The per-family model crates (flux, h3, paint, music, speech, stems,
vision, sfx, rife, trellis, beats, common) and the libs/diffusion
research binaries read some 190 environment variables that were
research knobs: tensor dumps, per-stage timing, oracle-parity and
fixture rigs, experiment toggles. The path taken with none of them set
is the one that ships; every such knob is deleted with the code it
gated, and every losing branch of an experiment toggle goes with its
toggle — dead kernels, fields and functions included. What remains are
the real configuration variables (the FLUX_*_MODE family, FLUX_GRAPH,
the VAE pool cap, the FLUX2 text-encoder residency, H3_VAE_BATCH, the
music3 caches and official modes, the stems/beats f16 switches, the
weight and data roots) and the build-script variables.
Rebased on the runtime cleanup: precision stays explicit everywhere
(GemmPrecision, f16_attention_operands, the H3 text precision, DA3's
StrictF32 in code); no act16, no H3_ACT_F16, no FLUX_ATTN_F16 or
FLUX_VAE_CONV_GEMM reads survive.
Reviewed by the delegate reviewer (APPLY, no findings) and gated on the
Windows CUDA box: all seventeen model crates check, motion and vision
tests, the hub and the diffusion bins — the gate caught one CUDA-only
tap marker the Mac never compiles, removed here. On this Mac: the same
checks plus motion 24, paint 158 and vision 23 tests.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Squashed from work; the fine-grained history is under tag archive/work-2026-08-26:
- vj: the tweener learns — RIFE runs on the Mac and feeds the warp
- asset-ai: the chat tells the truth while it works
- llm: a cold turn on the solo slot takes the session-native speculative path — think-mode turns no longer re-ingest the whole conversation through the draft head (66 → 122 tok/s on the four-lane box)
- llm: the step cost model is chosen per device — the RTX PRO 6000's measured verify curve (13.7 + 3.17·B ms) beside the 5090's; the bench warms every tail shape and times two windows
- vj: local store, lyrics and model plumbing, and the frame-interpolator's device parity check
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.
