QHexRT/NPU generate fails with InvalidArgument when prompt exceeds max_ctx (LFM2.5-230M: 599 > 512)

Summary

During device E2E testing (physical Snapdragon dev unit, serial 8977b1dd, Canoe_for_arm64, Android 16) after downloading and loading LFM2.5 230M on the QHexRT/NPU backend, a generate call failed natively with InvalidArgument because the assembled prompt (599 tokens) exceeded the model's max_ctx (512). The app did not crash — a subsequent plain-chat generate on the same loaded model succeeded cleanly (223 tok/s, 47ms ttft) — but the failure surfaces as a raw native error rather than a graceful, user-facing message.

Suspected trigger

The failing call appears to be a request that included the tool-calling / "web and tools" prompt template (see the accompanying racToolCallingRunLoopProto error in the same log burst), which pushes prompt length well past what a 512-ctx small model can hold. Worth checking whether the "enable web and tools" toggle should be disabled/warned for models whose catalog-declared context window can't fit the tool-calling template overhead.

Logcat evidence

08-17 15:16:50.146 I/qhexrt  ( 1375): rpcmem ion allocator ready
08-17 15:16:50.158 I/qhexrt  ( 1375): ion buffers bound (24 in, 21 out)
08-17 15:16:50.158 I/qhexrt  ( 1375): ion buffers bound (1 in, 1 out)
08-17 15:16:50.164 I/qhexrt  ( 1375): ion buffers bound (4 in, 21 out)
08-17 15:16:50.164 I/qhexrt  ( 1375): Session: built IoTensors for 3 graph(s), ion-persistent
08-17 15:16:50.548 I/qhexrt  ( 1375): rpcmem ion allocator ready
08-17 15:16:50.557 I/qhexrt  ( 1375): ion buffers bound (24 in, 21 out)
08-17 15:16:50.557 I/qhexrt  ( 1375): ion buffers bound (1 in, 1 out)
08-17 15:16:50.562 I/qhexrt  ( 1375): ion buffers bound (4 in, 21 out)
08-17 15:16:50.562 I/qhexrt  ( 1375): Session: built IoTensors for 3 graph(s), ion-persistent
08-17 15:16:50.566 E/qhexrt  ( 1375): lfm_generate: bad prompt len 599 (max_ctx 512)
08-17 15:16:50.566 E/qhexrt  ( 1375): plan: host-op 'lfm_generate' failed (InvalidArgument)
08-17 15:16:50.566 E/QHexRT  ( 1375): qhx_generate failed: InvalidArgument
08-17 15:16:50.566 I/System.out( 1375): [ERROR] [QHexRT] [Native] qhx_generate failed: InvalidArgument | source_file=qhexrt_llm_ops.cpp:282, source_function=qhexrt_llm_generate, device_model=Canoe for arm64, os_version=16, platform=android
08-17 15:16:50.566 I/System.out( 1375): [ERROR] [JNI.Commons] [Native] racToolCallingRunLoopProto: native proto API failed with code -130 (racToolCallingRunLoopProto) | source_file=runanywhere_commons_jni.cpp:524, source_function=makeProtoBufferByteArray, device_model=Canoe for arm64, os_version=16, platform=android

Start by reproducing the 599-token prompt failure on the QHexRT/NPU backend, then inspect qhexrt_llm_ops.cpp at qhexrt_llm_generate and the racToolCallingRunLoopProto path in runanywhere_commons_jni.cpp. The issue is done when prompts exceeding max_ctx no longer surface as a raw native error and the user receives graceful handling, including appropriate behavior for the web-and-tools option.