CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
INFER_CUDA_DEVICES=0,1,2,3,4,5,6,7 \
ARLE_DSV4_MOE_BACKEND=deepep \
ARLE_DSV4_INCREMENTAL_KV=1 \
ARLE_DSV4_FUSED_DISPATCH_PAYLOAD=1 \
ARLE_DSV4_ROUTE_GROUPED_EXPERTS=1 \
ARLE_CUDA_DISABLE_MARLIN_W4_FP8=1 \
nsys profile --trace cuda,nvtx,osrt --capture-range=cudaProfilerApi --capture-range-end=stop --export=sqlite --kill=none /root/arle/target/release/infer --model-path /root/DeepSeek-V4-Flash --port 18175 --num-slots 1 --max-seq-len 4096 --mem-fraction-static 0.10 --kv-cache-dtype fp8 --deepseek-distributed-layers 43
