[bug]incorrect output
CUDA_DEVICE_ORDER=PCI_BUS_ID
CUDA_VISIBLE_DEVICES=4,5,6,7
vllm serve MiniMax-M3-NVFP4
--served-model-name minimax-m3
--tool-call-parser minimax_m3
--reasoning-parser minimax_m3
--enable-auto-tool-choice
--tensor-parallel-size 4
--gpu-memory-utilization 0.9
--max-model-len -1
--port 14501
--enable-prefix-caching
--enable-chunked-prefill
--block-size 128
--limit-mm-per-prompt.image 1
--limit-mm-per-prompt.video 0
--max-num-seqs 4
with 4x blackwell rtx pro 6000
Exact same issue on the same hardware (4XRTXPRO6000)
I believe that you'd need vllm with https://github.com/vllm-project/vllm/pull/46380 but as far as I can tell the patches aren't in master yet...
这个问题你解决了吗？如何解决的，我部署了这个补丁，手工安装的vllm启动正常，api调用还是乱码
I was basing my previous comment on whats inside readme + https://huggingface.co/nvidia/MiniMax-M3-NVFP4/discussions/1 I've tried to build custom version of https://github.com/eugr/spark-vllm-docker from https://github.com/vllm-project/vllm/pull/46380 which seems to be cummulative branch for these fixes but I wasn't able to do it due to too many merge conflicts against main + too little time.