Refresh TorchVision FP32 and channelwise INT8 exports

Apply validated graph cleanup, replay the established quantization recipe, and document CPU/GPU/NPU checks. Preserve the existing dynamic and Google Tensor G5 variants.

README.md

@@ -44,12 +44,77 @@ num_params: 28589128
 
 `convnext_tiny_int8_channelwise.tflite`: Mixed INT8/FP32 with channelwise INT8 weights. LayerNorm remains FP32.
 
+## 2026-09-22 maintenance update
+
+This maintenance update removes ten redundant reshapes from the LayerNorm fallback bodies in the published FP32 graph. The source checkpoint and public FP32 input/output interface are unchanged. The channelwise INT8 file is regenerated from the refreshed FP32 export.
+
+| File | Size (bytes) | SHA-256 |
+|---|---:|---|
+| `convnext_tiny.tflite` | 114422656 | `17d67ff1731ea2a2f9726fc8a8170cd3168c7be2ce1bf36258db3add06316f18` |
+| `convnext_tiny_int8_channelwise.tflite` | 29859024 | `d22d0fd223a678d8af582cf54ae28377d08807922b548e65068e8d35eddffc53` |
+
+### Source and preprocessing
+
+TorchVision `ConvNeXt_Tiny_Weights.IMAGENET1K_V1` ([checkpoint](https://download.pytorch.org/models/convnext_tiny-983f1562.pth)).
+Convert to RGB, resize the short side to 236 with bilinear interpolation and antialiasing,
+then center crop to 224 × 224. Divide pixels by 255, subtract `[0.485, 0.456, 0.406]`,
+and divide by `[0.229, 0.224, 0.225]`. The input layout is NCHW, shape `[1, 3, 224, 224]`;
+output is ImageNet class scores with shape `[1, 1000]`.
+
+### Quantization
+
+`convnext_tiny_int8_channelwise.tflite`: Mixed INT8/FP32 with channelwise INT8 weights and INT8 input/output. LayerNorm remains FP32; this is not a fully integer model.
+The base recipe is `ai_edge_quantizer.recipe.static_wi8_ai8()`, with local scale/bias
+safeguards to avoid overflowing INT32 biases. Intermediate calibration ranges use an
+exponential moving average; the output logits use the global observed range.
+
+- Calibration images: 500 (manifest image IDs: 0–499).
+- Calibration manifest SHA-256: `6a8d04c842352ef41d7795e3ec6acd30a4cbad520466d69b8b0503a521d5d1e5`.
+- Base quantization recipe SHA-256: `3db5fac189c1707c9a0c3a18fecaa8fabd8c9ef3681af553d9e0c86fc18ec0fb`.
+
+For INT8 input, apply the preprocessing above, then compute
+`clip(round(x / input_scale) + input_zero_point, -128, 127)` and cast to INT8,
+using nearest rounding with ties away from zero, as in this validation.
+Dequantize output scores with `(q - output_zero_point) * output_scale`.
+Read scales and zero points from the selected model file; do not assume that they
+match an earlier revision. The existing FP32 usage example, where present, requires
+this additional I/O handling before it can be used with the INT8 variant.
+
+### Validation of the refreshed artifacts
+
+CPU evaluation used 1000 images from ImageNet-1k validation (manifest image IDs: 1000–1999). These are subset measurements, not a new full-validation accuracy result.
+
+| Variant | CPU backend | Top-1 correct | Top-5 correct |
+|---|---|---:|---:|
+| FP32 | xnnpack | 847 | 958 |
+| FP32 | builtin | 847 | 958 |
+| Mixed INT8/FP32 | xnnpack | 837 | 958 |
+| Mixed INT8/FP32 | builtin | 833 | 956 |
+
+For both variants on both CPU backends, all 1,000 output vectors are bit-identical to the preceding published revision. Public input/output metadata, quantization parameters and signatures are unchanged.
+
+The model-index values of 82.46% top-1 and 96.13% top-5 in this card's metadata are historical results retained from the earlier publication. They were not remeasured over the full validation set for this maintenance release.
+
+Validation used LiteRT 2.3.0 on host CPU. The 1,000-image subset is disjoint from the 500 calibration images, but was used during earlier recipe development; it is a regression set, not an untouched test set.
+
+Saved calibration arrays were restored as float32. Replaying the same recipe and INT32-bias safeguards on the previous source exports reproduced both previous quantized files byte for byte.
+
+FP32 GPU validation used 10 images on TECNO LJ9 / MT6897 with both default and explicit FP32 precision. Explicit FP32 relative L2 error versus host CPU was 1.77091e-06, with no top-1 changes. This release makes no new latency claim.
+
+Quantized MT6897 NPU execution used 10 images: 0 top-1 changes versus the quantized host CPU reference, relative L2 error 0.188453. The NPU output scores are bit-identical to the retained checks of the preceding published quantized file; the CPU/NPU numerical difference is unchanged. SM8750 and MT6993 coverage is host AOT compilation only; corresponding devices were unavailable.
+
+The existing `convnext_tiny_dynamic_wi8_afp32.tflite` and
+`convnext_tiny_Google_Tensor_G5_apply_plugin.tflite` files are retained unchanged.
+The validation results in this update apply only to the two files listed in the artifact table.
+
 ## Compatibility
 
 | File | CPU | GPU | NPU |
 |---|---|---|---|
-| `convnext_tiny.tflite` | Supported | Supported | N/A |
-| `convnext_tiny_int8_channelwise.tflite` | Supported | Not supported | Qualcomm / MediaTek |
+| `convnext_tiny.tflite` | Validated: builtin and XNNPACK, 1,000 images | Validated on TECNO LJ9 / MT6897: full delegation, 10 images | Not evaluated for this FP32 artifact |
+| `convnext_tiny_int8_channelwise.tflite` | Validated: builtin and XNNPACK, 1,000 images | Not supported on the tested MLDrift stack; use FP32 | Full AOT: MT6897, MT6993, SM8750; MT6897 device smoke passed |
+
+Compatibility entries above describe the 2026-09-22 artifacts. Earlier releases had FP32 GPU smoke checks and channelwise INT8 Qualcomm/MediaTek AOT compilation checks; those historical checks do not certify changed files. AOT compilation coverage alone does not establish on-device numerical accuracy. Earlier INT8 runs on the tested LiteRT MLDrift GPU stack produced large numerical errors; use the FP32 variant for GPU when its current validation supports that backend.
 
 ## Intended uses & limitations
 

convnext_tiny.tflite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42f4181cae412cd35fc769d4baa25d13d5e38a399fe37b0f8872b5b849571188
-size 114425184
+oid sha256:17d67ff1731ea2a2f9726fc8a8170cd3168c7be2ce1bf36258db3add06316f18
+size 114422656

convnext_tiny_int8_channelwise.tflite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:95e9aac495675b7e73cc91a7dfcb3b8474b8c06ae53e0815ee290c6cad9a25c7
-size 29861536
+oid sha256:d22d0fd223a678d8af582cf54ae28377d08807922b548e65068e8d35eddffc53
+size 29859024