diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_100.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_100.yaml new file mode 100644 index 0000000000000000000000000000000000000000..72cbac8217e7a66f616bc03bd86b0047015a7168 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_100.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-100 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-100 +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-100-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="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_100.sh + HOSTSFILE=$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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_101.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_101.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b98e7ea31dc0be5297ee2d4db6798b5b8494c22e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_101.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-101 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-101 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-101-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_101.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_120.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_120.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5fbb29e4220d884af98d697fe787b7ed9ab29839 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_120.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-120 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-120 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-120-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_120.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_133.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_133.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7db62ef7e848a4ad8f249c1fcd866a2247379839 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_133.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-133 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-133 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-133-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_133.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_149.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_149.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a49c7dd0cb6e151037a34a17527a462d00611893 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_149.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-149 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-149 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-149-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_149.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_153.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_153.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1db6a3997474e6737054e82e60ed9e6f22655ecd --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_153.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-153 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-153 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-153-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_153.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_158.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_158.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c726525b290affea3fc9a56540f54b4d2e09ca0b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_158.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-158 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-158 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-158-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_158.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_20.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_20.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6862c5fcf215c70584211e49839782249752c722 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_20.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-20 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-20 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-20-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_20.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_204.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_204.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe8775d22644b551752fb2a79ffae97b1e4982ae --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_204.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-204 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-204 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-204-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_204.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_209.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_209.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b8c3105fc69c21d89fc506528d80609c82eb1c37 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_209.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-209 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-209 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-209-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_209.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_213.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_213.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eb7e7a4ed945adff0feec20955bc02a68e6fdf0e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_213.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-213 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-213 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-213-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_213.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_230.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_230.yaml new file mode 100644 index 0000000000000000000000000000000000000000..477e530b1e92d6ae0078ceeff99233006a1319eb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_230.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-230 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-230 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-230-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_230.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_231.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_231.yaml new file mode 100644 index 0000000000000000000000000000000000000000..36ccd193c5fa6110d162701df51ce1d17c999170 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_231.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-231 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-231 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-231-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_231.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_234.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_234.yaml new file mode 100644 index 0000000000000000000000000000000000000000..52147be059da18b244fd7358eae74067b21c2893 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_234.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-234 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-234 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-234-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_234.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_236.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_236.yaml new file mode 100644 index 0000000000000000000000000000000000000000..64ad8d20c2461d51b099893341e8e985fc6032b8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_236.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-236 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-236 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-236-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_236.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_240.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_240.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f8056eac9bab10f354e146a9461d0f36f8510a8c --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_240.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-240 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-240 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-240-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_240.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_242.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_242.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5f4a71d6606ade97c9f8949e798c0c4757434235 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_242.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-242 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-242 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-242-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_242.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_243.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_243.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2af17140798aafdb703942a0fb5a6063b1d57072 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_243.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-243 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-243 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-243-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_243.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_298.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_298.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8bae1ab98bbf402f1df8824e680108f12b09f92a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_298.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-298 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-298 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-298-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_298.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_308.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_308.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9f89dddb88795ed4f35373c7cdc919e42893b808 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_308.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-308 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-308 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-308-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_308.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_321.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_321.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bf3414daacc1badbc261eda98215f22cb885986d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_321.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-321 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-321 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-321-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_321.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_328.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_328.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3b036c75acd77c3df7c049c3ab51e395490e7c98 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_328.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-328 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-328 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-328-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_328.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_369.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_369.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2249e928889f07184c3b84318dfca6765f0332f8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_369.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-369 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-369 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-369-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_369.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_387.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_387.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5d5614f7f259fd1c895dbbe8777bea9f6b7f1dda --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_387.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-387 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-387 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-387-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_387.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_388.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_388.yaml new file mode 100644 index 0000000000000000000000000000000000000000..160c3804cd9dbc5e25d2bb0df0c838e454c736e4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_388.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-388 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-388 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-388-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_388.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_393.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_393.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6a4cb8568e02ec898533498a2059fe0972bb5694 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_393.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-393 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-393 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-393-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_393.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_396.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_396.yaml new file mode 100644 index 0000000000000000000000000000000000000000..33a9ad5f9f416e3ccfbf2c5231901eae3904e6f5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_396.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-396 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-396 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-396-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_396.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_43.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_43.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d21fd9ad7af4be3bd588b3381fc5d8a9c4461673 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_43.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-43 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-43 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-43-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_43.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_433.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_433.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2357b211315044a40dee61757187fe2fda97064e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_433.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-433 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-433 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-433-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_433.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_450.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_450.yaml new file mode 100644 index 0000000000000000000000000000000000000000..558072109116c6e0bf2dc92aeb0d71fab1c266e1 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_450.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-450 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-450 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-450-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_450.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_451.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_451.yaml new file mode 100644 index 0000000000000000000000000000000000000000..086c4d887eb34006ac5071d64089e56153e1e808 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_451.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-451 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-451 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-451-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_451.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_477.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_477.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8973591f7a8bc1bbb4365ddbf10748d7a7456503 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_477.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-477 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-477 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-477-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_477.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_481.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_481.yaml new file mode 100644 index 0000000000000000000000000000000000000000..429d489050143382f0a86a0c378d8345eeea61f0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_481.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-481 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-481 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-481-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_481.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_521.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_521.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6e0752ed01b5d0d66808c772d2cd8f2e89e64d16 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_521.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-521 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-521 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-521-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_521.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_532.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_532.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eee4dabf187637164633992819faa6e8709c93b4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_532.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-532 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-532 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-532-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_532.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_546.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_546.yaml new file mode 100644 index 0000000000000000000000000000000000000000..79abd133fe091581d6807704a61a72b7492f1e26 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_546.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-546 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-546 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-546-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_546.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_575.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_575.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a9f1d2d7edb05f9255ac753a5050d215e374ea41 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_575.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-575 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-575 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-575-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_575.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_58.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_58.yaml new file mode 100644 index 0000000000000000000000000000000000000000..81346e778ca923020d95dcf91f4477828fbd011b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_58.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-58 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-58 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-58-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_58.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_582.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_582.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0e44d2a10096e3ae6326dbc022b7ac332905e1ca --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_582.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-582 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-582 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-582-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_582.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_587.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_587.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c3338fe291a2f119d6b4189f3d68a0fc2001b069 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_587.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-587 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-587 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-587-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_587.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_589.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_589.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a50baf732b26ade7001ee54a959c0d62c9d48e23 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_589.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-589 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-589 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-589-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_589.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_604.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_604.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b59e7a59db466d13ceda9a9f4c9cdd0c398d1b32 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_604.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-604 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-604 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-604-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_604.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_619.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_619.yaml new file mode 100644 index 0000000000000000000000000000000000000000..75d6da09f44c90adbbdc39fa309d0584b77ce181 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_619.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-619 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-619 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-619-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_619.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_620.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_620.yaml new file mode 100644 index 0000000000000000000000000000000000000000..315c69db89fc5b4bcf5beb7b33330237707aea48 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_620.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-620 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-620 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-620-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_620.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_625.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_625.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fa20219393d3f18be697e659fcf4f212da5d0bac --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_625.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-625 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-625 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-625-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_625.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_626.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_626.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f01231c9600d33c4ccb639d773cd35d80d56da5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_626.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-626 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-626 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-626-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_626.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_632.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_632.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4567e0350d5a2bc6d310556ebd9c8def2a6490be --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_632.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-632 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-632 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-632-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_632.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_636.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_636.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7839131c8dc4d059fe716657b97ff7df1f2b722d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_636.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-636 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-636 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-636-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_636.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_643.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_643.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6ea0c6a08a23ff3d9c66f609c80eeca46c4915ab --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_643.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-643 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-643 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-643-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_643.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_649.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_649.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a4ca9935a211026584df22eecb3a070a0c2e37de --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_649.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-649 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-649 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-649-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_649.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_65.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_65.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9ec1c962c8b1e64627182d0532391fed5a7ae146 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_65.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-65 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-65 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-65-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_65.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_655.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_655.yaml new file mode 100644 index 0000000000000000000000000000000000000000..38d25043ea7ebc07b577123e1a1a9b39373e3712 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_655.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-655 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-655 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-655-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_655.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_659.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_659.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9d8440708cf0880097de83651107453e0dda3490 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_659.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-659 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-659 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-659-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_659.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_662.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_662.yaml new file mode 100644 index 0000000000000000000000000000000000000000..983ed7d5e91597ee031551f0805a3a1d1382d240 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_662.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-662 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-662 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-662-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_662.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_683.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_683.yaml new file mode 100644 index 0000000000000000000000000000000000000000..331f5c9b2cd1f64ddf561db63d027546516a99c9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_683.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-683 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-683 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-683-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_683.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_690.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_690.yaml new file mode 100644 index 0000000000000000000000000000000000000000..565380705693af958f5040087c96c9be212e4d4e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_690.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-690 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-690 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-690-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_690.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_691.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_691.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b4621de75ff04836d476834d6cef7ac193f558cb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_691.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-691 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-691 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-691-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_691.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_692.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_692.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c77d368d1fa24c9d84235f32fb07e8018379b064 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_692.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-692 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-692 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-692-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_692.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_694.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_694.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f89fad01a7f8c90ea20f01622e1afddbceef65b0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_694.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-694 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-694 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-694-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_694.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_7.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_7.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1e2ee26ee07bc7ebf76d3bbd6bfdb61e28719d35 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_7.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-7 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-7 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-7-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_7.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_700.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_700.yaml new file mode 100644 index 0000000000000000000000000000000000000000..115d8c78e9b85ca604ab5eadc94c285d30b7f0a7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_700.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-700 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-700 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-700-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_700.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_71.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_71.yaml new file mode 100644 index 0000000000000000000000000000000000000000..670705bd7ef282ce9f451c18904c1f24bb6ed4e9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_71.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-71 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-71 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-71-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_71.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_729.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_729.yaml new file mode 100644 index 0000000000000000000000000000000000000000..06f86fac550f324f6ae27709e68b240e29382d33 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_729.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-729 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-729 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-729-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_729.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_737.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_737.yaml new file mode 100644 index 0000000000000000000000000000000000000000..464a83eff3fb5620d57bdcbf295af473d00b41cb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_737.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-737 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-737 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-737-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_737.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_748.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_748.yaml new file mode 100644 index 0000000000000000000000000000000000000000..97ff117b595e4358a81324f58c066f26ee501176 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_748.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-748 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-748 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-748-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_748.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_756.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_756.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ef9fee8a46b389a0c46fb95425a7437996a780a2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_756.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-756 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-756 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-756-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_756.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_760.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_760.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c846f4e78130d95aa91b1c2f17f4df2b0e84d865 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_760.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-760 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-760 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-760-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_760.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_791.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_791.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0b5d776405833776754711dafd05b815885acb3a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_791.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-791 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-791 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-791-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_791.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_796.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_796.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1d8c2e929b25eccd316614de236cf8471d45b7c5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_796.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-796 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-796 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-796-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_796.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_814.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_814.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ecac5d078ce50ab34d64c39d3c5ba504ba9365b2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_814.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-814 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-814 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-814-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_814.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_818.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_818.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0db0b33dc6546357492e74f5be71f104f727dd67 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_818.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-818 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-818 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-818-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_818.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_825.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_825.yaml new file mode 100644 index 0000000000000000000000000000000000000000..54496763f89072ebaba19af18f156e502cfc539e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_825.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-825 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-825 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-825-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_825.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_83.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_83.yaml new file mode 100644 index 0000000000000000000000000000000000000000..015bd7bf43bdc54e94638be56c97079be9256deb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_83.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-83 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-83 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-83-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_83.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_830.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_830.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d31cccc3270d07bad3e273508a5a378fbd3729cd --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_830.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-830 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-830 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-830-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_830.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_97.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_97.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0554484156687eeae3a27ba605d35b4ca9f68a6d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_97.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-97 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-97 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-97-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_97.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