PP-OCRv6: XNNPACK detect is now static int8

Previously fp32 on the belief that static-activation int8 was unstable across
dynamic input sizes. Re-measured: correlation 0.956-0.994 against eager fp32
from 640 to 1280 square, and on-device the int8 build returns the same
detections ~15% faster end to end at 23.9 MB instead of 31.1 MB. The SVTR
recognizer stays fp32.

.gitattributes

@@ -39,3 +39,4 @@ PP-OCRv6_xnnpack.pte filter=lfs diff=lfs merge=lfs -text
 coreml/pp_ocrv6_coreml_int8.pte filter=lfs diff=lfs merge=lfs -text
 vulkan/pp_ocrv6_vulkan_fp16.pte filter=lfs diff=lfs merge=lfs -text
 xnnpack/pp_ocrv6_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
+xnnpack/pp_ocrv6_xnnpack_int8.pte filter=lfs diff=lfs merge=lfs -text

README.md

@@ -62,7 +62,7 @@ whatever width it picks instead of inferring a ratio.
 
 | backend | target | detect | recognize | warm latency (detect @960² / recognize) |
 |---|---|---|---|---|
-| `xnnpack` | CPU | fp32, true-dynamic | fp32, true-dynamic | ~574 ms / ~28 ms (Galaxy S24) |
+| `xnnpack` | CPU | static int8, true-dynamic | fp32, true-dynamic | ~28 ms recognize (Galaxy S24) |
 | `coreml` | Apple ANE | weight-only int8, enumerated | weight-only int8, enumerated | ~12–15 ms / ~2 ms (Apple M-series ANE) |
 | `vulkan` | Android GPU | fp16, true-dynamic (resize) | fp32 on **XNNPACK** (mixed-delegate) | ~73 ms / ~27 ms (Galaxy S24, Xclipse 940) |
 
@@ -70,9 +70,12 @@ whatever width it picks instead of inferring a ratio.
 > (XNNPACK) — the 18 709-token vocab head is not Vulkan-safe, and int8 SVTR is lossy, so the
 > recognizer stays fp32 on CPU for correctness.
 
-> **Why fp32 on CPU?** Static-activation int8 quantization is not stable across dynamic input
-> sizes for the detector (measured broken at ≥960px — including in the old per-bucket builds);
-> fp32 is bit-exact at every shape.
+> **Why int8 detect but fp32 recognize?** The detector is static-activation int8, calibrated on
+> real pages at 1280. It was previously shipped as fp32 on the belief that int8 was unstable
+> across dynamic input sizes; re-measured in 2026-08 that no longer holds (correlation 0.956 to
+> 0.994 against eager fp32 from 640 to 1280 square, best at 1280), and on-device the int8 build
+> returns the same detections about 15% faster end to end at 23.9 MB instead of 31.1 MB. The
+> SVTR recognizer stays fp32: int8 is lossy on its attention stack.
 
 ## CoreML notes (iOS)
 

xnnpack/config.json

@@ -10,11 +10,11 @@
   "license": "apache-2.0",
   "variants": [
     {
-      "file": "pp_ocrv6_xnnpack_fp32.pte",
-      "precision": "fp32",
-      "quantized": false,
+      "file": "pp_ocrv6_xnnpack_int8.pte",
+      "precision": "int8",
+      "quantized": true,
       "default": true,
-      "size_bytes": 31066200,
+      "size_bytes": 23890904,
       "methods": {
         "detect": {
           "inputs": [

xnnpack/pp_ocrv6_xnnpack_int8.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:726d414858606c890e3909ce89d639e6bc70898b155f7042722fd13c3c2ea913
+size 23890904