Neuron Runtime Troubleshooting on Inf1, Inf2 and Trn1
Contents
This document is relevant for: Inf1, Inf2, Trn1, Trn1n
Neuron Runtime Troubleshooting on Inf1, Inf2 and Trn1#
This document aims to provide more information on how to fix issues you might encounter while using the Neuron Runtime 2.x or above. For each issue we will provide an explanation of what happened and what can potentially correct the issue.
If your issue is not listed below or you have a more nuanced problem, contact us via issues posted to this repo, the AWS Neuron developer forum, or through AWS support.
Table of contents
Generic Errors#
Neuron Driver installation fails#
aws-neuron-dkms is a driver package which needs to be compiled during
installation. The compilation requires kernel headers for the instance’s
kernel. uname -r can be used to find kernel version in the instance.
In some cases, the installed kernel headers might be newer than the
instance’s kernel itself.
Please look at the aws-neuron-dkms installation log for message like the following:
Building for 4.14.193-149.317.amzn2.x86_64
Module build for kernel 4.14.193-149.317.amzn2.x86_64 was skipped since the
kernel headers for this kernel does not seem to be installed.
If installation log is not available, check whether the module is loaded.
$ lsmod | grep neuron
If the above has no output then that means aws-neuron-dkms
installation is failed.
Solution#
Stop all applications using the NeuronCores.
Uninstall aws-neuron-dkms
sudo apt remove aws-neuron-dkmsorsudo yum remove aws-neuron-dkmsInstall kernel headers for the current kernel
sudo apt install -y linux-headers-$(uname -r)orsudo yum install -y kernel-devel-$(uname -r) kernel-headers-$(uname -r)Install aws-neuron-dkms
sudo apt install aws-neuron-dkmsorsudo yum install aws-neuron-dkms
Application fails to start#
Neuron Runtime requires Neuron Driver(aws-neuron-dkms package) to access Neuron devices. If the driver is not installed then Neuron Runtime wont able to access the Neuron devices and will fail with an error message in console and syslog.
If aws-neuron-dkms is not installed then the error message will be like the following:
2021-Aug-11 18:38:27.0917 13713:13713 ERROR NRT:nrt_init Unable to determine Neuron Driver version. Please check aws-neuron-dkms package is installed.
If aws-neuron-dkms is installed but does not support the latest runtime then the error message will be like the following:
2021-Aug-11 19:18:21.0661 24616:24616 ERROR NRT:nrt_init This runtime requires Neuron Driver version 2.0 or greater. Please upgrade aws-neuron-dkms package.
When using any supported framework from Neuron SDK version 2.5.0 and Neuron Driver (aws-neuron-dkms) versions 2.4 or older, Neuron Runtime will return the following error message:
2022-Dec-01 09:34:12.0559 138:138 ERROR HAL:aws_hal_tpb_pooling_write_profile failed programming the engine
Solution#
Please follow the installation steps in Setup Guide to install aws-neuronx-dkms.
This Neuron Runtime (compatibility id: X) is not compatible with the installed aws-neuron-dkms package#
This error is caused by incompatibility between the Neuron Driver (dkms package) and the Runtime Library (runtime-lib package). The driver remains backwards compatible with older versions of Neuron Runtime, but newer versions of the Runtime might rely on the functionality that is only provided by a newer driver. In that case, an update to the newer driver is required.
In some cases the compatibility error persists even after the driver has been updated. That happens when the update process fails to reload the driver at the end of the update. Note that $ modinfo neuron will misleadingly show the new version because modinfo reads the version information for neuron.ko file that’s been successfully replaced.
Reload failure happens because one of the processes is still using Neuron Devices and thus the driver cannot be reloaded.
Solution#
Check for any process that is still using the Neuron driver by running lsmod:
[email protected]:~$ lsmod | grep neuron
neuron 237568 0
[email protected]:~$
“Used by” counter, the second number, should be 0. If it is not, there is still a running process that is using Neuron. Terminate that process and either:
$ sudo rmmod neuron
$ sudo modprobe neuron
Or simply rerun the installation one more time. The driver logs its version in dmesg:
$ sudo dmesg
...
[21531.105295] Neuron Driver Started with Version:2.9.4.0-8a6fdf292607dccc3b7059ebbe2fb24c60dfc7c4
A common culprit is a Jupyter process. If you are using Jupyter on the instance, make sure to terminate Jupyter process before updating the driver.
