gwenn-ha-dev/Siliconed v0.14

Siliconed 0.14

on GitHub

The Neural Engine

• The Mac's Neural Engine joins the GPU in the image model's matrix products, where this Mac's
  measure finds the pair faster than the GPU alone. Beside the CPU's matrix unit, it takes a
  product's shape wherever it keeps 90 % of that unit's speed, for the energy it saves. It computes
  in fp16: each weight column and each activation row is lifted by a power of two into fp16's
  range, then brought back in fp32, and rows that overflow go back to the GPU. The text encoders
  and the VAE never use it.
• Your choice, per model: the fp32 tag of the rack's DiT module switches it. Off by
  default on the Standard, which stays the publisher's weights computed in fp32, its image checked
  against the reference; on by default on Compact and Light, whose 8- and 4-bit weights already
  make another image. The tag reads fp32 · ANE when it is on, and an image computed with it
  carries neural-engine in its metadata. Where the measure left the Neural Engine off, there is
  nothing to choose.
• On an M1 Pro, 1024², Standard, Faster Rendering: 16 to 26 % less energy per image, and as
  fast or faster — Krea 2 135.4–137.4 s instead of 144.6–144.9 s, Z-Image 83.4–84.0 s instead of
  86.8–93.9 s, Qwen-Image-2.1 99.9–100.2 s instead of 103.6 s. The image moves by a tenth of a
  level on average on Z-Image and Qwen-Image-2.1; Krea 2 keeps its scene, face and light, and its
  finest details move. Every figure: The Neural Engine.
• Eight bits against eight bits: Z-Image Compact at 1024², with Faster Rendering and the Neural
  Engine (both by default), renders in 81.1–83.2 s; Draw Things' default 8-bit q8p, 135.0–144.2 s
  on the same Mac (Next to Draw Things).
• This Mac is measured again before the first render after the update, about a minute, as after
  every version: the measure now includes the Neural Engine, on each model's own products.
• The remote control: silicontrol add --ane or --no-ane, like a click on the tag;
  neural_engine in models, in jobs and in images.

Open a Hugging Face model in the app

• The app opens siliconed://hf links, relayed by a page of the site that offers the .dmg
  when the app is not installed: what Hugging Face's « Use this model » button opens, once Hugging
  Face lists Siliconed. A family's own repository opens its installation; any other repository
  lists the .safetensors and .gguf files the app could import — each judged on its header
  before anything downloads —, with their size and license. The file chosen is downloaded, verified,
  then imported like a dropped file. Nothing in the link is trusted.
• More LoRAs import: an integer alpha (as sd-scripts saves it) is read instead of refused; a
  LoRA that touches only the attention finds its family from the installed models' shapes, when
  only one takes it all; a LyCORIS file or a text-encoder LoRA is refused with its own message
  instead of being taken for something else.

Fewer downloads

• A text encoder or a VAE you already have is taken instead of downloaded. Dropped on the app,
  a file equal in size and SHA-256 to one an installation fetches — the 8-bit encoders of Z-Image,
  Qwen-Image-2.1 and Krea 2, Krea 2's bf16 encoder, the Standard encoders, every family's VAE — is
  copied where the download would have put it, then verified like a download.

Longer prompts

• Every model reads up to 1,024 tokens of prompt, and refuses a longer one with a message. Until
  0.13, Z-Image, Anima and Krea 2 silently ignored everything past 512 tokens, as the reference
  pipelines do by default, and Qwen-Image-2.1 refused past 512. A prompt under 512 tokens gives the
  same image as before.

Fixed

• A LoRA trained with a different rank per layer showed « rank 0 » in the Models window and in
  the rack's tooltip; it now shows its range, « ranks 2–57 » for inCase on Anima.