The AI runtime crates read their precision, activation and kernel
choices from environment variables; the code path taken with none set
is the one that ships, so every knob that selected it is now an explicit
argument and every losing path is gone. GemmPrecision { f16_accumulate,
f16_activations } is passed by each caller: the default {true, true} is
the old unset Flux route; H3's DiT and text encoder pass {false, false}
(H3's >1e4 activations saturate f16 — the policy H3 used to set through
FLUX_GEMM_F16ACC=0 on itself), its VAE {true, false}; DA3's StrictF32
selects f32 packed attention in code; Hy-Motion carries an explicit
f16_attention_operands flag through its text refiner, its double and
single blocks and the CUDA backend (true in production, false only in
its full validator). The libs/diffusion bins — a separate workspace —
are migrated to the same shapes.
Benches and validators no longer set variables on themselves: llama's
skip-logits is a session option (the CUDA bench turns it on), OCR takes
explicit use_f16_gemm and tiled_roformer options, the lane speculative
probe reads its CLI. The live gates the first cut had deleted are back
as explicit-config tests: MMQ M=129, the strided-f32 MMV path, the
RMS+MUL CPU oracle. The loader's THREADS and CHUNK_MB stay real settings.
The Metal quantized-matmul experiment (metal_qmm and its vendored MLX
kernels) was reachable only through a knob and goes with it.
Reviewed in three rounds by the delegate reviewer (the last round
accepted everything but one Hy-Motion call site, fixed in round four
and reviewed here), and gated on the Windows CUDA box: lib checks of
common/paint/loader/cuda/llm/motion/vision, motion 24 and vision 23
tests, the hub check, the diffusion bins, llm 253 passed / 1 ignored.
On this Mac: the same checks plus the motion and vision tests.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ff62db3c5e