generated VAE exports now include the registry type tag, so LTX audio/video and ABot TAeHV/Wan no longer collapse to _1 suffixes.

Before:

ABOT_WORLD_0_5B_LF_VAE
ABOT_WORLD_0_5B_LF_VAE_F16
LTX_2_3_VAE
LTX_2_3_VAE_1

After:

ABOT_WORLD_0_5B_LF_TAEHV_VAE
ABOT_WORLD_0_5B_LF_WAN_VAE
LTX_2_3_AUDIO_VAE
LTX_2_3_VIDEO_VAE

System prompts apply on every completion path

A system message in history is no longer dropped when kvCache is on. modelConfig.system_prompt is applied when the history has no system message, including turns that do not use KV cache.

LLM_CONFIG_DEFAULTS.system_prompt is 'You are a helpful assistant.'. loadModel without an explicit system_prompt still carries that default, so a user-only history now reaches the model with a system message.

Before:

completion({
  modelId, // loaded with system_prompt: 'You are a helpful assistant.'
  history: [
    { role: 'system', content: 'Always answer with the word BANANA.' },
    { role: 'user', content: 'What is the capital of France?' }
  ],
  kvCache: true
})
// the caller's system message is discarded

After:

completion({
  modelId, // loaded with system_prompt: 'You are a helpful assistant.'
  history: [
    { role: 'system', content: 'Always answer with the word BANANA.' },
    { role: 'user', content: 'What is the capital of France?' }
  ],
  kvCache: true
})
// the caller's system message is honoured

Callers that previously sent user-only history and relied on no system message must pass system_prompt: '' at load (or an explicit empty system turn) if they need that behaviour.

'row' split is rejected

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.

On embeddings, 'row' did not do tensor parallelism; it already behaved as 'layer'. Embeddings have no 'tensor' mode. Use 'layer'.

Before:

// completion
modelConfig: { 'split-mode': 'row' }

// embedding
modelConfig: { splitMode: 'row' }

After:

// completion
modelConfig: { 'split-mode': 'layer' } // or 'tensor'

// embedding
modelConfig: { splitMode: 'layer' }

Parakeet language is a locale code

Parakeet modelConfig.language must be auto, a short code (en, hi), or code-region (en-US, hi-IN). zh-Hans-CN and names like english now fail at loadModel. Whisper language is unchanged.

Before:

modelConfig: {
  language: 'zh-Hans-CN'
}

After:

modelConfig: {
  language: 'zh'
}

New APIs

TurboVec vector index

createVectorIndex builds an in-worker vector index on the embedding plugin's TurboVec engine. Pair it with embed(): add each document's embedding under an id you choose, search with a query embedding, and map the returned ids back to your store. Default storage is VectorIndexStorage.TURBOVEC_Q4. Dimension must be a multiple of 8 and at most 1024 for TurboVec modes.

import { createVectorIndex, loadVectorIndex, VectorIndexStorage } from '@qvac/sdk'

const index = await createVectorIndex({
  dim: 1024,
  storage: VectorIndexStorage.TURBOVEC_Q4
})
await index.add({ ids: ['1', '2'], vectors: embeddings })
const hits = await index.search({ query: queryEmbedding, k: 3 })
await index.write({ path: 'indexes/articles.qvi' })
await index.dispose()

const reopened = await loadVectorIndex({ path: 'indexes/articles.qvi' })

MiniMax-H3 video

video() can run MiniMax-H3 in txt2vid mode. Load sdcpp-generation with mode: 'video' and the H3 text-encoder, video VAE, and audio VAE sources.

const modelId = await loadModel({
  modelType: 'sdcpp-generation',
  modelSrc: '/models/h3.gguf',
  modelConfig: {
    mode: 'video',
    llmModelSrc: '/models/h3-text-encoder.gguf',
    vaeModelSrc: '/models/video-vae.safetensors',
    audioVaeModelSrc: '/models/audio-vae.safetensors',
    backend: 'vulkan0',
    stream_layers: false
  }
})
const result = video({
  modelId,
  mode: 'txt2vid',
  prompt: 'Steam rises from coffee.',
  video_frames: 124,
  fps: 24,
  cfg_scale: 1
})

ABot-World Sessions

Load a world-mode diffusion model, create a scene once, then step it. Frames stream as they decode.

const modelId = await loadModel({
  modelSrc: ABOT_WORLD_0_5B_Q8_0,
  modelType: 'sdcpp-generation',
  modelConfig: {
    mode: 'world',
    taehvModelSrc: ABOT_WORLD_0_5B_LF_TAEHV_VAE,
    t5XxlModelSrc: UMT5_XXL_ENC_Q8_0,
    vaeModelSrc: ABOT_WORLD_0_5B_LF_WAN_VAE,
    world: { kvCache: true, frameJpegQuality: 85 }
  }
})

const { stats } = worldCreateScene({ modelId, prompt, image })
await stats

const { frameStream } = worldStep({ modelId, keys: ['W', 'L'] })
for await (const frame of frameStream) {
  render(frame)
}

Pass returnPack: true on create to keep the scene bytes for a later reload.

