from chutes.chute import NodeSelector
from chutes.chute.template.sglang import build_sglang_chute

chute = build_sglang_chute(
    tee=True,
    username="chutes",
    readme="Gemma-4-31B-IT NVFP4 with DFlash speculative decoding",
    model_name="nvidia/Gemma-4-31B-IT-NVFP4",
    name="google/gemma-4-31B-turbo-TEE",
    image="chutes/sglang:0.5.21",
    official_package=True,
    concurrency=24,
    revision="4135a98a9b728a548947683219633b25682223ac",
    node_selector=NodeSelector(
        gpu_count=1,
        include=["pro_6000"],
    ),
    engine_args=(
        "--trust-remote-code "
        "--reasoning-parser gemma4 "
        "--tool-call-parser gemma4 "
        "--mem-fraction-static 0.89 "
        f"--cuda-graph-max-bs-decode 24 "
        "--disable-custom-all-reduce"
    ),
)