Neuron Core is in use#
A Neuron Core cant be shared between two applications. If an application started using a Neuron Core all other applications trying to use the NeuronCore would fail during runtime initialization with the following message in the console and in syslog:
2021-Aug-27 23:22:12.0323 28078:28078 ERROR NRT:nrt_allocate_neuron_cores NeuronCore(s) not available - Requested:nc1-nc1 Available:0
Solution#
Terminate any other processes that are using NeuronCore devices and then try launching the application again. If you are using Jupyter, ensure that you only have a single Jupyter kernel attempting to access the NeuronCores by restarting or shutting-down any other kernels, which will release any NeuronCores that might be in use.
Unsupported NEFF Version#
While loading a model(NEFF), Neuron Runtime checks the version compatibility. If the version the NEFF is incompatible with Runtime then it would fail the model load with following error message:
NEFF version mismatch supported: 1.1 received: 2.0
Solution#
Use compatible versions of Neuron Compiler and Runtime. Updating to the latest version of both Neuron Compiler and Neuron Runtime is the simplest solution. If updating one of the two is not an option, please refer to the neuron-runtime-release-notes of the Neuron Runtime to determine NEFF version support.
Insufficient Memory#
While loading a model(NEFF), Neuron Runtime reserves both device and host memory for storing weights, ifmap and ofmap of the Model. The memory consumption of each model is different. If Neuron Runtime is unable to allocate memory then the model load would fail with the following message in syslog
kernel: [XXXXX] neuron:mc_alloc: device mempool [0:0] total 1073741568 occupied 960539030 needed 1272 available 768
Solution#
As the error is contextual to what’s going on with your instance, the exact next step is unclear. Try unloading some of the loaded models which will free up device DRAM space. If this is still a problem, moving to a larger Inf1 instance size with additional NeuronCores may help.
Insufficient number of NeuronCores#
The NEFF requires more NeuronCores than available on the instance.
Check for error messages in syslog similar to:
NRT: 26638:26638 ERROR TDRV:db_vtpb_get_mla_and_tpb Could not find VNC id n
NRT: 26638:26638 ERROR NMGR:dlr_kelf_stage Failed to create shared io
NRT: 26638:26638 ERROR NMGR:stage_kelf_models Failed to stage graph: kelf-a.json to NeuronCore
NRT: 26638:26638 ERROR NMGR:kmgr_load_nn_post_metrics Failed to load NN: xxxxxxx, err: 2
Solution#
The NeuronCores may be in use by models you are not actively using. Ensure you’ve unloaded models you’re not using and terminated unused applications. If this is still a problem, moving to a larger Inf1 instance size with additional NeuronCores may help.
Numerical Error#
Neuron Devices will detect any NaN generated during execution and report it. If Neuron Runtime sees NaNs are generated then it would fail the execution request with Numerical Error with the following message:
nrtd[nnnnn]: .... Error notifications found on NC .... INFER_ERROR_SUBTYPE_NUMERICAL
Memory Errors#
Transient memory errors#
Uncorrectable memory error is detected on Neuron device: 5:1 metadata: 0x2. The error might cause incorrect computational results and might affect training convergence. Please
terminate and restart from the last checkpoint if the convergence is impacted.
Neuron detected a single uncorrectable bit flip in the device memory. The execution can continue but there is a possibility of a numerical error. If this is a concern, terminate and restart from the last known good check point.
Persistent memory errors#
Uncorrectable memory error is detected on Neuron device: 5:1 metadata: 0x2. Failing execution.
Multiple uncorrectable errors are detected during execution. The execution cannot continue. This is most likely caused by faulty hardware. Terminate and move to a different instance.
Failure to initialize Neuron#
nd0 nc0 Timestamp program stop timeout (1000 ms)
nd0 nc0 Error while waiting for timestamp program to end on TPB eng 0
nd0 nc0 Failed to stop neuron core
nd0 nc0 Failed to end timestamp sync programs
TDRV not initialized
Failed to initialize devices, error:5
Previously executed application left Neuron devices in running state. Reset Neuron devices but reloading Neuron Driver. Note, this is a temporary workaround, future versions of Neuron will reset running devices automatically.
sudo rmmod neuron; sudo modprobe neuron
An application is trying to use more cores that are available on the instance#
Could not open the nd1
Use properly sized instance. trn1.32xlarge has 32 Neuron Cores, trn1.2xlarge has 2 Neuron Cores.
