nx_sampler.nx

nx_sampler.nx -- DPM++ 2M scheduler (substrate-native, no CUDA). Ships VRAM-track B-001 per docs/VRAM_OPTIMIZATION_REALISTIC_TRACKING.md: DPM++ 2M Karras at 8 steps replaces Euler at 12 steps in the Z-Image inference loop. Same or slightly better quality at ~30-40% lower latency. Working memory savings come from fewer kept-around step buffers; VRAM-bytes-per-step is unchanged but step-count drops. Per the sovereign-from-bits-up cardinal: pure NishiLang, no PyTorch / diffusers / sdcpp dependency. Composes against nx_tensor envelopes and the caller's i64+Q10 model evaluation. ===== Math reference ============================================= DPM-Solver++ second-order multi-step (Lu et al. 2022, arXiv:2211.01095, "DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models"). In σ-only parameterisation (α_t = 1 schedule, "model_x" prediction): λ_t = -log(σ_t) h_i = λ_{i+1} - λ_i = log(σ_i / σ_{i+1}) r_i = h_{i-1} / h_i D_i = model_x(x_i, σ_i) -- caller-supplied D'_i = (1 + 1/(2r_i)) D_i - (1/(2r_i)) D_{i-1} x_{i+1} = (σ_{i+1}/σ_i) x_i + (1 - σ_{i+1}/σ_i) D'_i (2nd order) First step has no D_{i-1}, so D'_0 := D_0: x_1 = (σ_1/σ_0) x_0 + (1 - σ_1/σ_0) D_0 (Euler) ===== Log-uniform σ spacing collapses r_i to 1 ================== If σ_i is geometric (σ_{i+1} = k σ_i for a constant ratio k < 1) then log(σ_i/σ_{i+1}) = -log k is constant, so all h_i are equal and r_i = 1 for every step. The 2nd-order update simplifies: D'_i = (1 + 1/2) D_i - (1/2) D_{i-1} = 1.5 D_i - 0.5 D_{i-1} And σ_{i+1}/σ_i = k everywhere, so the linear-interp coefficient