INFO 09-13 07:24:17 [__init__.py:52] Available plugins for group vllm.platform_plugins:
INFO 09-13 07:24:17 [__init__.py:54] - ascend -> vllm_ascend:register
INFO 09-13 07:24:17 [__init__.py:57] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.
INFO 09-13 07:24:17 [__init__.py:272] Platform plugin ascend is activated
INFO 09-13 07:24:17 [platform.py:70] Breakable CUDAGraph on Ascend is opt-in; using VLLM_USE_BREAKABLE_CUDAGRAPH=0.
{
  "command": [
    "vllm",
    "serve",
    "/data/shared_models/Qwen--Qwen2.5-14B-Instruct",
    "--gpu-memory-utilization",
    "0.85",
    "--max-model-len",
    "4096",
    "--served-model-name",
    "qwen14b",
    "--port",
    "8353",
    "--compilation-config",
    "{\"cudagraph_capture_sizes\":[32,64,128]}"
  ],
  "environment": {
    "VLLM_HUST_FIA_DEMASK": "1",
    "VLLM_ASCEND_ENABLE_CASCADE_DECODE": "1",
    "VLLM_ASCEND_ENABLE_CASCADE_GRAPH": "1",
    "VLLM_HUST_FI_PREFILL_OUT": "1",
    "VLLM_HUST_SKIP_COS_SIN": "1",
    "VLLMHUST_EXT_ENABLED_BUNDLES": "org.vllm-hust.fia-demask,org.vllm-hust.split-batch-full-graph,org.vllm-hust.zerocost-wiring"
  },
  "native_extension_manifests": {}
}
