|
62 | 62 | # /usr/local/Ascend/ascend-toolkit/latest
|
63 | 63 | "ASCEND_HOME_PATH":
|
64 | 64 | lambda: os.getenv("ASCEND_HOME_PATH", None),
|
65 |
| - # The path for HCCN Tool, the tool will be called by disaggregated prefilling |
66 |
| - # case. |
67 |
| - "HCCN_PATH": |
68 |
| - lambda: os.getenv("HCCN_PATH", "/usr/local/Ascend/driver/tools/hccn_tool"), |
69 | 65 | # The path for HCCL library, it's used by pyhccl communicator backend. If
|
70 | 66 | # not set, the default value is libhccl.so。
|
71 | 67 | "HCCL_SO_PATH":
|
72 |
| - # The prefill device id for disaggregated prefilling case. |
73 | 68 | lambda: os.environ.get("HCCL_SO_PATH", None),
|
74 |
| - "PROMPT_DEVICE_ID": |
75 |
| - lambda: os.getenv("PROMPT_DEVICE_ID", None), |
76 |
| - # The decode device id for disaggregated prefilling case. |
77 |
| - "DECODE_DEVICE_ID": |
78 |
| - lambda: os.getenv("DECODE_DEVICE_ID", None), |
79 |
| - # The port number for llmdatadist communication. If not set, the default |
80 |
| - # value is 26000. |
81 |
| - "LLMDATADIST_COMM_PORT": |
82 |
| - lambda: os.getenv("LLMDATADIST_COMM_PORT", "26000"), |
83 |
| - # The wait time for llmdatadist sync cache. If not set, the default value is |
84 |
| - # 5000ms. |
85 |
| - "LLMDATADIST_SYNC_CACHE_WAIT_TIME": |
86 |
| - lambda: os.getenv("LLMDATADIST_SYNC_CACHE_WAIT_TIME", "5000"), |
87 | 69 | # The version of vllm is installed. This value is used for developers who
|
88 | 70 | # installed vllm from source locally. In this case, the version of vllm is
|
89 | 71 | # usually changed. For example, if the version of vllm is "0.9.0", but when
|
|
100 | 82 | "VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP":
|
101 | 83 | lambda: bool(int(os.getenv("VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP", '0'))
|
102 | 84 | ),
|
| 85 | + # Whether to enable DBO feature for deepseek model. |
103 | 86 | "VLLM_ASCEND_ENABLE_DBO":
|
104 | 87 | lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_DBO", '0'))),
|
105 | 88 | # Whether to enable the model execute time observe profile. Disable it when
|
|
128 | 111 | "VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION":
|
129 | 112 | lambda: bool(
|
130 | 113 | int(os.getenv("VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION", '1'))),
|
131 |
| - |
132 | 114 | # `LLMDataDistCMgrConnector` required variable. `DISAGGREGATED_PREFILL_RANK_TABLE_PATH` is
|
133 | 115 | # used for llmdatadist to build the communication topology for kv cache transfer, it is
|
134 | 116 | # a required variable if `LLMDataDistCMgrConnector` is used as kv connector for disaggregated
|
|
0 commit comments