Grove Vision AI V2: NPU config mismatch (cmd_stream_version 2 vs 0) with official Studio build

Report bugs
#1

Board: Seeed Grove Vision AI Module V2 (Himax WiseEye2, Ethos-U55-64)
Host OS: macOS
Edge Impulse CLI: edge-impulse-daemon >= 1.27.1

Project: jegermnty-project-1
Project ID: 1103820
Impulse: Image classification, 96x96, 3 classes (hand, hold, unknown)

Deployment target: Seeed Grove Vision AI Module V2 (Himax WiseEye2)
Model optimization: Quantized (int8), TensorFlow Lite

Issue:
After flashing firmware.img + model_vela.tflite from the same Studio Build zip (tried deploy versions 4, 6, 8), AT+RUNIMPULSE always fails:

What works:

  • Flash via xmodem (921600) succeeds
  • AT+INFO shows correct project metadata (not Demo firmware)
  • Camera initializes (OV5647, “Taking photo…”)
  • edge-impulse-daemon connects; data acquisition works

What we tried:

  1. Multiple fresh Studio builds (Grove V2 target) — same error
  2. Ethos-U55-64 library deploy — embedded model is byte-identical to Grove prebuilt model_vela.tflite
  3. C++ library deploy (v8) + local recompile with ethos-u-vela 5.1.0 + Himax vela.ini — produces identical SHA256 to Studio model; still cmd_stream_version=0
  4. CDN base firmware (v1.75.3 Demo) + our model — different error: tensor arena / init (-6), not NPU mismatch

Conclusion:
Firmware runtime expects NPU command stream version 2, but Vela output (Studio and local vela 5.1.0) has optimizer cmd_stream_version 0. Appears to be a pipeline mismatch on the official Grove V2 deployment, not a hardware issue.

Request:
Is there a known fix, compatible Vela version, or updated firmware/model pair for Grove V2? Happy to provide build logs or zip hashes if needed.

Thanks!