EFA and Collective Communication Errors#
Missing aws-neuronx-collectives package#
aws-neuronx-collectives package is required to execute Collective Communication on a single instance and across multiple instances.
NCCL init error: Error opening libnccom.so, cannot use collective operations! Please set LD_LIBRARY_PATH to library location. Error: libnccom.so: cannot open shared object
file: No such file or directory
Please make sure to install correct version of aws-neuronx-collectives; for detailed installation instructions visit Neuron documentation
Install aws-neuornx-collectives package. If the installation used non-default destination set LD_LIBRARY_PATH.
Missing efa installer package.#
efa-installer package is required to execute Collective Communication across multiple instances.
Unable to run multi-instance workload. Ofi plugin is not installed or EFA is not enabled
Follow the directions to install efa-installer package. Make sure to add the path to to libfabric library to LD_LIBRARY_PATH
EFA is not enabled in trn1.32xlarage#
EFA is used as a transport for Collective Communication among multiple instances. EFA must be enabled on the instances used for multi-node training.
OFI plugin initNet() failed is EFA enabled?
Confirm that EFA is enabled by running lspci command and making sure there are eight EFA devices. For example:
[[email protected] ~]$ lspci -tv
-+-[0000:a0]-+-00.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-01.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-19.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1a.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1b.0 Amazon.com, Inc. NeuronDevice
| +-1c.0 Amazon.com, Inc. NeuronDevice
| +-1d.0 Amazon.com, Inc. NeuronDevice
| +-1e.0 Amazon.com, Inc. NeuronDevice
| \-1f.0 Amazon.com, Inc. NVMe SSD Controller
+-[0000:90]-+-00.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-01.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-19.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1a.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1b.0 Amazon.com, Inc. NeuronDevice
| +-1c.0 Amazon.com, Inc. NeuronDevice
| +-1d.0 Amazon.com, Inc. NeuronDevice
| +-1e.0 Amazon.com, Inc. NeuronDevice
| \-1f.0 Amazon.com, Inc. NVMe SSD Controller
+-[0000:20]-+-00.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-01.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-19.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1a.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1b.0 Amazon.com, Inc. NeuronDevice
| +-1c.0 Amazon.com, Inc. NeuronDevice
| +-1d.0 Amazon.com, Inc. NeuronDevice
| +-1e.0 Amazon.com, Inc. NeuronDevice
| \-1f.0 Amazon.com, Inc. NVMe SSD Controller
+-[0000:10]-+-00.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-01.0 Amazon.com, Inc. Elastic Network Adapter (ENA)
| +-19.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1a.0 Amazon.com, Inc. Elastic Fabric Adapter (EFA)
| +-1b.0 Amazon.com, Inc. NeuronDevice
| +-1c.0 Amazon.com, Inc. NeuronDevice
| +-1d.0 Amazon.com, Inc. NeuronDevice
| +-1e.0 Amazon.com, Inc. NeuronDevice
| \-1f.0 Amazon.com, Inc. NVMe SSD Controller
\-[0000:00]-+-00.0 Intel Corporation 440FX - 82441FX PMC [Natoma]
+-01.0 Intel Corporation 82371SB PIIX3 ISA [Natoma/Triton II]
+-01.3 Intel Corporation 82371AB/EB/MB PIIX4 ACPI
+-03.0 Amazon.com, Inc. Device 1111
+-04.0 Amazon.com, Inc. NVMe EBS Controller
\-1f.0 Amazon.com, Inc. NVMe EBS Controller
Launch instances with EFA enabled and try again. If not planning to use the instances for multi-node training or running on trn1.2xlarge, this error message can be ignored.
Communication timeout#
Ranks exchange information during NEFF loading and before the start of the execution. The loading/execution cannot move forward until all ranks are ready.
Timeout waiting for RX (waited 120 sec) - retrying
Timeout waiting for incoming connection (waited 120 sec) - retrying
Connect to localhost:33666 failed - retrying
The communication timeouts are not fatal. The ranks will continue waiting forever. In most case the timeouts are caused by one of the ranks getting delayed, usually be recompilation of a graph. The execution is resumed after the graph is compiled (might take significant amount of time). It is possible to determine if compilation is in progress by checking the logs on all nodes.
