Environment Variables#
vllm-ascend uses the following environment variables to configure the system:
env_variables: Dict[str, Callable[[], Any]] = {
# max compile thread number for package building. Usually, it is set to
# the number of CPU cores. If not set, the default value is None, which
# means all number of CPU cores will be used.
"MAX_JOBS":
lambda: os.getenv("MAX_JOBS", None),
# The build type of the package. It can be one of the following values:
# Release, Debug, RelWithDebugInfo. If not set, the default value is Release.
"CMAKE_BUILD_TYPE":
lambda: os.getenv("CMAKE_BUILD_TYPE"),
# Whether to compile custom kernels. If not set, the default value is True.
# If set to False, the custom kernels will not be compiled. Please note that
# the sleep mode feature will be disabled as well if custom kernels are not
# compiled.
"COMPILE_CUSTOM_KERNELS":
lambda: bool(int(os.getenv("COMPILE_CUSTOM_KERNELS", "1"))),
# The CXX compiler used for compiling the package. If not set, the default
# value is None, which means the system default CXX compiler will be used.
"CXX_COMPILER":
lambda: os.getenv("CXX_COMPILER", None),
# The C compiler used for compiling the package. If not set, the default
# value is None, which means the system default C compiler will be used.
"C_COMPILER":
lambda: os.getenv("C_COMPILER", None),
# The version of the Ascend chip. If not set, the default value is
# ASCEND910B1(Available for A2 and A3 series). It's used for package building.
# Please make sure that the version is correct.
"SOC_VERSION":
lambda: os.getenv("SOC_VERSION", "ASCEND910B1"),
# If set, vllm-ascend will print verbose logs during compilation
"VERBOSE":
lambda: bool(int(os.getenv('VERBOSE', '0'))),
# The home path for CANN toolkit. If not set, the default value is
# /usr/local/Ascend/ascend-toolkit/latest
"ASCEND_HOME_PATH":
lambda: os.getenv("ASCEND_HOME_PATH", None),
# The path for HCCN Tool, the tool will be called by disaggregated prefilling
# case.
"HCCN_PATH":
lambda: os.getenv("HCCN_PATH", "/usr/local/Ascend/driver/tools/hccn_tool"),
# The path for HCCL library, it's used by pyhccl communicator backend. If
# not set, the default value is libhccl.so。
"HCCL_SO_PATH":
# The prefill device id for disaggregated prefilling case.
lambda: os.environ.get("HCCL_SO_PATH", None),
"PROMPT_DEVICE_ID":
lambda: os.getenv("PROMPT_DEVICE_ID", None),
# The decode device id for disaggregated prefilling case.
"DECODE_DEVICE_ID":
lambda: os.getenv("DECODE_DEVICE_ID", None),
# The port number for llmdatadist communication. If not set, the default
# value is 26000.
"LLMDATADIST_COMM_PORT":
lambda: os.getenv("LLMDATADIST_COMM_PORT", "26000"),
# The wait time for llmdatadist sync cache. If not set, the default value is
# 5000ms.
"LLMDATADIST_SYNC_CACHE_WAIT_TIME":
lambda: os.getenv("LLMDATADIST_SYNC_CACHE_WAIT_TIME", "5000"),
# The version of vllm is installed. This value is used for developers who
# installed vllm from source locally. In this case, the version of vllm is
# usually changed. For example, if the version of vllm is "0.9.0", but when
# it's installed from source, the version of vllm is usually set to "0.9.1".
# In this case, developers need to set this value to "0.9.0" to make sure
# that the correct package is installed.
"VLLM_VERSION":
lambda: os.getenv("VLLM_VERSION", None),
# Whether to enable the trace recompiles from pytorch.
"VLLM_ASCEND_TRACE_RECOMPILES":
lambda: bool(int(os.getenv("VLLM_ASCEND_TRACE_RECOMPILES", '0'))),
# Whether to enable fused_experts_allgather_ep. MoeInitRoutingV3 and
# GroupedMatmulFinalizeRouting operators are combined to implement EP.
"VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP":
lambda: bool(int(os.getenv("VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP", '0'))
),
"VLLM_ASCEND_ENABLE_DBO":
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_DBO", '0'))),
# Whether to enable the model execute time observe profile. Disable it when
# running vllm ascend in production environment.
"VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE":
lambda: bool(int(os.getenv("VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE", '0'))
),
# MOE_ALL2ALL_BUFFER:
# 0: default, normal init.
# 1: enable moe_all2all_buffer.
"MOE_ALL2ALL_BUFFER":
lambda: bool(int(os.getenv("MOE_ALL2ALL_BUFFER", '0'))),
# Some models are optimized by vllm ascend. While in some case, e.g. rlhf
# training, the optimized model may not be suitable. In this case, set this
# value to False to disable the optimized model.
"USE_OPTIMIZED_MODEL":
lambda: bool(int(os.getenv('USE_OPTIMIZED_MODEL', '1'))),
# The tolerance of the kv cache size, if the difference between the
# actual kv cache size and the cached kv cache size is less than this value,
# then the cached kv cache size will be used.
"VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE":
lambda: int(
os.getenv("VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE", 64)),
# Whether to enable the topk optimization. It's enabled by default. Please set to False if you hit any issue.
# We'll remove this flag in the future once it's stable enough.
"VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION":
lambda: bool(
int(os.getenv("VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION", '1'))),
# `LLMDataDistCMgrConnector` required variable. `DISAGGREGATED_PREFILL_RANK_TABLE_PATH` is
# used for llmdatadist to build the communication topology for kv cache transfer, it is
# a required variable if `LLMDataDistCMgrConnector` is used as kv connector for disaggregated
# pd. The rank table can be generated by adopting the script `gen_ranktable.sh`
# in vllm_ascend's example folder.
"DISAGGREGATED_PREFILL_RANK_TABLE_PATH":
lambda: os.getenv("DISAGGREGATED_PREFILL_RANK_TABLE_PATH", None),
# `LLMDataDistCMgrConnector` required variable. `VLLM_ASCEND_LLMDD_RPC_IP` is used as the
# rpc communication listening ip, which will be used to receive the agent metadata from the
# remote worker.
"VLLM_ASCEND_LLMDD_RPC_IP":
lambda: os.getenv("VLLM_ASCEND_LLMDD_RPC_IP", "0.0.0.0"),
# `LLMDataDistCMgrConnector` required variable. `VLLM_LLMDD_RPC_PORT` is used as the
# rpc communication listening port, which will be used to receive the agent metadata from the
# remote worker.
"VLLM_LLMDD_RPC_PORT":
lambda: int(os.getenv("VLLM_LLMDD_RPC_PORT", 5557)),
# Whether to enable mla_pa for deepseek mla decode, this flag will be removed after its available torch_npu is public accessible
# and the mla_pa will be the default path of deepseek decode path.
"VLLM_ASCEND_MLA_PA":
lambda: int(os.getenv("VLLM_ASCEND_MLA_PA", 0)),
# Whether to enable MatmulAllReduce fusion kernel when tensor parallel is enabled.
# this feature is supported in A2, and eager mode will get better performance.
"VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE":
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE", '0'))),
# Whether to enable the alltoall_seq flag, this provides a basic framework on the basis of alltoall for easy expansion.
# 0: default, normal init.
# 1: enable moe all2all seq.
"VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ":
lambda: bool(int(os.getenv('VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ', '0'))),
}