The Metal tensor backend is host-Vec based: every op copied its inputs up
and its result back, which put the DINOv3 backbone at 2.1 s a frame on an
M3 Max. try_vit_backbone_resident_f32 runs a whole pre-norm ViT stack
(LayerNorm, q/k/v, rotate-half rope from tables, flash attention, out,
LayerNorm, SwiGLU, down; residuals in place) inside one command buffer
against cached bf16 weights, with a new kernel_makepad_rope_half_tables_f32
kernel for the rope. The precompiled metallib now carries the bf16 GEMM
kernels the runtime source compile already enabled on bfloat devices.
The common backend exposes it as gpu_vit_backbone_resident (CUDA declines;
its per-op path is already resident) and the body backbone tries it first:
233 ms a frame on the M3 Max, same oracle parity (kp3d 1.5 mm).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The model code was spread across eight crates that had grown into each other:
ggml and cuda and mlx each owned part of a tensor runtime, llama and tts and
voice2 each owned part of a model, and libs/diffusion owned everything else.
They are now one tree with an explicit shape:
libs/ai/cuda — kernels and launch surface
libs/ai/metal — Metal shaders and the shim
libs/ai/llm — the language-model runtime (sessions, lanes, contexts,
the CUDA and Metal executors, the compiled Metal path)
libs/ai/models/ — common, flux, h3, music, paint, speech, stems, vision
libs/diffusion is not deleted but demoted: what remains is the VALIDATOR
crate — several dozen `*_validate.rs` oracles that check a native
implementation against a reference, which is where they belong now that the
implementations live next door.
The functional work inside the move is mostly in the LLM runtime: N lanes that
draft while one verify batch serves all of them, per-slot prefill over a shared
folded attention arena, speculation that survives batching, and a scheduler
that reports rather than publishes. And in the CUDA build: a machine without
usable CUDA must still LINK (and say so), the default kernel arch is the
building machine's GPU, `NO_CUDA` forces the stub even where the toolkit
exists, and kernels compile in parallel with progress.
libs/video_flow is new here: classical optical flow estimation and the `mkfl`
motion-field payload — a flow field measured from a clip without a model,
which is what drives free-rate bounce-looping playback and the uprez/tween
enhance pipe.
