Is W4A4KV4 inference (NVFP4 KV Cache) on 5090 still not supported by Trt-LLM?

System Info

nvidia-smi: NVIDIA-SMI 570.86.10 Driver Version: 570.86.10 CUDA Version: 12.8
python --version: 3.12.11
tensorrt-llm: Version: 0.21.0

How would you like to use TensorRT-LLM

I have quantized a llama3-8B on 5090 using Tensor-model-optimizer to nvfp4 (w4a4kv4) using

CUDA_VISIBLE_DEVICES=2 python hf_ptq.py --pyt_ckpt_path /models/Meta-Llama-3-8B-Instruct --qformat nvfp4 --kv_cache_qformat nvfp4 --export_fmt hf --export_path /nvfp4-models/llama3-8b --dataset cnn_dailymail.

After that I tried to evaluate the throughput and latency of the nvfp4 model using Trt-LLM, but I found that Trt-LLM only supports FP8/BF16 KV Cache inference and there is no option to change the kv cache precision during inference.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

Start by reviewing the TensorRT-LLM documentation and examples referenced in the issue, then inspect how the supplied hf_ptq.py command exports the NVFP4 model. Verify whether TensorRT-LLM 0.21.0 can load the exported W4A4KV4 model on the stated 5090 setup; done means confirming support or documenting the missing inference path.