tensorflow多流问题 #1
Description
首先抱歉，tensorflow仓库不能提issue，只能在这里提了：
想请教一个问题：
我使用r2.2_multistream分支代码编译so移植到我们的服务中，现在发现一个问题，就是我们的模型使用with tf.device('/GPU:0')或with tf.device('/GPU:1') 水平把并行部分拆到了两个卡上进行并行计算，使用这个分支的so加载我们的模型发现只能把模型加载到一张卡，推理时只有一张卡有使用率，另外一个一直为0。这个多流技术无法支持多卡吗？
启动时确实好像能看到session启动了多个stream：
339703 2025-03-31 17:00:29.328625: I tensorflow/core/common_runtime/direct_session.cc:190] New DirectSession, devices:
339704 2025-03-31 17:00:29.328636: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:CPU:0
339705 2025-03-31 17:00:29.328642: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:STREAM_GPU_0:1
339706 2025-03-31 17:00:29.328649: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:STREAM_GPU_0:0
339707 2025-03-31 17:00:29.328655: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:STREAM_GPU_1:1
339708 2025-03-31 17:00:29.328661: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:STREAM_GPU_1:0
339709 2025-03-31 17:00:29.328667: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:XLA_CPU:0
339710 2025-03-31 17:00:29.328674: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:XLA_GPU:0
339711 2025-03-31 17:00:29.328680: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:GPU:0
339712 2025-03-31 17:00:29.328686: I tensorflow/core/common_runtime/direct_session.cc:192] /job:localhost/replica:0/task:0/device:GPU:1
如果能帮忙看下就太感谢了！