Distributed Parallel

View Source On Gitee

Q: What should I do if the error message Init plugin so failed, ret = 1343225860 is displayed during the HCCL distributed training?

A: When the user starts distributed training on the Ascend and meets the error that HCCL fails to be initialized, the possible cause is that rank_table.json is incorrect. You can use the tool in hccl_tools.py to generate new rank_table.json. Alternatively, set the environment variable export ASCEND_SLOG_PRINT_TO_STDOUT=1 to enable the log printing function of HCCL and check the ERROR log information.


Q: In the GPU distributed training scenario, if the number of environment variables CUDA_VISIBLE_DEVICES set incorrectly is less than the number of processes executed, the process blocking problem may occur.

A: In this scenario, some training processes will prompt the following error:

[ERROR] DEVICE [mindspore/ccsrc/runtime/device/gpu/cuda_driver.cc:245] SetDevice] SetDevice for id:7 failed, ret[101], invalid device ordinal. Please make sure that the 'device_id' set in context is in the range:[0, total number of GPU). If the environment variable 'CUDA_VISIBLE_DEVICES' is set, the total number of GPU will be the number set in the environment variable 'CUDA_VISIBLE_DEVICES'. For example, if export CUDA_VISIBLE_DEVICES=4,5,6, the 'device_id' can be 0,1,2 at the moment, 'device_id' starts from 0, and 'device_id'=0 means using GPU of number 4.
[ERROR] DEVICE [mindspore/ccsrc/runtime/device/gpu/gpu_device_manager.cc:27] InitDevice] Op Error: Failed to set current device id | Error Number: 0

The remaining processes may normally execute to the initialization NCCL step due to the successful allocation of GPU resources, and the log is as follows:

[INFO] DEVICE [mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc:90] Initialize] Start initializing NCCL communicator for device 1

In this step, the NCCL interface ncclCommInitRank is called, which blocks until all processes agree. So if a process doesn't call ncclCommInitRank, it will cause the process to block. We have reported this issue to the NCCL community, and the community developers are designing a solution. The latest version has not been fixed, see issue link. Solution: Manually kill the training process. According to the error log, set the correct card number, and then restart the training task.


Q: What can we do when in the GPU distributed training scenario, if a process exits abnormally, it may cause other processes to block?

A: In this scenario, the abnormal process exits due to various problems, and the remaining processes are executed normally to the initialization NCCL step due to the successful allocation of GPU resources. The log is as follows:

[INFO] DEVICE [mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc:90] Initialize] Start initializing NCCL communicator for device 1

In this step, the NCCL interface ncclCommInitRank is called, which blocks until all processes agree. So if a process doesn't call ncclCommInitRank, it will cause the process to block. We have reported this issue to the NCCL community, and the community developers are designing a solution. The latest version has not been fixed, see issue link. Solution: Manually kill the training process and then restart the training task.


Q: When executing a GPU stand-alone single-card script, when the process is started without mpirun, calling the mindspore.communication.init method may report an error, resulting in execution failure, how to deal with it?

[CRITICAL] DISTRIBUTED [mindspore/ccsrc/distributed/cluster/cluster_context.cc:130] InitNodeRole] Role name is invalid...

A: In the case where the user does not start the process using mpirun but still calls the init() method, MindSpore requires the user to configure several environment variables and verify according to training and dynamic cluster startup methods. If without configuring, MindSpore may display the above error message. Therefore, it is suggested that only when performing distributed training, mindspore.communication.init is called, and in the case of not using mpirun, it is configured the correct environment variables according to the documentation to start distributed training.


Q: What can we do when performing multi-machine multi-card training via OpenMPI, the prompt fails due to MPI_Allgather?

pml_ucx.c:175 Error: Failed to receive UCX worker address: Not found (-13)
pml_ucx.c:452 Error: Failed to resolve UCX endpoint for rank X

A: This problem is that OpenMPI cannot communicate with the peer address when communicating on the Host side, which is generally caused by the different configuration of the NIC between the machines, and can be solved by manually setting the NIC name or subnet:

mpirun -n process_num --mca btl tcp --mca btl_tcp_if_include eth0 ./run.sh

The above instruction starts the process_num of run.sh processes, and selects the Host side communication mode as tcp. The network card selects eth0, so that the network card used on each machine is the same, and then the communication abnormal problem is solved.

You can also select subnets for matching:

mpirun -n process_num --mca btl tcp --mca btl_tcp_if_include 192.168.1.0/24 ./run.sh

The subnet range needs to include the IP addresses used by all machines.


Q: What can we do when performing distributed training via OpenMPI, stand-alone multi-card training is normal, but during the multi-machine multi-card training, some machines prompt the GPU device id setting to fail?

