IMAGE=ai01/vllm-v9-b12x-tp3:20260710-v7-routing STANDARD_MODEL=/root/.cache/huggingface/local/DeepSeek-V4-Flash NAME=deepseek-v4-flash PORT=8000 GPUS=0,1,2 TP=3 BACKEND=b12x-a16 MODE=standard-mtp2 MAX_NUM_SEQS=64 MAX_MODEL_LEN=524288 MAX_BATCHED=8192 GPU_MEM=0.94 GRAPH=512 PREFIX_CACHE=1 CACHE=/srv/ai/vllm/deepseek-v4-flash/cache CONTAINER_TMP=/srv/ai/tmp/deepseek-v4-flash ENABLE_TOPO_PIN=1 PIN_CPUSET_CPUS=0-47 PIN_CPUSET_MEMS=0 AOT_COMPILE=0 BREAKABLE_CUDAGRAPH=1 PCIE_ONESHOT_SINGLE_CHANNEL=1 PCIE_ONESHOT_MIN_SIZE=64KB DEBUG_B12X_CAPTURE_SEQUENCE=0