[sandlbn]

What

• FuseToConv also folds the 2D im2col chain that ggml_conv_2d exports (Pad + ExtractImagePatches + Transpose/Reshape) into one Convolution. Before, it matched only the 1D chain. GGML_OPENVINO_FUSE_CONV2D=0 keeps the old graph for A/B.
• GGML_OPENVINO_NPU_PROPS="K=V,..." passes extra properties into the NPU compile config, for example NPU_TURBO=YES.

Why

• The NPU runs ExtractImagePatches on its SHAVE cores: 349.9 ms of a 382.8 ms SmolVLA SigLIP view at 512 px, by the per-op profile. The same tower built natively with a Convolution runs in 32.9 ms. SDPA fusion, f16 weights and the execution-mode, performance-hint and dynamic-quantization settings were each tested on their own; none moved the tower by more than ~5%.
• The im2col cost grows with resolution: ~350 ms per view at 512 px, ~30 at 448, ~4 at 224. GR00T N1.6, VLA-JEPA and TurboVLA use a linear patch embed and do not change.
• The NPU drops its clock between requests. With a server's 50–500 ms gaps that costs 50–70 ms per request, and NPU_TURBO=YES prevents it.

Tests

• Latency, vla-bench p50, warmup 3, reps 20, back to back, before → after:

  Model	Before	After
  SmolVLA (2 × 512 px)	926 ms	177 ms (vision 810 → 63)
  Evo-1 (3 × 448 px)	572 ms	482 ms (vision 354 → 263)
  π0.5	500 ms	486 ms
  GR00T N1.5	263 ms	255 ms
  GR00T N1.6	241 ms	241 ms
  VLA-JEPA	161 ms	161 ms
  TurboVLA	27 ms	27 ms

• NPU_TURBO=YES with 200 ms gaps (vla-bench --gap-ms 200, PR 4): SmolVLA 247 → 178 ms, π0.5 549 → 487 ms, GR00T N1.6 292 → 241 ms. No change back to back, and no change in numerics.

• Correctness, max action error vs CPU f32, two fresh processes per model, all finite:

  Model	Error
  π0.5	0.0015
  SmolVLA	0.0021
  GR00T N1.6	0.0023
  TurboVLA	0.0032
  Evo-1	0.0065
  VLA-JEPA	0.0068
  GR00T N1.5	0.0176 (0.0198 with the fusion off)

  On the CPU plugin, the fused graph is within 3e-5 of the im2col graph.

• LIBERO-object on the NPU, PRs 1 + 2 + 3, NPU_TURBO=YES, median server time per request:

  Model	Success	Median per request
  SmolVLA	89/100	182 ms (was 90/100 at 981 ms)
  Evo-1	98/100	481 ms (was 97/100 at 648 ms)
  GR00T N1.5	95/100	383 ms (was 94/100 at 445 ms)
  π0.5, GR00T N1.6, VLA-JEPA, TurboVLA	in progress

• Patch script: applies to a fresh llama.cpp checkout and is idempotent.