📦 NPM: https://www.npmjs.com/package/@qvac/sdk/v/0.19.0

QVAC SDK 0.19.0 is the first release after @qvac/inference became the in-process engine. You can assess whether a model will fit before downloading it, walk an ABot-World session, generate MiniMax music, and transcribe with Parakeet Unified. Delegated DHT inference is gone; no_mmap is load_mode, and batch translations return an array instead of a newline-joined string. @qvac/bare-sdk is no longer part of the lockstep pipeline.

Table of Contents

• Breaking Changes
  • Inference Is the In-Process Engine
  • Delegated Inference Removed
  • no_mmap Became load_mode
  • Batch Translations Return an Array
  • n_discarded Dropped
  • MiniMax Python Config Types
  • Language Detection Package Rename
  • e2e Test Suite Package Rename
• New APIs
  • assessModelFit
  • ABot-World Sessions
  • MiniMax Music Generation
  • Parakeet Unified Transcription
  • Hugging Face Download Checksums
  • Worker Startup Timeout
  • Tensor Split and Flash Attention
  • Injected TurboVec RAG Index
  • Config Schema Descriptions
• Features
• Bug Fixes
• Model Changes
  • Added Models

Breaking Changes

Inference Is the In-Process Engine

The SDK worker and Bare in-process path now sit on @qvac/inference. @qvac/bare-sdk is a thin re-export and is no longer versioned with the SDK. Import lifecycle helpers from @qvac/sdk/worker-lifecycle. The worker entry is dist/src/worker/index.js. ./commands and ./worker-core are gone; importing them throws with migration guidance. ./onnx-tts/plugin remains as an alias of the TTS plugin.

Before:

import { … } from "@qvac/sdk/worker-core";
const worker = "<sdk_root>/dist/server/worker.js";

After:

import { … } from "@qvac/sdk/worker-lifecycle";
const worker = "<sdk_root>/dist/src/worker/index.js";

@qvac/sdk now depends on @qvac/inference@^0.19.0. Install both at 0.19.0.

Delegated Inference Removed

Provider mode and DHT delegation are gone. Models load and run locally only.

Before:

await startQVACProvider({ firewall })
const id = await loadModel({ modelSrc, delegate: { providerPublicKey } })
await heartbeat({ delegate: { providerPublicKey } })
await stopQVACProvider()

After:

const id = await loadModel({ modelSrc })
await heartbeat()

Removed: startQVACProvider, stopQVACProvider, loadModel/heartbeat delegate options, hasActiveProviders on unload, isDelegated/providerInfo on loaded-model info, profiler origin / resourceOrigin, and the provider/delegate error classes. Python load_model(delegate=...) is also gone.

no_mmap Became load_mode

modelConfig.no_mmap is replaced by load_mode. Do not keep a boolean under the new key.

Before:

await loadModel({
  modelSrc: MODEL,
  modelType: 'llm',
  modelConfig: { ctx_size: 2048, no_mmap: true }
})

After:

await loadModel({
  modelSrc: MODEL,
  modelType: 'llm',
  modelConfig: { ctx_size: 2048, load_mode: 'none' }
})

The same mapping applies to deviceDefaults.llm and deviceDefaults["llamacpp-completion"]. load_mode also accepts "mlock", "mmap+mlock", and "dio".

Batch Translations Return an Array

A translate call with several strings used to join results with \n. It now returns translations: string[]. Streaming emits one whole translation per token, in input order.

Before:

const result = translate({
  modelId,
  text: ['Good morning', 'Good night'],
  stream: false
})
const translations = (await result.text).split('\n')

After:

const result = translate({
  modelId,
  text: ['Good morning', 'Good night'],
  stream: false
})
const translations = await result.translations

n_discarded Dropped

modelConfig.n_discarded is no longer accepted. Context overflow now reports requiredTokens, cachedTokens, promptTokens, and ctxSize on ContextOverflowError.

MiniMax Python Config Types

AudioGen load-config generated names are engine-specific. ACE-Step classes gained an Acestep infix; MiniMax has its own Minimax types. Update Python imports accordingly.

Language Detection Package Rename

@qvac/langdetect-text-cld2 is removed. Import from @qvac/langdetect-text. detectOne is synchronous.

e2e Test Suite Package Rename

SDK e2e depends on @qvac/test-suite instead of @qvac/qvac-test-suite. Update imports and Metro resolvers. The old package remains installable for 0.10.x pins.