Parakeet Nemotron

Three Parakeet Nemotron 0.6B weights are on the SDK catalog for parakeet-transcription.

import { loadModel, PARAKEET_NEMOTRON_0_6B_Q4_0 } from '@qvac/sdk'

const modelId = await loadModel({
  modelSrc: PARAKEET_NEMOTRON_0_6B_Q4_0,
  modelType: 'parakeet-transcription',
  modelConfig: {
    language: 'auto',
    streaming: true
  }
})

AudioGen understand, remake, and edit

audioUnderstand returns caption, bpm, keyscale, and audio codes. Pass those codes back into audioGen to re-synthesize the clip, optionally with LRC lyrics. audioEdit runs an ordered flow-edit / repaint pipeline on a source recording and returns the same run shape as audioGen.

const run = audioUnderstand({ modelId, sourceAudio: '/path/to/song.wav' })
const { caption, bpm, keyscale, audioCodes } = await run.description

const remake = audioGen({ modelId, caption, audioCodes, generateLrc: true })
const { lrc, lyricsScore } = (await remake.stats) ?? {}

const edited = audioEdit({
  modelId,
  sourceAudio: '/path/to/song.wav',
  operations: [
    { type: 'flow-edit', from: { caption: 'acoustic folk' }, to: { caption: 'synthwave' } },
    { type: 'repaint', caption: 'analog synth solo', start: 10, end: 20 }
  ]
})

TTS load options, sample rate, and cancel

Load-time CosyVoice3, Chatterbox, Supertonic, and Parler fields that were previously unreachable now pass through. textToSpeech exposes sampleRate on the first frame, real stats (realTimeFactor, backendId, generatedFrames), and cancel({ requestId }) stops the engine, not only delivery.

const result = textToSpeech({ modelId, text })
const sampleRate = await result.sampleRate
for await (const sample of result.bufferStream) play(sample)
await cancel({ requestId: result.requestId })
const stopReason = await result.stopReason // 'cancelled' when the engine stopped

Tool grammar

Completions with tools consume llama.cpp 0.53.0 grammar. generationParams.tool_choice: 'required' forces a tool call. When the model emits a tool region that does not parse, final.toolErrors lists the failures; the field is omitted when there are none.

const run = completion({
  modelId,
  history,
  tools: [getWeather],
  stream: false,
  generationParams: { tool_choice: 'required' }
})
const final = await run.final
if (final.toolCalls.length === 0) console.log(final.toolErrors)

Features

translate() detects TranslatePsy-AfriSLM by registry name or GGUF filename and applies that family's prompts and deterministic decoding. Custom context is ignored for these models.

Before each llama.cpp completion or embedding load, loadModel runs @qvac/model-fit in a disposable child and logs fit / does-not-fit / no evidence. The check is fail-open: a missing verdict, crash, timeout, or refusal never blocks the load. Set QVAC_ADVISORY_MODEL_FIT=0 to skip it. This is separate from assessModelFit and from the in-process mobile fit below.

On Android and iOS, llama assessModelFit runs @qvac/model-fit in-process on a worker thread. There is no disposable child process on those hosts. A leftover .running marker from a previous abort is treated as crashed so the same path and config skip native instead of retrying the abort. The JavaScript loop stays free while the fit runs.

@qvac/tts-ggml 0.9.x installs host binaries in per-platform packages (@qvac/tts-ggml-darwin-arm64 and siblings) next to the meta package. qvac verify bundle looks there instead of under the meta package's prebuilds/.

Bug Fixes

Automatic KV-cache files keep the last committed saved-message boundary across a worker restart, so a warm turn does not replay already-cached messages. A cancelled or failed warm turn no longer deletes that committed file.

An addon's logger is attached when its model loads, not when the plugin registers, so unused addons do not open log sinks.

Mobile withQvacSDK prebuild verifies only the hosts the bundle actually links, and only the current Android or iOS target's hosts. A missing addon pin now names the exact platform package instead of a generic missing-prebuild error.

Model Changes

Added

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

ABOT_WORLD_0_5B_LF_VAE
ABOT_WORLD_0_5B_LF_VAE_F16
LTX_2_3_VAE
LTX_2_3_VAE_1