FluidInference
/

gliner2-5-base-coreml

@@ -43,4 +43,6 @@ The original L128/K8 classification packages remain available, with a separate `
 An optional **embedding-only W8** classifier keeps the encoder and heads FP16. It reduces the package from **388,981,604 to 291,053,994 bytes** (25.2%). On an M5 Pro plugged into AC, it agreed with FP16 on all 14 bucket-fitting requests in a fixed 20-request manifest (largest probability difference 0.00992), and matched the native classifier on 100/100 selected eligible application requests (largest confidence difference 0.01370). The 14-request paired full-request p50 was 3.662 ms FP16 versus 3.632 ms W8 with automatic device selection; this is a size option, not a demonstrated speedup. Forced CPU+ANE was slower (9.168 versus 9.666 ms). The W8 compute plan put 496/516 executable ops on ANE under forced CPU+ANE, with 20 CPU boundary/constant ops. Use `runtime.py --precision embedding_w8` for this classification package; extraction continues to use its separate FP32 or FP16 packages.
 The Core ML deployment target is iOS 17/macOS 14. Conversion scripts, pinned dependencies, asset hashes, and selected verification reports are included. The source revision is a current pinned snapshot; identity with the historical Decision Index evaluation checkpoint has not been established.

 An optional **embedding-only W8** classifier keeps the encoder and heads FP16. It reduces the package from **388,981,604 to 291,053,994 bytes** (25.2%). On an M5 Pro plugged into AC, it agreed with FP16 on all 14 bucket-fitting requests in a fixed 20-request manifest (largest probability difference 0.00992), and matched the native classifier on 100/100 selected eligible application requests (largest confidence difference 0.01370). The 14-request paired full-request p50 was 3.662 ms FP16 versus 3.632 ms W8 with automatic device selection; this is a size option, not a demonstrated speedup. Forced CPU+ANE was slower (9.168 versus 9.666 ms). The W8 compute plan put 496/516 executable ops on ANE under forced CPU+ANE, with 20 CPU boundary/constant ops. Use `runtime.py --precision embedding_w8` for this classification package; extraction continues to use its separate FP32 or FP16 packages.
+When downloading with `huggingface_hub.snapshot_download`, pass `local_dir="./gliner2-base-coreml"` and point the runtime there. Core ML compilation on the tested macOS release rejected the symlinked weight file in the default Hub cache snapshot; a materialized local directory passed the downloaded-package smoke test.
 The Core ML deployment target is iOS 17/macOS 14. Conversion scripts, pinned dependencies, asset hashes, and selected verification reports are included. The source revision is a current pinned snapshot; identity with the historical Decision Index evaluation checkpoint has not been established.