New APIs

assessModelFit

assessModelFit estimates whether a set of models will fit before you download them. The same function is exported from @qvac/sdk and @qvac/inference.

import { assessModelFit, QWEN3_8B_INST_Q4_K_M } from '@qvac/sdk'

const result = await assessModelFit({
  models: [{ model: QWEN3_8B_INST_Q4_K_M, workload: { kind: 'llm', contextTokens: 8192 } }],
  execution: 'sequential',
  policy: 'interactive-v1'
})

result.verdict // "likely-fits" | "likely-too-large" | "unknown"
result.basis // "system-memory" | "process-memory" | "device-memory" | "device-budget"
result.budget?.availableBytes // headroom before the policy reserve

On a discrete GPU, basis is device-memory (Linux VRAM) or device-budget (Windows DXGI). Integrated GPUs stay on system-memory because they allocate from RAM. Multi-GPU machines require likely-fits on the smallest usable card and likely-too-large on the largest; in between the verdict is unknown. VM display adapters are not counted as GPUs. The reserve is 20% of budget.availableBytes, capped at 2 GiB on desktop and 1 GiB on mobile. iOS uses per-process memory and may return unknown when that metric is missing. Catalog resource profiles (getModelResourceProfile) back the estimator; an unknown checksum is undefined, not a guess.

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_VAE,
    t5XxlModelSrc: UMT5_XXL_ENC_Q8_0,
    vaeModelSrc: ABOT_WORLD_0_5B_LF_VAE_F16,
    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.

MiniMax Music Generation

AudioGen can load MiniMax (engine: "minimax") alongside ACE-Step.

const modelId = await loadModel({
  modelType: 'audiogen',
  modelConfig: {
    engine: 'minimax',
    lmModelSrc: '/models/mm3-lm-q8.gguf',
    synthModelSrc: '/models/mm3-synth-q8.gguf'
  }
})

const run = audioGen({
  modelId,
  caption: 'warm cinematic piano',
  maxFrames: 250,
  inferenceSteps: 12,
  cfgScale: 1.8
})

audioGen results now include diagnostics (selectedBackend, selectedDevice, optional fallback.reason when a GPU request landed on CPU). Progress total may be 0 for indeterminate stages.

Parakeet Unified Transcription

import { loadModel, transcribe, PARAKEET_UNIFIED_0_6B_Q8_0 } from '@qvac/sdk'

const modelId = await loadModel({
  modelSrc: PARAKEET_UNIFIED_0_6B_Q8_0,
  modelType: 'parakeet-transcription'
})
const text = await transcribe({ modelId, audioChunk: 'audio.wav' })

Hugging Face Download Checksums

Hugging Face HTTP downloads are verified against the Hub SHA-256. requireHttpChecksum and requireSecureTransport can also be set globally or per loadModel / downloadAsset call. Plain HTTP to a private origin is unchanged unless you opt in.

await loadModel({
  modelSrc: 'https://huggingface.co/org/repo/resolve/main/model.gguf',
  modelType: 'llamacpp-completion',
  requireHttpChecksum: true,
  requireSecureTransport: true
})

Worker Startup Timeout

rpcInitTimeoutMs (config file) and QVAC_RPC_INIT_TIMEOUT_MS control the worker handshake. A timeout whose worker already exited attaches WorkerStartupError as cause, with exitCode, exitSignal, and stderrTail, so you can tell a dead worker from a slow one.

Tensor Split and Flash Attention

llama.cpp loads accept split-mode: "tensor" and flash-attn: "on" in modelConfig.

Injected TurboVec RAG Index

Set ragTurbovec: true in qvac.config.json. Embedding plugins can supply a turbovecIndexProvider with create / load.

Config Schema Descriptions

Every modelConfig field now carries a description, exported from @qvac/sdk/schemas. configSchemaForModelType("whisper") (or "llm", "tts-ggml", "diffusion", …) returns that type's schema so tools such as qvac configure can document options without a per-addon list.

Features

Qwen3.8 tool calls go through the Qwen parser. Darwin-arm64 calibration uses a persistent-based fit with an audio guard. Desktop calibration for assessModelFit covers darwin-x64, linux-arm64, and win32-x64, including integrated GPUs; AMD linux stays unknown. @qvac/tts-ggml 0.8.0 can select CUDA on linux-x64 NVIDIA without a new backend key. @qvac/diffusion-cpp is ^0.21.0. @qvac/bci-whispercpp is 0.8.0.

Bug Fixes

A worker RPC init timeout now preserves the exit signal instead of dropping it. Expo prebuild refreshes iOS addon links. ContextOverflowError reports how many tokens the request needed versus the effective ctx_size / parallel ceiling.

Model Changes

This release adds Parakeet Unified 0.6B transcription constants and Qwen3.8 Flash Next 177B multimodal shards.

Added Models

MMPROJ_QWEN3_8_FLASH_NEXT_177B_MULTIMODAL_F16
PARAKEET_UNIFIED_0_6B_F16
PARAKEET_UNIFIED_0_6B_Q4_0
PARAKEET_UNIFIED_0_6B_Q8_0
QWEN3_8_FLASH_NEXT_177B_MULTIMODAL_UD_Q2_K_XL_SHARD
QWEN3_8_FLASH_NEXT_177B_MULTIMODAL_UD_Q4_K_XL_SHARD