Skip to content

Troubleshooting

Common build, link, and runtime issues — and how to fix them.

Build Issues

Missing HELIA backend module

FATAL_ERROR: CONFIG_HELIA_RT_BACKEND_HELIA requires Ambiq's HELIA
acceleration module, currently distributed as ns-cmsis-nn.

Cause: The HELIA backend needs the ns-cmsis-nn module, which is not bundled in the public repo.

Fix: Either:

  • Switch to the open CMSIS-NN backend: CONFIG_HELIA_RT_BACKEND_CMSIS_NN=y
  • Switch to Reference: CONFIG_HELIA_RT_BACKEND_REFERENCE=y
  • Contact support.aitg@ambiq.com for access to ns-cmsis-nn

Missing CMSIS-NN module (Zephyr)

FATAL_ERROR: CONFIG_HELIA_RT_BACKEND_CMSIS_NN requires an open CMSIS-NN
Zephyr module.

Fix: Add the CMSIS-NN module to your west.yml:

- name: cmsis-nn
  url: https://github.com/zephyrproject-rtos/cmsis-nn
  revision: main
  path: modules/lib/cmsis-nn

Third-party download failures

Something went wrong with the CMSIS download: ...

Cause: The Makefile build auto-downloads CMSIS and ns-cmsis-nn on first run. This can fail behind corporate proxies.

Fix:

  1. Set https_proxy / http_proxy environment variables

  2. Or manually download and set CMSIS_PATH / NS_CMSIS_NN_PATH:

    export CMSIS_PATH=/path/to/cmsis
export NS_CMSIS_NN_PATH=/path/to/ns_cmsis_nn

Flash / ITCM overflow

region `FLASH' overflowed by 12345 bytes

Fix options:

  1. Switch to the SIZE build variant (-Os / -Oz)
  2. Reduce your operator resolver — only register operators your model actually uses
  3. Use the prebuilt archive for the SIZE variant
  4. Place code in a larger memory region via linker script

Duplicate symbol errors

multiple definition of `tflite::ops::micro::Register_CONV_2D()'

Cause: Usually from linking both a prebuilt heliaRT archive and compiling kernels from source.

Fix: Use one or the other — never both. Remove either the .a or the source files from your build.

Runtime Issues

Arena too small

AllocateTensors() failed

or

Failed to allocate memory for tensor arena

Fix:

  1. Increase the arena size in your application code
  2. Use interpreter.arena_used_bytes() after AllocateTensors() to find the actual minimum
  3. Align arena to 16 bytes: alignas(16) uint8_t tensor_arena[ARENA_SIZE];

Incorrect output values

Possible causes:

  • Quantization mismatch: model expects int8 input but you're feeding float (or vice versa). Check input->type.
  • Wrong input scaling: the input must match the model's quantization parameters (input->params.scale and zero_point).
  • Model not compatible: ensure the .tflite was quantized for int8/int16 LiteRT for Micro, not float-only LiteRT.

Model loads but no inference output

Check: Did you call interpreter.AllocateTensors() before interpreter.Invoke()? Forgetting this is the most common beginner mistake.

Getting More Help