diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_10.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_10.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3ced48b64e9231f39055dc00c98cdfe82b19e766 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_10.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-10 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-10 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-10-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_10.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_110.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_110.yaml new file mode 100644 index 0000000000000000000000000000000000000000..34862dc26f22d21d928b80988c0aeb454f4f5b8e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_110.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-110 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-110 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-110-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_110.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_115.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_115.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dd6791fd96a4d27c760e5a9aff06fffaa804018a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_115.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-115 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-115 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-115-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_115.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_119.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_119.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d85c7810ea21e9485e284488164951ead2d1bd48 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_119.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-119 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-119 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-119-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_119.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_142.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_142.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b6563b2491dd1a971c95eb2650b976078e676754 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_142.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-142 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-142 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-142-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_142.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_152.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_152.yaml new file mode 100644 index 0000000000000000000000000000000000000000..af11134f02daf85a7933a87177ab0398cc5defb7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_152.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-152 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-152 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-152-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_152.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_160.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_160.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f519cb9d6c7c8bdf71aef5598977e5f54feeeacd --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_160.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-160 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-160 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-160-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_160.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_18.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_18.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a68a0a863c7d67165e5a75f58cbbc6a4ad84bfbd --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_18.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-18 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-18 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-18-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_18.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_184.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_184.yaml new file mode 100644 index 0000000000000000000000000000000000000000..179a314da4bec940ab60ece0ddf4fb40639dc724 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_184.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-184 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-184 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-184-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_184.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_190.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_190.yaml new file mode 100644 index 0000000000000000000000000000000000000000..470ab1b6ed18fadb0e4452cf7674f2f06dc8cbaf --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_190.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-190 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-190 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-190-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_190.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_192.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_192.yaml new file mode 100644 index 0000000000000000000000000000000000000000..243eb0f02f00419c72684df38ef2d3787b4196b6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_192.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-192 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-192 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-192-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_192.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_199.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_199.yaml new file mode 100644 index 0000000000000000000000000000000000000000..38bf1a14c3c347027528104ea8374733af7088d3 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_199.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-199 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-199 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-199-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_199.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_21.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_21.yaml new file mode 100644 index 0000000000000000000000000000000000000000..153d55f4255504614b9b2d82926e17bb5a068358 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_21.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-21 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-21 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-21-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_21.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_22.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_22.yaml new file mode 100644 index 0000000000000000000000000000000000000000..57f22a737ff7fce4965dcc004d6ba88bb8265e20 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_22.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-22 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-22 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-22-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_22.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_226.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_226.yaml new file mode 100644 index 0000000000000000000000000000000000000000..92d10fa83ba8936ff01c9587f3eb214a97a8c04f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_226.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-226 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-226 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-226-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_226.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_244.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_244.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a596d440eb58af5960a0962b4127b7496f00bdda --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_244.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-244 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-244 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-244-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_244.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_252.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_252.yaml new file mode 100644 index 0000000000000000000000000000000000000000..64baebedaaaf707937d303ff6f1bf55e54b92f3e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_252.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-252 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-252 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-252-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_252.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_268.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_268.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de12356bc8b50a8d23b928221b7b13e8c03c8db9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_268.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-268 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-268 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-268-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_268.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_277.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_277.yaml new file mode 100644 index 0000000000000000000000000000000000000000..935909199b0e49e20442ccbc8b91d4bce52360ed --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_277.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-277 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-277 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-277-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_277.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_281.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_281.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f7e4640c4e76c019504011c408b830fc84be8227 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_281.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-281 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-281 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-281-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_281.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_304.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_304.yaml new file mode 100644 index 0000000000000000000000000000000000000000..12cd5dbf2b01c90fa5ea7a1deeee2aac200dd157 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_304.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-304 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-304 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-304-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_304.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_305.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_305.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8fdac4e19c8e2ea52151f148505fbd96a44fc729 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_305.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-305 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-305 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-305-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_305.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_310.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_310.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d3750b5922af40440af6392a8baf593dc344dc91 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_310.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-310 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-310 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-310-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_310.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_312.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_312.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0d82f92e492932063781c6f98d3d437542b3d971 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_312.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-312 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-312 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-312-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_312.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_315.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_315.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ba98c537cdb71cc07c5762bd88cd533240d93570 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_315.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-315 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-315 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-315-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_315.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_322.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_322.