docker run --rm --name sg-http --publish 127.0.0.1:8710:8000 --device /dev/tenstorrent:/dev/tenstorrent -v /dev/hugepages:/dev/hugepages -v /dev/hugepages-1G:/dev/hugepages-1G -v /var/tmp/gemma4-12b/native/p3d-pc-gate-s1/gemma-4-12B-it:/model:ro -v /var/tmp/gemma4-12b/serve-cache:/cache --workdir /home/container_app_user/tt-metal --entrypoint /home/container_app_user/tt-metal/python_env/bin/python -v /var/tmp/gemma4-12b/native/p3d-pc-gate-s1/gemma-4-12B-it-assistant:/drafter:ro -e GEMMA4_DECODE_KV_ROWS=1 -e GEMMA4_DECODE_MM_DN=1 -e GEMMA4_FUSE_GELU_MUL=1 -e GEMMA4_NATIVE=/model -e GEMMA4_SINGLE_USER=1 -e GEMMA4_VERIFY_SDPA=batched -e GEMMA_PREFIX_SNAPSHOTS=4 -e HF_HUB_OFFLINE=1 -e HF_MODEL=/model -e TORCHDYNAMO_DISABLE=1 -e TRANSFORMERS_OFFLINE=1 -e TT_METAL_CACHE=/cache/6677ccdf80ee229b -e VLLM_TARGET_DEVICE=tt gemma4-12b:p3d-pc-gate-s1 -m vllm.entrypoints.openai.api_server --model /model --served-model-name google/gemma-4-12B-it --host 0.0.0.0 --port 8000 --max-model-len 262144 --max-num-seqs 1 --block-size 64 --max-num-batched-tokens 16384 --enable-chunked-prefill --enable-prefix-caching --seed 0 --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4 --additional-config '{"tt": {"trace_mode": "all", "enable_model_warmup": true, "trace_region_size": 200000000}}' --speculative-config '{"method": "gemma4_assistant", "model": "/drafter", "num_speculative_tokens": 5}'