Communication timeouts might also indicate that one of the nodes or ranks is hang. If that is the case, terminate the run and restart from the last known good check point.
Communication errors.#
RX, connection closed by remote peer
There could be other similar messages indicating that ranks failed to communicate.
One of the ranks or nodes encountered a problem and terminated. Terminate the run and restart from the last known check point.
EFA Kernel messages (dmesg) after process termination.#
[298850.502143] neuron:npid_detach: neuron:npid_detach: pid=90193, slot=0
[298850.919248] efa 0000:a0:1a.0 rdmap160s26: Failed to process command DEREG_MR (opcode 8) comp_status 7 err -22
When a process that executed Collective Communication terminates it deregisters buffers that were registered with the networking stack. There is a race condition because the Neuron driver deregisters buffers owned by terminating process as part of the memory cleanup. The error is benign and will be removed in the future releases.
Failure to find bootstrap interface#
No interface found in the same subnet as remote address fe80::1461:22ff:fe33:b471<45015>
No usable listening interface found
Bootstrap code incorrectly trying to use link-local IPv6 address for communication. This error will be fixed in the next Neuron release. In the meantime, as a workaround, disable IPv6 on the instances.
sudo sysctl -w net.ipv6.conf.all.disable_ipv6=1
sudo sysctl -w net.ipv6.conf.default.disable_ipv6=1
Name resolution failure#
WARN Invalid NCCL_COMM_ID [compute1-st-kaena-training-0-1.pcluster-trn1-24-pdx80-2n.pcluster:41211], please use format: <ipv4>:<port> or [<ipv6>]:<port>
Verify that the name can be resolved by DNS by using nslookup or dig. Currently released version fails to resolve FQDN longer than 63 characters. This error will be fixed in the upcoming Neuron release. In the mean time use shorter names to ensure that FQDN length does not exceed the maximum of 63 characters.
Usage of Neuron Custom C++ Operators#
Neuron Runtime timeout or GPSIMD exception#
At this point, reset of Neuron Runtime is required after running a model which invoked a Neuron Custom C++ operator. Otherwise, a Neuron Runtime timeout or GPSIMD exception may occur.
Example Neuron Runtime timeout:
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_tpb_status_notifications Missing infer_status notification: (end:1)
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_tpb_status_notifications Missing infer_status notification: (end:2)
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_tpb_status_notifications Missing infer_status notification: (end:3)
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_tpb_status_notifications Missing infer_status notification: (end:4)
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_tpb_status_notifications Missing infer_status notification: (end:0)
2023-Jan-09 20:27:41.0593 15042:15042 ERROR TDRV:exec_consume_infer_status_notifications (FATAL-RT-UNDEFINED-STATE) inference timeout (600000 ms) on Neuron Device 0 NC 0, waiting for execution completion notification
2023-Jan-09 20:27:41.0600 15042:15042 ERROR NMGR:dlr_infer Inference completed with err: 5
Example GPSIMD exception:
2023-Jan-06 22:28:01.0845 137472:137472 ERROR TDRV:pool_stdio_queue_consume_all_entries Printing stderr from GPSIMD:
GPSIMD EXCEPTION OCCURRED: ILLEGAL INSTRUCTION
Subtype/Type/Cause: 0x201
Exception PC: 0x840001E8
Solution#
If either of the above errors are seen, and NEURON_RT_RESET_CORES is set to
0, either unset it or set it to 1. This will enable the default runtime
behaviour of resetting NeuronCores when initializing applications. See
Neuron Runtime Configuration for more information.
Also note that the timeout period can be changed by setting
NEURON_RT_EXEC_TIMEOUT. See Neuron Runtime Configuration for more information.
FI_EFA_FORK_SAFE#
Older Linux (<5.15) kernels require environment variable FI_EFA_FORK_SAFE to be set to 1 for the libfabric to operate correctly. Specifically Amazon Linux 2 uses 5.10 kernel and requires the variable to be set.
When the variable is not set multi-node collective communication will be disabled. Intra-node collective communication is still possible. The following error message will be logged the first time a model containing collective communication is loaded:
Linux kernel 5.10 requires setting FI_EFA_FORK_SAFE=1 environment variable. Multi-node support will be disabled.
Please restart with FI_EFA_FORK_SAFE=1 set."
This document is relevant for: Inf1, Inf2, Trn1, Trn1n