Release ViT-L FP16 CUDA weights and document platform deployment choices

README.md

@@ -25,7 +25,10 @@ DINOv2 ViT-L/14 encoder and a square-query decoder. The model takes
 
 ## Files
 
-- `chessqueries-vitL14-644-joint.safetensors` — recommended, tensor-only
+- `chessqueries-vitl-644-safetensors-fp16-r1.safetensors` — CUDA/web deployment
+  weights (743.35 MB), FP16 weights/input/logits; SHA-256
+  `45edb69bf156f94299256fc51d1c75805f1de745ecfe34dc0615ef21265ab506`.
+- `chessqueries-vitL14-644-joint.safetensors` — FP32 source/reference, tensor-only
   inference weights; SHA-256
   `6151bdd98fbe25f32080c097eba7ae75808b5615160e4867240931085fc762e5`.
 - `chessqueries-vitL14-644-joint.ckpt` — original full PyTorch Lightning
@@ -36,6 +39,39 @@ DINOv2 ViT-L/14 encoder and a square-query decoder. The model takes
 - `chessquerieslite-vits-644-int8.onnx` — the same ViT-S/14 model,
   INT8 ONNX (36.03 MB).
 
+- `chessquerieslite-vits-644-coreml-fp16-r2.mlpackage.zip` — dedicated iOS
+  Lite Core ML package (62.47 MB), applied EXIF orientation and float32 tensor I/O;
+  SHA-256 `e01695f8adb0de7216bb1aae3a20ac2bb2f988a6c269a430181440e5d6f3292f`.
+
+## FP16 validation
+
+The ViT-L FP16 artifact was reloaded and compared with the released FP32 source
+on all 3,196 test boards: ChessReD 2,129, ChessCog 342, SLCC 373 and CVChess 352.
+All 204,544 square predictions matched; exact-board accuracies were unchanged
+(99.624%, 98.538%, 87.668% and 87.216%, respectively). Both models ran at batch
+one on an RTX 4090 with PyTorch 2.8.0, timm 1.0.27 and TF32 disabled. This
+validates this artifact/runtime combination, not every backend or future export.
+
+## Deployment selection
+
+CUDA/web serving selects ViT-L FP16 weights, targeting roughly 20 ms per warm
+batch-one model inference on an RTX 4090. Use an FP16-aware ChessQueries loader:
+older packages support only FP32. Preserve float32 RGB decode, bilinear
+antialiased resize and ImageNet normalization, then cast to FP16 immediately
+before inference. Safetensors serving ignores EXIF orientation. The dedicated
+iOS Core ML package applies orientation and has its own preprocessing contract.
+
+iOS selects the dedicated Core ML export. Device-specific goals are around
+100 ms on a powerful iPhone and 400–500 ms on a less powerful iPhone; validate
+the exact export on the target phone. Android runtime selection is deferred.
+The best existing Pixel 9 Pro result is about 629–633 ms with LiteRT FP16 CPU,
+but a common-workload CPU FP16/INT8 comparison and complete device validation
+remain necessary. These timing scopes differ and are not a cross-platform benchmark.
+
+New releases follow these deployment choices. FP32 source/reference checkpoints
+and existing published artifacts remain available; new FP32 companions are not
+required for every deployment release.
+
 Setup, inference, evaluation, and reproduction instructions are in the
 [code repository](https://github.com/JSeytre/chessqueries).
 

chessqueries-vitl-644-safetensors-fp16-r1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45edb69bf156f94299256fc51d1c75805f1de745ecfe34dc0615ef21265ab506
+size 743353530