← ALL TAGS

Tag: metal

5 articles tagged “metal” on Oak Oliver Engineering, newest first. The latest is “A Neuron Atlas for a Speech Model: What a Local 4B Can Label, What Replays Bit for Bit, and the Voice Levers We Had to Re-check”.

Engineering

A Neuron Atlas for a Speech Model: What a Local 4B Can Label, What Replays Bit for Bit, and the Voice Levers We Had to Re-check

A one-day feasibility probe on LFM3-Audio-2B, a 2B speech-to-speech model: a local 4B model wrote labels for 300 of its 184,320 MLP neurons, a simulator scored them against a shuffled baseline, steering checked whether they are causal, and every test replayed bit for bit in new processes and across Metal and CUDA. Labels work for selective neurons only, families built from descriptions do not, labels transfer to a text model with another tokenizer, and of the voice neurons we first reported, three survived a 16-seed rerun against random neurons.

October 30, 2026 16 min read
Read Story →
Engineering

Memory in the Weights: Teaching a 4B Model New Facts While It Sleeps, Without Teaching It to Lie

A lab note on a wake–sleep memory for a small language model: a teacher distils the day's messages at night and the facts are trained into per-person experts behind a learned router that can stay shut. Dense memory hallucinated; routed memory kept the model's own abstention. With Qwen3-VL-4B the memory held 100 episodes at best; the first Spark-X2.6-4B seed still passes at 250, and the three-seed answer is not in yet. All of it runs on our Rust engine, on Metal and CUDA.

October 23, 2026 15 min read
Read Story →
Engineering

Sounds and Videos as Prompts: A 107M Adapter That Lets EmbeddingGemma 2 Drive Qwen-Image and Iris, Retrained on Permissive Data

We replaced the 8B text encoder of Qwen-Image 2.1 with the 0.74B EmbeddingGemma 2 plus a small adapter, so a sound or a short video can be referenced in a prompt like an image. Then we threw away every non-commercial training item and retrained on CC0 / CC BY data only. The permissive adapter is better on text, the best we have on video, and the licensing surprise was not in the data.

October 21, 2026 11 min read
Read Story →
Engineering

Porting TRELLIS.2 From Python to Rust and Metal: 2D Sprites to 3D Models in One Command

I moved Microsoft's 4-billion-parameter image-to-3D pipeline off PyTorch and onto our Rust and Metal runtime. One command now turns a 2D sprite into a textured GLB in one to three minutes with about 6 GB of memory instead of 33. This is how the port was organised, the three bugs worth retelling, the sparse-convolution kernel that decided it, and what parity means when the sampler is chaotic.

October 5, 2026 12 min read
Read Story →
Engineering

Porting ERNIE-Image-Turbo to Zig and Metal: An 8B Diffusion Transformer From Scratch on Apple Silicon

I rewrote the hot path of Baidu's 8B-parameter ERNIE-Image-Turbo (DiT + scheduler) from scratch in Zig with hand-written Metal kernels. 13.06s/step at 1024×1024 on M2 Max. The bug archaeology behind getting 36 layers numerically correct, plus the honest accounting of what still has Python in it.

April 23, 2026 18 min read
Read Story →