基础功能

graph_run_mode

图执行模式，取值：

• 0：在线推理场景下，请配置为0。
• 1：训练场景下，请配置为1，默认为1。

配置示例：

custom_op.parameter_map["graph_run_mode"].i = 1

训练/在线推理

session_device_id

当用户需要将不同的模型通过同一个脚本在不同的Device上执行，可以通过该参数指定Device的逻辑ID。

通常可以为不同的图创建不同的Session，并且传入不同的session_device_id，该参数优先级高于ASCEND_DEVICE_ID。

配置示例：

config_0 = tf.ConfigProto()
custom_op = config_0.graph_options.rewrite_options.custom_optimizers.add()
custom_op.name = "NpuOptimizer"
custom_op.parameter_map["session_device_id"].i = 0
config_0.graph_options.rewrite_options.remapping = RewriterConfig.OFF
config_0.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF
with tf.Session(config=config_0) as sess_0:
    sess_0.run(...)

config_1 = tf.ConfigProto()
custom_op = config_1.graph_options.rewrite_options.custom_optimizers.add()
custom_op.name = "NpuOptimizer"
custom_op.parameter_map["session_device_id"].i = 1
config_1.graph_options.rewrite_options.remapping = RewriterConfig.OFF
config_1.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF
with tf.Session(config=config_1) as sess_1:
    sess_1.run(...)

config_7 = tf.ConfigProto()
custom_op = config_7.graph_options.rewrite_options.custom_optimizers.add()
custom_op.name = "NpuOptimizer"
custom_op.parameter_map["session_device_id"].i = 7
config_7.graph_options.rewrite_options.remapping = RewriterConfig.OFF
config_7.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF
with tf.Session(config=config_7) as sess_7:
    sess_7.run(...)

训练/在线推理