gemma-4-31B-it on Intel Arc Pro B65 32GB

docker run --rm -d --name lsv-g31-iq4-tp1-ctx131k --net=host --shm-size=32g --device /dev/dri:/dev/dri -v /dev/dri:/dev/dri -v /models/weights:/llm/models:ro -v /models/bench/lmx-ctx-20261019/session-a/g31-iq4-tp1-ctx131k/session:/bench -e ZE_AFFINITY_MASK=0 -e VLLM_XPU_ENABLE_XPU_GRAPH=1 -e DISABLE_ESIMD_GDN_OUTPROJ=0 -e DISABLE_ESIMD_INT4=0 -e DISABLE_ESIMD_INT4_GEMM=0 -e VLLM_OFFLOAD_WEIGHTS_BEFORE_QUANT=1 -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 -e VLLM_WORKER_MULTIPROC_METHOD=spawn -e CCL_ZE_IPC_EXCHANGE=sockets --entrypoint vllm intel/llm-scaler-vllm:0.31.0-b1-experimental@sha256:a397e04d66494637862b1927ee92ab4dd8b2f03f0dc9d359ec9536959a88e704 serve --model /llm/models/gemma-4-31B-it --host 0.0.0.0 --served-model-name g31 --tensor-parallel-size 1 --pipeline-parallel-size 1 --data-parallel-size 1 --gpu-memory-utilization 0.92 --max-model-len 8192 --max-num-seqs 16 --max-num-batched-tokens 8192 --block-size 64 --kv-cache-dtype auto --mamba-ssm-cache-dtype auto --dtype float16 --quantization sym_int4 --compilation-config \{\"cudagraph_mode\":\"FULL_DECODE_ONLY\"\,\"mode\":0\} --enable-prefix-caching --trust-remote-code --limit-mm-per-prompt \{\"image\":0\,\"video\":0\} --cpu-offload-gb 0 --port 8000

TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.