From 7691868e2becdd2b80216f8e2b0c6be2320c8730 Mon Sep 17 00:00:00 2001 From: huilan li Date: Tue, 15 Jul 2025 11:01:03 +0800 Subject: [PATCH] master_hccl_buffer_size --- .../source_en/api_python/env_var_list.rst | 14 ++++++++++++++ .../source_zh_cn/api_python/env_var_list.rst | 14 ++++++++++++++ 2 files changed, 28 insertions(+) diff --git a/docs/mindspore/source_en/api_python/env_var_list.rst b/docs/mindspore/source_en/api_python/env_var_list.rst index ac13abca24..0b5d80d557 100644 --- a/docs/mindspore/source_en/api_python/env_var_list.rst +++ b/docs/mindspore/source_en/api_python/env_var_list.rst @@ -350,6 +350,20 @@ Graph Compilation and Execution No setting or use other value: Asynchronous launch operators. - + * - MS_DEV_HCCL_CONF + - Configure the memory allocation. + - String + - Configuration items, with the format of "key: value", multiple configuration items separated by commas, for example, "export MS_DEV_HCCL_CONF=enable_hccl_config:True,hccl_customized_default:100MB,hccl_list_config:0-1-2-3=200MB,hccl_stride_config:4-7:2=50MB". + + enable_hccl_config: Whether to enable virtual memory, with a default value of false. + + hccl_customized_default: Set the virtual memory alignment size in MB. + + hccl_list_config: Sets the HCCL_BUFFSIZE for the specified rank_list in MB. + + hccl_stride_config: Sets the specified rank range, splits the communication group according to the specified stride, and configures the HCCL_BUFFSIZE for this rank_list in MB. + - + * - MS_DEV_HOST_BLOCKING_RUN - Control whether the operator is launched in a single thread in PyNative mode. When enabled, the operator will be launched in a single thread in PyNative mode. - Integer diff --git a/docs/mindspore/source_zh_cn/api_python/env_var_list.rst b/docs/mindspore/source_zh_cn/api_python/env_var_list.rst index 23c1c278d9..5eebe342be 100644 --- a/docs/mindspore/source_zh_cn/api_python/env_var_list.rst +++ b/docs/mindspore/source_zh_cn/api_python/env_var_list.rst @@ -350,6 +350,20 @@ 不设置或其他值：不开启算子同步下发 - + * - MS_DEV_HCCL_CONF + - 设置HCCL_BUFFSIZE显存策略 + - String + - 配置项，格式为key:value，多个配置项以逗号分隔，例如 `export MS_DEV_HCCL_CONF=enable_hccl_config:True,hccl_customized_default:100MB,hccl_list_config:0-1-2-3=200MB,hccl_stride_config:4-7:2=50MB`。 + + enable_hccl_config: 是否使能HCCL_BUFFSIZE显存，默认值为false。 + + hccl_customized_default: 设置HCCL_BUFFSIZE全局默认大小，单位为MB。 + + hccl_list_config: 设置指定rank_list的HCCL_BUFFSIZE，单位为MB。 + + hccl_stride_config: 设置指定rank范围、按照指定stride切分通讯组，并配置此rank_list的HCCL_BUFFSIZE，单位为MB。 + - + * - MS_DEV_HOST_BLOCKING_RUN - 控制动态图算子是否单线程下发。开启后，动态图算子将采用单线程下发。 - Integer -- Gitee