To identify number of Inferentia devices in a given instance use the neuron-ls
.
$ neuron-ls
+--------------+---------+--------+-----------+-----------+------+------+
| PCI BDF | LOGICAL | NEURON | MEMORY | MEMORY | EAST | WEST |
| | ID | CORES | CHANNEL 0 | CHANNEL 1 | | |
+--------------+---------+--------+-----------+-----------+------+------+
| 0000:00:1f.0 | 0 | 4 | 4096 MB | 4096 MB | 0 | 0 |
+--------------+---------+--------+-----------+-----------+------+------+
The above output is taken Inf1.xlarge instance. The first column shows the PCI Bus Device Function ID. The second column shows logical ID assigned to the device. This logical ID is used during multiple Neuron-rtd configuration. The third columns shows the number of NeuronCores in the inferentia device. The last two columns shows the connection to any other inferentia devices; since this is a single inferentia device, those are empty.
Multiple NeuronCores(NC) can be combined to form a NeuronCore Group (NCG). Neuron framework layer will automatically create a default NeuronCore Group. To view list of available NCGs the following command can be used.
$ neuron-cli list-ncg
Device 1 NC count 4
+-------+----------+--------------------+----------------+
| NCG ID| NC COUNT | DEVICE START INDEX | NC START INDEX |
+-------+----------+--------------------+----------------+
| 1 | 1 | 0 | 0 |
| 2 | 1 | 0 | 1 |
| 3 | 2 | 0 | 2 |
+-------+-----------------+----------+-----------------+----------------+
If for some reason, need to delete the framework created NCGs neuron-cli destroy-ncg
can be used.
Models can be loaded into a NCG. Multiple models can be loaded into a single NCG but only one can be in STARTED state at a given moment. Inference can be done only on the models with STARTED state.
To view all the models neuron-cli list-model
can be used.
$ neuron-cli list-model
Found 3 models
10003 MODEL_STATUS_LOADED 1
10001 MODEL_STATUS_STARTED 1
10002 MODEL_STATUS_STARTED 1
In the above output 10001 and 10002 are unique identifier for models loaded in Neuron device.
To start/stop/unload a model neuron-cli start/stop/unload
command can be used.
Each model loaded consumes different amount of memory (host and device), NeuronCore and CPU usage. This usage can be viewed by using neuron-top.
$ neuron-top
neuron-top - 2019-11-13 23:57:08
NN Models: 3 total, 2 running
Number of VNCs tracked: 2
0000:00:1f.0 Utilizations: Neuron core0 0.00, Neuron core1 0.00, Neuron core2 0.00, Neuron core3 0.00,
DLR Model Node ID Subgraph Exec. Unit Host Mem Device Mem Neuron core %
10003 1 0 0000:00:1f.0:0 384 135660544 0.00
10001 3 0 0000:00:1f.0:0 384 67633152 0.00
10002 1 0 0000:00:1f.0:1 384 135660544 0.00
Please see Neuron Gatherinfo