Remove MLX artifacts: 7.5-19.6x slower than Core ML on device

MLX measured substantially slower than Core ML on an iPhone 16 for every convolutional vision model tested (efficientnet-v2-s 19.6x, yolo26-s 16.0x, yolo26-seg-s 7.5x, yolo26-pose-n 10.0x). MLX weight quantization also cannot shrink a conv model, since it only rewrites nn.Linear/embedding modules.

The .pte and config.json are removed; each mlx/ dir keeps a NOTES.md with the measurements. The exporter is retained in export-scripts so this can be re-tested against a future ExecuTorch/MLX release.

Earlier tags (v0.9.0 and below) are untouched and still contain the MLX artifacts.

mlx/NOTES.md

@@ -0,0 +1,50 @@
+# MLX is not shipped for this model
+
+The MLX artifacts for this model were removed deliberately. The exporter in
+`export-scripts` is kept so the decision can be re-tested against a future
+ExecuTorch/MLX release, but no `.pte` or `config.json` is published here.
+
+## Why
+
+Measured on device (iPhone 16, ExecuTorch 1.4.1, Release build, 3 warmup runs +
+15 timed runs, median):
+
+| backend | precision | median latency |
+|---|---|---|
+| Core ML | fp16 | **2.4 ms** |
+| MLX | fp16 | 47.0 ms |
+
+MLX is **19.6x slower than Core ML** for EfficientNet-V2-S @384.
+
+This is not specific to this model. MLX was measured behind Core ML on every
+convolutional vision model tested, by 7.5x to 19.6x:
+
+| model | Core ML | MLX | MLX penalty |
+|---|---|---|---|
+| efficientnet-v2-s @384 | 2.4 ms | 47.0 ms | 19.6x |
+| yolo26-s @640 | 3.7 ms | 59.2 ms | 16.0x |
+| yolo26-seg-s @640 | 11.0 ms | 82.9 ms | 7.5x |
+| yolo26-pose-n @384 | 1.2 ms | 12.0 ms | 10.0x |
+
+Two things compound the latency result:
+
+1. **MLX quantization cannot shrink a conv model.**
+   `executorch.backends.mlx.llm.quantization.quantize_model_` only rewrites
+   `nn.Linear` and embedding modules. EfficientNet-V2-S holds just 6.0% of its
+   parameters (1,281,000 of 21,458,488) in its single `nn.Linear`, so 4-bit
+   weight quantization buys about 5% file size. `nvfp4` is a literal no-op: it
+   logs `Skipping classifier.1: weight shape [1000,1280] incompatible with
+   nvfp4 (group_size=16)` and emits a byte-identical artifact.
+2. **Core ML quantizes the whole model.** Core ML's
+   `op_linear_quantizer_config` reaches the convolutions, so int8/per_channel
+   halves the artifact again (43.9 MB fp16 -> 22.1 MB) at identical latency.
+
+## Scope
+
+This applies to convolution-dominated vision models. It does **not** generalize:
+
+- `rfdetr-nano` (conv + transformer) measured only **1.24x** behind Core ML.
+- The `lfm2.5` text encoders measured **2.7-3.1x faster on MLX** than XNNPACK.
+
+MLX suits matmul-heavy graphs. The right backend depends on model class, so
+this file is not a statement about MLX in general.

mlx/config.json

@@ -1,42 +0,0 @@
-{
-  "$schema": "https://huggingface.co/software-mansion/react-native-executorch-spec/resolve/main/config.schema.json",
-  "model": "efficientnet_v2_s",
-  "family": "efficientnet",
-  "capabilities": [
-    "classification"
-  ],
-  "backend": "mlx",
-  "license": "bsd-3-clause",
-  "variants": [
-    {
-      "file": "efficientnet_v2_s_mlx_fp16.pte",
-      "precision": "fp16",
-      "quantized": false,
-      "default": true,
-      "methods": {
-        "forward": {
-          "inputs": [
-            {
-              "shape": [
-                1,
-                3,
-                384,
-                384
-              ],
-              "dtype": "float32"
-            }
-          ],
-          "outputs": [
-            {
-              "shape": [
-                1,
-                1000
-              ],
-              "dtype": "float32"
-            }
-          ]
-        }
-      }
-    }
-  ]
-}

mlx/efficientnet_v2_s_mlx_fp16.pte

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:fc41b9a923b89b27f95efc43d59c9651a3fe713619f61dd1d870bdd0713d7faf
-size 43509376