visual_quantization: invalid choice: 'mxfp8'

(trt_edgellm) lz@10-60-57-2:~/hf/models$ tensorrt-edgellm-quantize llm \
  --model_dir /home/lz/hf/models/Qwen_Qwen3-VL-2B-Instruct \
  --output_dir /home/lz/hf/models/q3vl_mxfp8 \
  --quantization mxfp8 \
  --lm_head_quantization mxfp8 \
  --visual_quantization mxfp8
usage: tensorrt-edgellm-quantize llm [-h] --model_dir MODEL_DIR --output_dir OUTPUT_DIR
                                     [--quantization {fp8,int4_awq,nvfp4,mxfp8,int8_sq}]
                                     [--lm_head_quantization {fp8,int4_awq,nvfp4,mxfp8}]
                                     [--visual_quantization {fp8}] [--audio_quantization {fp8}]
                                     [--kv_cache_quantization {fp8}] [--dtype {fp16}] [--device DEVICE]
                                     [--dataset DATASET] [--num_samples NUM_SAMPLES]
tensorrt-edgellm-quantize llm: error: argument --visual_quantization: invalid choice: 'mxfp8' (choose from 'fp8')

Why is MXFP8 quantization not supported for the visual encoder?

If I want to customize support for MXFP8 quantization for the visual encoder, what should I do?

Start at the tensorrt-edgellm-quantize llm CLI entry point and inspect how the --visual_quantization choices are defined and handled. Reproduce the command shown with --visual_quantization mxfp8, then trace the visual encoder quantization path. Done means MXFP8 is intentionally supported there, or its unsupported status and customization path are documented.