[ERROR] DEVICE [mindspore/ccsrc/runtime/device/gpu/cuda_driver.cc:245] SetDevice] SetDevice for id:7 failed, ret[101], invalid device ordinal. Please make sure that the 'device_id' set in context is in the range:[0, total number of GPU). If the environment variable 'CUDA_VISIBLE_DEVICES' is set, the total number of GPU will be the number set in the environment variable 'CUDA_VISIBLE_DEVICES'. For example, if export CUDA_VISIBLE_DEVICES=4,5,6, the 'device_id' can be 0,1,2 at the moment, 'device_id' starts from 0, and 'device_id'=0 means using GPU of number 4.
[ERROR] DEVICE [mindspore/ccsrc/runtime/device/gpu/gpu_device_manager.cc:27] InitDevice] Op Error: Failed to set current device id | Error Number: 0

A: In the multi-machine scenario, each process card number needs to be calculated after the host side AllGatherand HOSTNAME. If the same HOSTNAME is used between machines, the process card number will be calculated incorrectly, causing the card number to cross the boundary and the setting to fail. This can be resolved by setting the HOSTNAME of each machine to its respective IP address in the execution script:

export HOSTNAME=node_ip_address

Q: What can we do when performing multi-machine multi-card training via OpenMPI, the NCCL error message displays that the network is not working?

include/socket.h:403 NCCL WARN Connect to XXX failed: Network is unreachable

A: This problem is that NCCL cannot communicate with the peer address when synchronizing process information or initializing the communication domain on the Host side, which is generally caused by the different configuration of the network card between the machines, and the network card can be selected by setting the NCCL environment variable NCCL_SOCKET_IFNAME:

export NCCL_SOCKET_IFNAME=eth

The above command sets the NCCL to select the network card name with eth in the Host side to communicate.


Q: After selecting RDMA NIC with a specific name (set via NCCL_SOCKET_IFNAME) for communication for multiple machines and multiple cards, the training still reports an error:

misc/ibvwrap.cc:284 NCCL WARN Call to ibv_modify_qp failed with error Invalid argument
...
include/socket.h:403 NCCL WARN Connect to XXX failed: Connection refused

A: Generally this problem is the existence of differences in the configuration of the RDMA network card between multiple machines, which need to be analyzed on a case-by-case basis. However, the common reason is there are IB protocol and RoCE protocol on some host NICs at the same time, and there may be connection establishment failure. Solution:

The following command is required to specify the name of the RDMA NIC to be used as beginning with ib:

export NCCL_IB_HCA=mlx

Q:Single-machine multi-card training can be successful, but after expanding the script to multi-machine multi-card, other hosts prompt all kinds of errors:

There are various types of errors reported. Here are a few typical ones:

  1. The installed whl package could not be found.

  2. IB NIC communication failure.

  3. Cuda library load failure.

A: These problems are caused by the fact that when mpirun starts other hosts, the environment variables of the other hosts (including the NCCL's NIC selection configurations) are not synchronized with the local machine, resulting in the phenomenon that single-machine multi-card executes normally while multi-machine multi-card fails. The solution is to export specific environment variables via mpirun -x option:

mpirun --hostfile /path/to/hostfile -n 64 -x PYTHONPATH -x GLOG_v -x LD_LIBRARY_PATH -x NCCL_SOCKET_IFNAME -x NCCL_IB_HCA -x NCCL_DEBUG=INFO python train.py

The above command exports some environment variables that have been set on this machine to other hosts, ensuring that the environment variables of all hosts remain the same before executing the training script to achieve the goal of multi-machine multi-card training.


Q: Performing the distributed training via OpenMPI on Ascend, got HcclCommInitRootInfo error message:

Ascend collective Error: "HcclCommInitRootInfo failed. | Error Number 2

A: Currently, when training via OpenMPI, hccl needs to allocate about 300M device memory for each card within a communicator. The more communicators one card involved in, the more extra device memory needed. This probably cause memory issue. You can set variable_memory_max_size in contextto reduce variable memory for Ascend processes, so that hccl will have enough memory to create communicators.


Q: When executing a distributed network under auto_parallel, an error is reported that the tensor cannot be split by the strategy. How can I solve it?

np_tensor can not be split by strategy!

A: This error indicates that a strategy is configured for a parameter on the network, but a certain dimension of the parameter is not devisible by the strategy. There are two possible problems: 1. The parameter is used as the input of an operator, and the shard interface is called to set an illegel strategy for this operator. 2. When dataset_strategy="data_parallel" or full_batch=False is set in auto_parallel_context, the framework will automatically set a data-parallel strategy for network input. This error is also reported if the network input contains parameter whose shape cannot be divisible by the data-parallel strategy. However, auto-parallel only supports Tensor as network input, and you need to make adjustments to your script.


Q: What should I do if a process exits abnormally during the execution of multi-card training on a Linux environment and there is a shared memory residue through the ipcs command?

A: In the case of multi-card training and enabling graph operator fusion, the framework uses a shared memory mechanism for unified compilation of operators among multiple cards, and the shared memory is not effectively freed if the process ends unexpectedly due to internal or external exceptions during the compilation process. The ipcs command shows that nattch of the residual shared memory is 0. The framework will take over the shared memory again when the training script is re-executed, and it can be released normally as long as no exception occurs. You can also release the shared memory by ipcrm command, which will not affect the training script execution.


Q: I try to train a small network on Ascend platform but during initializing distributed module, device memory not enough exception is still thrown. How should I solve this issue?

A: This is because under the Ascend platform, the MindSpore backend defaults to pre-allocate a block of memory, with approximately 80% of NPU memory occupied and the remaining 20% used for initializing the HCCL collection communication library. Each HCCL communication group occupies 200 MB memory by default, so in scenarios with more communication groups, it is easy to encounter device side memory shortage errors. The solution is to set up HCCL_BUFFSIZE environment variable to change communication domain memory usage. Specific configuration method can refer to HCCL official document.