engine = "sglang"
name = "GLM-5.2 NVFP4"
docker_image = "scitrera/dgx-spark-sglang-mm:v0"
model = "Mapika/GLM-5.2-NVFP4"
required_nodes = 4
tensor_parallel_size = 4
distributed_backend = "mp"

entrypoint = ["sglang", "serve"]

args = [
    "--quantization", "modelopt_fp4",
    "--mem-fraction-static", "0.85",
    "--context-length", "32768",
    "--fp4-gemm-backend", "cutlass",
    "--kv-cache-dtype", "auto",
    "--reasoning-parser", "glm45",
    "--tool-call-parser", "glm47",
    "--moe-runner-backend", "flashinfer_cutlass",
    "--trust-remote-code",
    "--cuda-graph-max-bs", "16",
    "--cuda-graph-bs", "1", "2", "4", "8", "12", "16",
    "--served-model-name", "glm52",
]

[env]
TORCH_NCCL_ASYNC_ERROR_HANDLING = "1"