build(deps): bump @qvac/sdk from 0.14.0 to 0.16.0

[dependabot]

Bumps @qvac/sdk from 0.14.0 to 0.16.0.

Release notes

Sourced from @​qvac/sdk's releases.

QVAC Embed Addon v0.16.0

Changed

• qvac-fabric >= 8189.0.2: Mali/Adreno F16 coopmat1 NaN fix, Vulkan VMA migration, accumulated upstream fixes since 7248.x.
• OpenCL backends default flash-attn=off (not reliably supported on OpenCL); user flash-attn/flash_attn overrides are honored.

Internals (no behaviour change)

• ABI port: common_init_result → common_init_result_ptr.

QVAC SDK v0.16.0

📦 NPM: https://www.npmjs.com/package/@​qvac/sdk/v/0.16.0

QVAC SDK 0.16.0 adds Ideogram 4 image generation, per-phase timing stats for the diffusion pipeline, and first-class GR00T support in the VLA SDK. It includes a full-fidelity Python SDK client (not yet published) that mirrors the TypeScript surface, reuses the KV cache across turns in the OpenAI-compatible serve endpoints for faster multi-turn chat, and refreshes the TTS and transcription model registry with new Parler-TTS voices and multimodal LLMs.

Table of Contents

• New APIs
  • Ideogram 4 Diffusion
  • Per-Phase Diffusion Timing Stats
  • Python SDK Client
  • GR00T Support in the VLA SDK
• Features
  • KV Cache Reuse Across Turns in OpenAI-Compatible Endpoints
• Bug Fixes
• Model Changes
  • Added Models
  • Removed Models

New APIs

Ideogram 4 Diffusion

Image generation now supports Ideogram 4 through the sdcpp-generation model type. Load the diffusion model together with its companion LLM, VAE, and unconditional models, then generate as with other diffusion backends.

const modelId = await loadModel({
  modelSrc: IDEOGRAM_MODEL_URL,
  modelType: 'sdcpp-generation',
  modelConfig: {
    llmModelSrc: QWEN3_VL_MODEL_URL,
    vaeModelSrc: VAE_MODEL_URL,
    uncondModelSrc: IDEOGRAM_UNCOND_MODEL_URL
  }
})

Per-Phase Diffusion Timing Stats

... (truncated)

Changelog

Sourced from @​qvac/sdk's changelog.

[0.16.0]

📦 NPM: https://www.npmjs.com/package/@​qvac/sdk/v/0.16.0

QVAC SDK 0.16.0 adds Ideogram 4 image generation, per-phase timing stats for the diffusion pipeline, and first-class GR00T support in the VLA SDK. It includes a full-fidelity Python SDK client (not yet published) that mirrors the TypeScript surface, reuses the KV cache across turns in the OpenAI-compatible serve endpoints for faster multi-turn chat, and refreshes the TTS and transcription model registry with new Parler-TTS voices and multimodal LLMs.

New APIs

Ideogram 4 Diffusion

Image generation now supports Ideogram 4 through the sdcpp-generation model type. Load the diffusion model together with its companion LLM, VAE, and unconditional models, then generate as with other diffusion backends.

const modelId = await loadModel({
  modelSrc: IDEOGRAM_MODEL_URL,
  modelType: 'sdcpp-generation',
  modelConfig: {
    llmModelSrc: QWEN3_VL_MODEL_URL,
    vaeModelSrc: VAE_MODEL_URL,
    uncondModelSrc: IDEOGRAM_UNCOND_MODEL_URL
  }
})

Per-Phase Diffusion Timing Stats

Diffusion results now include a per-phase timing breakdown, so you can see where generation time goes — conditioning, the denoising loop, VAE decode, post-processing, and overall throughput.

const { stats } = await result
console.log(stats.conditionerMs) // prompt conditioning time (ms)
console.log(stats.denoiseMs) // denoising loop time (ms)
console.log(stats.vaeMs) // VAE decode time (ms)
console.log(stats.postProcessMs) // encode/upscale/mux time (ms)
console.log(stats.stepsPerSecond) // denoising throughput (steps/s)

Python SDK Client

Note: The Python client has landed in the SDK source but is not yet published as an installable package. The example below previews the upcoming client; it is not installable yet.

A full-fidelity Python client mirrors the TypeScript SDK surface, with an async Client, load_model, and a streaming completion that yields events and resolves a final result.

from tetherto.qvac_sdk import Client, load_model, completion
from tetherto.qvac_sdk.models import LLAMA_3_2_1B_INST_Q4_0
async with Client() as client:
transport = client.transport
model_id = await load_model(transport, model_src=LLAMA_3_2_1B_INST_Q4_0)
</tr></table>

... (truncated)

Commits

• a6ac843 QVAC-22171 chore: release @​qvac/sdk 0.16.0 (#3403)
• 61453d9 QVAC-21303 infra: simplify SDK e2e target selection (#3391)
• 21c6550 QVAC-22687 doc: add GR00T and imageInputMode/patches to VLA docs page (#3407)
• 2234534 QVAC-21691 fix[api]: export TranslationFailedError and harden Python connect ...
• 043e55a QVAC-22514 fix: make registry GGUF OCR models load via documented ocr() path ...
• 3ea98c1 doc: clarify GR00T continuous-state and required noise in VLA docs (#3392)
• d0cff00 QVAC-21980 test: add GR00T VLA usage example + desktop e2e (#3395)
• 9f479e3 QVAC-22636 feat[mod]: expose latest tts-ggml APIs and refresh models (#3393)
• 2190f31 QVAC-21980 feat[api]: first-class GR00T exposure in VLA SDK (registry + hpara...
• 9e1231d QVAC-19547 test[skiplog]: add strict Snap SDK e2e coverage (#3290)
• Additional commits viewable in compare view

[edycutjong]

@dependabot rebase