# Official Prime RL inference entrypoint; requires the released Linux GPU stack.
[deployment]
type = "single_node"
gpus_per_node = 1
[vllm]
dtype = "bfloat16"
quantization = "fp8_per_block"
model = "Qwen/Qwen3.5-9B"
revision = "c202236235762e1c871ad0ccb60c8ee5ba337b9a"
max_model_len = 50000
max_num_seqs = 16
tensor_parallel_size = 1
tool_call_parser = "qwen3_coder"
reasoning_parser = "qwen3"
[server]
host = "127.0.0.1"
port = 8000