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9af84f28178f042569312e896e9093b1d992d8fa --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_322.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-322 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-322 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-322-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_322.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_338.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_338.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dcf7760ab4134460b1e88fb9379b510ef9dd25da --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_338.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-338 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-338 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-338-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_338.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_34.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_34.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c5de114ad209c4a269201d9b43a74d6f1387702d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_34.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-34 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-34 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-34-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_34.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_350.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_350.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d047114bac21ab3eaa7b32ff7502077d51f4df8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_350.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-350 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-350 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-350-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_350.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_361.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_361.yaml new file mode 100644 index 0000000000000000000000000000000000000000..40506148210fdf151c8f579a3621a3beb019cdd4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_361.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-361 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-361 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-361-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_361.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_362.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_362.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d8ea7b757e91814a88461c5d66ceccac435188aa --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_362.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-362 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-362 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-362-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_362.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_38.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_38.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a2426dc0d58b67898103f8080c38dcf38f16b3b5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_38.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-38 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-38 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-38-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_38.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_380.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_380.yaml new file mode 100644 index 0000000000000000000000000000000000000000..21e14143dadcc630edf0020a3d43fbdbd9e2b159 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_380.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-380 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-380 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-380-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_380.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_385.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_385.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f0b3d72b5a364b0feba2ff056dacb3aeaec54f6d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_385.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-385 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-385 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-385-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_385.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_413.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_413.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe1582b3d3b3215493045ba0fcde6fcffa6ec1c4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_413.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-413 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-413 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-413-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_413.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_435.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_435.yaml new file mode 100644 index 0000000000000000000000000000000000000000..88cb15a5ab80a01ccfbc7f1448326553d7b426b4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_435.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-435 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-435 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-435-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_435.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_44.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_44.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d0b936853a693f8c8123271476bc1bdf5dd3809b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_44.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-44 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-44 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-44-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_44.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_441.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_441.yaml new file mode 100644 index 0000000000000000000000000000000000000000..105412a3209d3185f6f1a2fc78267d1cfabb7dfe --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_441.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-441 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-441 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-441-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_441.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_444.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_444.yaml new file mode 100644 index 0000000000000000000000000000000000000000..63fe2152645771f4638d3a6bde877e027470f006 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_444.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-444 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-444 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-444-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_444.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_452.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_452.yaml new file mode 100644 index 0000000000000000000000000000000000000000..84af0f84c44cdd91e55d15dc3beea5014bc7912d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_452.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-452 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-452 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-452-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_452.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_46.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_46.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ace0f3f1c7a79c7a912dac3f5c6718549c6a5341 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_46.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-46 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-46 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-46-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_46.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_471.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_471.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f4e2963b97a7d26c3f24fc7d76bf0393a1e686e5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_471.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-471 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-471 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-471-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_471.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_473.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_473.yaml new file mode 100644 index 0000000000000000000000000000000000000000..909f8f6a9bd4263e6be3c8f04496f12445be4e61 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_473.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-473 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-473 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-473-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_473.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_476.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_476.yaml new file mode 100644 index 0000000000000000000000000000000000000000..81f931a10d60d809632b923a8c788749ecc45fa2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_476.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-476 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-476 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-476-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_476.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_487.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_487.yaml new file mode 100644 index 0000000000000000000000000000000000000000..633ffb053150aced068713c83888a258ae287c3a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_487.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-487 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-487 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-487-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_487.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_515.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_515.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3989d5d420e333169124c62a937348a3f72ed412 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_515.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-515 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-515 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-515-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_515.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_517.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_517.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f44a4bb1f4381b3c97783e9d4c72f6b159ad4f61 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_517.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-517 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-517 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-517-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_517.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_525.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_525.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1e2a907dad5675042c824a1d794c31b4d6a95356 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_525.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-525 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-525 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-525-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_525.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_526.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_526.yaml new file mode 100644 index 0000000000000000000000000000000000000000..aaa33b4e665e7fa507858661ce9912b000dcfb4e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_526.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-526 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-526 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-526-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_526.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_53.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_53.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2c74b6058f28cbe73263f1a220a572a0e70e1140 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_53.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-53 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-53 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-53-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_53.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_535.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_535.yaml new file mode 100644 index 0000000000000000000000000000000000000000..be8814ce949da425758d9959c606bbd12080f4d3 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_535.