Troubleshooting

Check the Core version, target platform, model, and inputs, then reproduce with CPU. Error logs and the specific platform are usually more useful than a generic error message.

Symptom	Possible cause	What to check
Model fails to load	Incorrect path, missing file, or corrupt model	Check local file permissions and size; on the web inspect the model URL's request and cross-origin response.
JVM cannot load the LiteRT library	Missing library, incompatible architecture or ABI	Check the package and process architecture; set kmplitert.native.path if needed. See the JVM guide.
Native linking or startup fails	Link path, deployed libraries, or dependencies do not match	Check architecture, libraries and search paths, and Apple embedding and signing.
Web initialization fails	CDN unavailable, incorrect WASM path, or version mismatch	Inspect browser network requests; configure a self-hosted path before initialization if needed.
Tensor not found	Input/output name or signature mismatch	Use the model's names and note Native metadata and Web signature limitations.
Incorrect input or output	Type, shape, length, order, or quantization mismatch	Check the model specification and tensor guide; do not blindly reuse example data.
Output not initialized	Web output read before inference completes	Write inputs, await run(), then read outputs.
GPU/NPU unavailable or slow	Unsupported plugin, device, browser, or model	Inspect logs, verify CPU first, then test on the target device. See hardware acceleration.
Operation fails after close	Compiler or buffer already released	Adjust resource scope; do not reuse closed objects. See lifecycle.

If the problem persists, prepare a shareable minimal reproduction and complete error details before reporting an issue; do not upload sensitive models or data.