Does eIQ Neutron support Vision Transformer (RF-DETR / DETR) conversion for NPU acceleration on i.MX

[vijayranaACL]

Goal

• Deploy RF-DETR Nano (vision transformer + DETR head) on NXP FRDM-IMX95
• Run inference with eIQ Neutron NPU via libneutron_delegate.so
• Need confirmation: Does eIQ support vision-transformer-based models for NPU acceleration?

My Board Details

• Board: NXP FRDM-IMX95 (imx95-15x15-lpddr4x-frdm)
• Kernel: Linux 6.18.2-1.0.0-gf49f45233f7b (aarch64, Feb 2026)
• CPU: 6× ARM Cortex-A55 (500–1800 MHz)
• NPU: eIQ Neutron-S (libneutron_delegate.so)
• Guide used: FRDM i.MX95 Getting Started – NPU section

Model & Pipeline Tried

• Model: RF-DETR Nano (DINOv2 ViT backbone + DETR decoder)
• Input: 384×384, COCO 80 classes
• Flow:
  1. PyTorch → ONNX (opset 18)
  2. ONNX → TFLite via onnx2tf (-fdosm for SavedModel)
  3. TFLite dynamic-range quant (float32 I/O) — CPU inference works
  4. Host: neutron-converter --input model.tflite --output model_neutron.tflite --target imx95
  5. Board: TFLite + Neutron delegate

What Works

• ONNX export succeeds
• TFLite conversion succeeds (~30 MB model)
• CPU inference is correct — detections match PyTorch (dog/person/car on test image)
• Model runs on FRDM with and without delegate

What Fails / Blocked

• No NPU graph created after conversion
  • Inspecting converted .tflite: no neutronGraph / NeutronOperator markers
  • Neutron delegate appears to offload 0 nodes → full CPU fallback
• Full INT8 quantization fails on transformer ops (RANGE, DIV, etc.)
• Used dynamic-range quant instead (weights quantized, float32 I/O)

Why I Think It’s Unsupported (Need Confirmation)

• RF-DETR is transformer-heavy: attention, BatchMatMul, Softmax, LayerNorm, Gather, etc.
• Neutron docs list mainly CNN ops (Conv2D, DepthwiseConv2D, Pooling, Add…)
  • Neutron Converter
  • Supported operators
• Similar split-model approach needed on Rockchip NPU (backbone on NPU, decoder on CPU):
  rfdetr-on-rockchip-npu

My Specific Questions

1. Does eIQ Neutron on i.MX95 support end-to-end vision transformer models (ViT / DETR / RF-DETR)?
2. If not, what is the recommended split (e.g. DINOv2 backbone on NPU, DETR head on CPU)?
3. Are BatchMatMul / Multi-Head Attention supported on Neutron-S for i.MX95 in current SDK?
4. Which eIQ Toolkit version matches BSP kernel 6.18.2-1.0.0 for FRDM-IMX95?
5. Should we use NXP tflite-profiler + tflite-quantizer instead of onnx2tf dynamic-range quant for better NPU compatibility?
   • Model Deployment Quick Start
6. Any reference example for transformer or DETR-style models on i.MX95 Neutron

Commands Used (for reproducibility)

Host (convert):

neutron-converter \

--input rf_detr_nano_full_int8.tflite \

--output rf_detr_nano_neutron_imx95.tflite \

--target imx95

Board (inference):

python3 redetr_test_neutron_imx95.py \

--model rf_detr_nano_neutron_imx95.tflite \

--image dog.jpg \

--delegate /usr/lib/libneutron_delegate.so

Expected issue: converted model has no neutronGraph → no NPU acceleration.

Please confirm:

• Is vision-transformer NPU acceleration supported on i.MX95 today?
• If not, what is the official workaround (model split, different export path, or future roadmap)?