I am trying to build deep neural network models natural for language processing models like GPT2 or BERT.
The request for GPUS_ALL_REGIONS is partially approved for 4 GPUs, I want to use multiple GPUs ( at most 4 ones ) in a single instance.
However, all requests which assign more than 2 PREEMPTIBLE_NVIDIA_{A100 or V100}_GPUS into a single region are denied.
How could I obtain multiple GPUs’ quota into a single region?
What did you try?:
-
Checked billing account and credit card is correctly registered.
-
Confirmed not to log-in as a free trial user.
-
Repeated to send Quota Request many times for various regions and various GPU types(A100, V100, T4, P100)
-
Confirmed to log-in
gcloud auth log-in
as the account of 1.
What’s the Expected Status?:
The number of GPUs “1” is increased into 4 for each single region.
GPUS_ALL_REGIONS is increased into 4 times of the number of regions.
The following gcloud-cli would be successfully executed with ACCELERATOR==”count=”${NUM_GPUS}”,type=nvidia-tesla-a100″
gcloud compute instances create ${VM_NAME} --preemptible
--zone ${ZONE}
--machine-type ${MACHINE_TYPE}
--network-interface=${NETWORK_INTERFACE}
--maintenance-policy=TERMINATE
--provisioning-model=SPOT
--service-account="**********@developer.gserviceaccount.com"
--accelerator=${ACCELERATOR}
--enable-display-device --tags=http-server,https-server
--create-disk=auto-delete=yes,boot=yes,device-name=instance-1,image=projects/debian-cloud/global/images/debian-11-bullseye-v20221206,mode=rw,size=${BOOT_DISK_SIZE},type=projects/${PROJECT_NAME}/zones/${ZONE}/diskTypes/${BOOT_DISK_TYPE}
--image-project ${IMAGE_PROJECT}
--image-family ${IMAGE_FAMILY}
--no-shielded-secure-boot --shielded-vtpm --shielded-integrity-monitoring --reservation-affinity=any
--threads-per-core=2 --visible-core-count=6