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-535 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-535 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-535-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_535.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_556.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_556.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cb1a7ce2c48d7ff880d59d602908615735b344e7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_556.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-556 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-556 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-556-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_556.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_571.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_571.yaml new file mode 100644 index 0000000000000000000000000000000000000000..69265f0af8384bc304d358e0559b1f124fd94d03 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_571.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-571 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-571 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-571-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_571.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_588.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_588.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f32bb19e060af44487f48d64489fd9521a70716e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_588.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-588 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-588 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-588-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_588.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_592.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_592.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5390f73dd824620fd9df1c818304c01090fcdfb6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_592.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-592 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-592 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-592-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_592.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_593.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_593.yaml new file mode 100644 index 0000000000000000000000000000000000000000..509a6cf0ae244961fa6f53fd1999845455ec2f84 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_593.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-593 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-593 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-593-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_593.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_602.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_602.yaml new file mode 100644 index 0000000000000000000000000000000000000000..93d0f93c9e71f331402890d6cae98f6b1ebc2ca3 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_602.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-602 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-602 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-602-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_602.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_603.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_603.yaml new file mode 100644 index 0000000000000000000000000000000000000000..086fcce9563882c8ab51ea01942b7dfd8a8d273e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_603.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-603 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-603 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-603-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_603.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_608.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_608.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e7c55ddded9828d154570392a5bd513c7a3051f6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_608.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-608 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-608 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-608-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_608.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_63.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_63.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c8b086084721e5dc8084c06b1d14a20e3d8b08e5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_63.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-63 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-63 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-63-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_63.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_637.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_637.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6c1b6c88a4b5ae4a709674f5192fe0e5ae44b8e7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_637.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-637 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-637 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-637-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_637.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_642.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_642.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5b80024b967276ba00bb09eff6941605a5c2a1cc --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_642.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-642 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-642 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-642-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_642.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_713.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_713.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dacf2198df246ea9ee5d05c6fbc5eaf26cd6d856 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_713.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-713 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-713 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-713-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_713.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_72.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_72.yaml new file mode 100644 index 0000000000000000000000000000000000000000..62b9195f9fac6cd41b14064a374cc61a6a770c28 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_72.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-72 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-72 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-72-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_72.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_722.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_722.yaml new file mode 100644 index 0000000000000000000000000000000000000000..736d4459408b31e17f68306730b83f765e4e3d3b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_722.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-722 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-722 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-722-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_722.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_73.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_73.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f36c7d5ebe8a6eff5a3551e7c6f90259c8006a4d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_73.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-73 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-73 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-73-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_73.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_747.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_747.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cb3f1ce8520e4a16f62126427d88a8655b20442b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_747.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-747 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-747 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-747-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_747.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_751.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_751.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6a47a93c4a8e00a6b07d98c55f9f51f90c26569b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_751.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-751 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-751 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-751-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_751.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_770.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_770.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7b56be2be73594efd0d028d702f9d8d588abbac5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_770.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-770 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-770 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-770-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_770.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_790.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_790.yaml new file mode 100644 index 0000000000000000000000000000000000000000..57552ea6dc325e28befac7b26c6805c4c32335e7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_790.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-790 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-790 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-790-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_790.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_797.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_797.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6fd2bf52d320fcab2103a7d04d7aadc3eed19c28 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_797.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-797 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-797 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-797-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_797.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_824.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_824.yaml new file mode 100644 index 0000000000000000000000000000000000000000..87ab27f913445fcffcb47c061abaa395710544d0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_824.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-824 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-824 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-824-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_824.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_833.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_833.yaml new file mode 100644 index 0000000000000000000000000000000000000000..14e687ac7b41be921166d259514a5c3cf55eb82e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_833.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-833 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-833 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-833-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_833.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_84.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_84.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5ed778bf49754f4c52fb2407f3b0bd9d641f6e8e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_84.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-84 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-84 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-84-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_84.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_9.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_9.yaml new file mode 100644 index 0000000000000000000000000000000000000000..13b8cb45fceeac0079bdac5b485d519fae0cba6f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_9.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-9 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-9 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-9-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_9.sh + HOSTSFILE=$OMPI_MCA_orte_default_hostfile; + echo "HOSTSFILE=$HOSTSFILE"; + MASTER_ADDR="$(head -n 1 $HOSTSFILE | sed -n s/[[:space:]]slots.*//p)"; + NUM_NODES=$(wc -l < $HOSTSFILE); + CARDS_PER_NODE=8; + N_CARDS=$((NUM_NODES*CARDS_PER_NODE)); + echo "MPI_ROOT=$MPI_ROOT"; + echo "N_CARDS=$N_CARDS"; + echo "MASTER_ADDR=$MASTER_ADDR"; + sleep 20; + + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $MEGATRON_SETUP_CMD; + + mpirun -np $N_CARDS -npernode 8 \ + --tag-output \ + --allow-run-as-root \ + --prefix $MPI_ROOT \ + -x WORKER_DIR=$WORKER_DIR \ + -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock