diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_109.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_109.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7ccfbbbe8b9b02e78d88bfd22a21424930caa927 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_109.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-109 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-109 +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-109-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="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_109.sh + HOSTSFILE=$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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_125.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_125.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2de688a2f003c9adc7f2017c39dd77ef23187aa4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_125.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-125 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-125 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-125-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_125.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_126.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_126.yaml new file mode 100644 index 0000000000000000000000000000000000000000..166fce37d7d354e62596e2d5f2bed739c045e0ef --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_126.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-126 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-126 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-126-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_126.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_127.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_127.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3fcfec62cf4237fe4c01ff7e8ec1788d76c71487 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_127.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-127 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-127 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-127-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_127.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_14.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_14.yaml new file mode 100644 index 0000000000000000000000000000000000000000..987acdcc2ae02f44aaf30ac90f649a8f3e986071 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_14.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-14 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-14 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-14-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_14.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_151.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_151.yaml new file mode 100644 index 0000000000000000000000000000000000000000..43e643508e22b2209101732901b40648c5254a34 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_151.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-151 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-151 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-151-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_151.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_161.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_161.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f292af99ebdd276a1ab00e0bde16eb2147d79fbe --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_161.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-161 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-161 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-161-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_161.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_168.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_168.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e00a99fc7d3d037f6952c05842046ae1d11b7b0b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_168.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-168 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-168 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-168-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_168.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_174.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_174.yaml new file mode 100644 index 0000000000000000000000000000000000000000..efc25ec044a9a7aa9192780fcbf1011fb2b68331 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_174.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-174 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-174 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-174-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_174.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_185.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_185.yaml new file mode 100644 index 0000000000000000000000000000000000000000..68401733570ba28f717591347495cd6941b97c46 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_185.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-185 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-185 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-185-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_185.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_189.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_189.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d94beb539cc4179624fc71cff09b0faff16fab0b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_189.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-189 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-189 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-189-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_189.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_19.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_19.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c24bf0f7e01e6a6c0ee77997e4e4a4c2af356b0f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_19.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-19 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-19 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-19-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_19.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_202.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_202.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f51ad15c906026144a1a89480dca47e13374e418 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_202.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-202 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-202 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-202-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_202.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_208.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_208.yaml new file mode 100644 index 0000000000000000000000000000000000000000..10477b0884ce1a4c1fd9cdb0e60396f2bf3e158d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_208.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-208 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-208 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-208-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_208.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_216.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_216.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a700c75995bc28ebc7c6ebc65ed6af6db51edc86 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_216.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-216 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-216 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-216-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_216.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_223.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_223.yaml new file mode 100644 index 0000000000000000000000000000000000000000..39d05139d4ed6806dea168fd23929af8ba8db32e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_223.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-223 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-223 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-223-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_223.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_228.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_228.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fd49350631dbc37b69bcc3575c75b637f1fedb9e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_228.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-228 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-228 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-228-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_228.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_247.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_247.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c2321fb14198036a06a2dbec28b18b7b954d5fb9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_247.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-247 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-247 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-247-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_247.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_248.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_248.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ff25ed16b0f020f6bfe25bc874d740fabd7c8e54 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_248.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-248 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-248 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-248-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_248.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_25.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_25.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8b220a2c230341ece435c81d2ca47f4042e7213a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_25.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-25 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-25 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-25-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_25.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_251.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_251.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe36652d345849f7c120fe23adc3fc7154abe6cb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_251.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-251 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-251 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-251-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_251.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_261.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_261.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eb257bdd0b649fc966b202fdec8ced416722044e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_261.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-261 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-261 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-261-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_261.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_264.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_264.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ff1ee01a5245703e654f6ac6dc6ce581eaadc1a8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_264.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-264 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-264 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-264-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_264.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_273.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_273.yaml new file mode 100644 index 0000000000000000000000000000000000000000..eb2af1899a66b9824a18992a78e47c4629d922d2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_273.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-273 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-273 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-273-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_273.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_297.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_297.yaml new file mode 100644 index 0000000000000000000000000000000000000000..504ac63e8e4eec863d8bba0e5108cefbfc84d3b0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_297.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-297 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-297 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-297-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_297.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_301.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_301.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4a42a89b00f907143861394f2490c9f08833af40 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_301.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-301 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-301 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-301-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_301.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_316.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_316.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8e3c5c8fbb1fac188e1ef67918ed499d05bf96c9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_316.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-316 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-316 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-316-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_316.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_335.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_335.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fab066fe755dfe018e24e886fe6ebf6ee97cd8e1 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_335.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-335 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-335 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-335-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_335.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_339.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_339.yaml new file mode 100644 index 0000000000000000000000000000000000000000..925b8ca0d6e61f2ffb0fb0ea0eca1cc42e9d9529 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_339.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-339 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-339 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-339-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_339.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_340.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_340.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6a9ee94873331a7ec404568360a87562d7c28f43 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_340.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-340 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-340 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-340-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_340.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_355.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_355.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6c9b896a3b03f9cbc23309be007fc69c23aa409c --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_355.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-355 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-355 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-355-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_355.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_359.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_359.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ed2311b09589c0ad4a5a3a3eb4b9b45ec45088b5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_359.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-359 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-359 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-359-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_359.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_39.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_39.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e0a8b75d3a27753539ff749596cd931ed631e172 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_39.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-39 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-39 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-39-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_39.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_408.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_408.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0710105af4c82e1936e5c071c91e280154bdb6af --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_408.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-408 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-408 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-408-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_408.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_428.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_428.yaml new file mode 100644 index 0000000000000000000000000000000000000000..548aa9c7f97d759532715194bcbe70164254b34a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_428.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-428 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-428 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-428-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_428.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_440.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_440.yaml new file mode 100644 index 0000000000000000000000000000000000000000..02ce36daef3b1ab3bd0ab89f40672b77de26f487 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_440.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-440 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-440 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-440-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_440.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_489.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_489.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e3e63437e266312f407a5320c08da97ed6e2b222 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_489.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-489 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-489 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-489-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_489.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_49.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_49.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7f9b069c182c1e13441e435ee4d9bab77842e3cd --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_49.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-49 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-49 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-49-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_49.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_508.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_508.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0334b1852dab26179a6ba712a58b886e04633fbe --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_508.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-508 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-508 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-508-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_508.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_52.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_52.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe1200551ab366762347d5580960bc33e074e34d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_52.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-52 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-52 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-52-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_52.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_530.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_530.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b79a06b02a87d05847f93ed2c732b35b24bce0d9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_530.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-530 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-530 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-530-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_530.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_539.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_539.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ab1bad2d865947af95454cd14d0568d8ef3752d6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_539.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-539 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-539 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-539-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_539.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_54.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_54.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1ed716cd1ce4cede6342955309bb83640f330f1f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_54.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-54 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-54 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-54-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_54.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_541.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_541.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7b457c2132fb17dacb58bdd933f6747ae571b093 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_541.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-541 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-541 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-541-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_541.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_543.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_543.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f58bd04194a56c04951272cc6a64325466b554d8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_543.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-543 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-543 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-543-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_543.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_550.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_550.yaml new file mode 100644 index 0000000000000000000000000000000000000000..69ec5a2483f2459a6008464b09cc87cb2d338a36 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_550.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-550 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-550 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-550-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_550.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_551.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_551.yaml new file mode 100644 index 0000000000000000000000000000000000000000..798924a094255a6a185bbf75ce3c2440f2829415 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_551.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-551 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-551 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-551-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_551.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_567.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_567.yaml new file mode 100644 index 0000000000000000000000000000000000000000..33bae93e9011713e1176a4adc29c705bcbe1b0e6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_567.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-567 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-567 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-567-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_567.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_576.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_576.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a87941b9c0bca62de43e6ae7a3e3ed3386ea7a02 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_576.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-576 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-576 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-576-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_576.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_59.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_59.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f505abd19f62b0ec9ebce682b052f015d0a03873 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_59.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-59 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-59 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-59-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_59.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_590.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_590.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2a49117b4c224500f744272eb3cc345d4a8ce9ed --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_590.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-590 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-590 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-590-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_590.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_591.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_591.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ca39752722d25eb1e56ac9b0ce7b55ea7d3d4ad0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_591.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-591 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-591 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-591-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_591.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_601.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_601.yaml new file mode 100644 index 0000000000000000000000000000000000000000..db5ac8ea95083e99a28d34c3cce6ed240ca93765 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_601.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-601 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-601 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-601-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_601.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_612.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_612.yaml new file mode 100644 index 0000000000000000000000000000000000000000..30a67ce6a275db98ad8aaa2cb91cab0b230f0bb8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_612.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-612 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-612 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-612-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_612.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_616.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_616.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d66254c3157dfcf1b77a1d25e63464685b94e07e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_616.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-616 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-616 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-616-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_616.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_617.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_617.yaml new file mode 100644 index 0000000000000000000000000000000000000000..46c3530420f83da182f0bdc6253e899fe87d7776 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_617.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-617 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-617 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-617-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_617.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_62.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_62.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2bea14c043535bf7d0c073d28c589323e14f964a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_62.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-62 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-62 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-62-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_62.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_64.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_64.yaml new file mode 100644 index 0000000000000000000000000000000000000000..795b600f49c72e8edd893be5945e14caac9985ca --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_64.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-64 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-64 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-64-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_64.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_652.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_652.yaml new file mode 100644 index 0000000000000000000000000000000000000000..64043d30a86043c4fc850f4a58c53f4a30849b75 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_652.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-652 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-652 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-652-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_652.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_654.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_654.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f179426195453036ee4a7e58700153f6da0b377 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_654.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-654 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-654 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-654-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_654.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_656.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_656.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cae67dffbe2c4ad8e9af62b1d5f07555dd7601ad --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_656.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-656 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-656 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-656-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_656.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_667.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_667.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cc74661a531f6d16027f1d374cd29053522c1ca4 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_667.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-667 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-667 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-667-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_667.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_668.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_668.yaml new file mode 100644 index 0000000000000000000000000000000000000000..67a62ecf68065bbffe03c2672ad8a953c99f9f19 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_668.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-668 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-668 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-668-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_668.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_703.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_703.yaml new file mode 100644 index 0000000000000000000000000000000000000000..87c147e8c383e2c7c06f3523c8eb904b5f923ce5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_703.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-703 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-703 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-703-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_703.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_709.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_709.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bb57ab8577080d35c08bfe020cfac05960ea2693 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_709.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-709 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-709 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-709-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_709.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_714.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_714.yaml new file mode 100644 index 0000000000000000000000000000000000000000..847d37dcbc11ec9e2dc105182acd1cfd9fb2ced5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_714.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-714 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-714 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-714-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_714.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_740.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_740.yaml new file mode 100644 index 0000000000000000000000000000000000000000..504a19c830e2687e7e85d96341ab69ee6d3b80a8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_740.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-740 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-740 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-740-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_740.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_742.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_742.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c6759fecf6a4725f898c323ffca25b234e0e4261 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_742.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-742 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-742 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-742-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_742.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_753.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_753.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5ba15a83d8feb98c952961260a93c06b7ac9506f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_753.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-753 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-753 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-753-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_753.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_779.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_779.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c9d807f2c0e1f3570e6cfb84540a17bdcf9b91f6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_779.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-779 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-779 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-779-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_779.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_782.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_782.yaml new file mode 100644 index 0000000000000000000000000000000000000000..301d43e730151af670de691110c8a36b19bc5188 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_782.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-782 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-782 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-782-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_782.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_817.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_817.yaml new file mode 100644 index 0000000000000000000000000000000000000000..965d3f4c2636faa25f871e0f941d641de45f218d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_817.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-817 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-817 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-817-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_817.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_826.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_826.yaml new file mode 100644 index 0000000000000000000000000000000000000000..77294ef78b30055d11b0e587caae1ea8b39785ac --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_826.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-826 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-826 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-826-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_826.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_90.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_90.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fd866530862792179ab8e25ecc4edc28911bb196 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_90.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-90 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-90 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-90-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_90.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_91.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_91.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a3ee55c9c4b95d1ff1a158992fa512c5143c03a8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_91.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-91 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-91 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-91-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_91.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