NVIDIA Model Optimizer (ModelOpt) quantization recipes

GROUP SIZE LEFT NULL 2026-10-04: ModelOpt is a recipe SELECTOR, not a format - the record's own bits field is 'w4a16 / w4a8 / w8a8 / w4a4-fp4 depending on the selected recipe', spanning group sizes 16, 32, 64, 128 and -1 depending on which recipe and export target is chosen. Recording one number would attribute a specific format's group size to a tool that has many. ModelOpt is a quantization *toolkit*, not a format: it emits checkpoints for every recipe in the TensorRT-LLM list, which is why its scheme value in this enum is the honest 'modelopt' rather than a specific bit-width. Worth knowing because it is now the front door for NVIDIA-format inference - TensorRT-LLM's docs tell you to quantize offline with ModelOpt and then run, including for the NVFP4 KV cache, where they note only FP8 weight/activation quantization is currently supported alongside an NVFP4 KV cache (quant-trtllm-quantization). The doc also documents a real constraint that shows up in practice: vision components of multimodal models stay FP16 by default and the language component decides which quantization methods are available. bits is therefore a set, not a number, and weight_group_size is null because it follows the recipe.