diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_107.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_107.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5dc0bc8f2c453895eba9bb739204d4c6b5d431f6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_107.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-107 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-107 +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-107-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="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_107.sh + HOSTSFILE=$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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_113.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_113.yaml new file mode 100644 index 0000000000000000000000000000000000000000..adaa683000959f81e7b86dbd3c104d6beab592aa --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_113.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-113 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-113 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-113-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_113.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_123.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_123.yaml new file mode 100644 index 0000000000000000000000000000000000000000..19474750166fb4cae88c8171dc871f9a94c31e40 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_123.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-123 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-123 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-123-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_123.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_134.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_134.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ddb79847a862f0a428e9c02701ac44d867871fec --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_134.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-134 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-134 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-134-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_134.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_138.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_138.yaml new file mode 100644 index 0000000000000000000000000000000000000000..462a9eea50d664611b6dd14c6063bef9a496d51b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_138.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-138 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-138 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-138-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_138.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_147.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_147.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7f5e80c3cb007e5cf1b7ae0815465ca30c22044a --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_147.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-147 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-147 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-147-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_147.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_16.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_16.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6231b83718c179c1ac3c41301d5d0c0ac4dcb7df --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_16.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-16 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-16 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-16-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_16.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_165.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_165.yaml new file mode 100644 index 0000000000000000000000000000000000000000..301ecbafc13c6063135ed1094f64e01875e337b8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_165.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-165 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-165 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-165-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_165.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_166.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_166.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fea8de88ef2d3b993912ec346c06ea6346a2b7d5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_166.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-166 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-166 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-166-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_166.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_169.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_169.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fc2af2ed17322fa84f6c5a4a106cf24fb4fc9bba --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_169.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-169 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-169 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-169-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_169.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_178.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_178.yaml new file mode 100644 index 0000000000000000000000000000000000000000..77dab8d49fe08ec0ff5cf42d508a5f0113a69837 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_178.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-178 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-178 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-178-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_178.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_195.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_195.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f1ff460404bd8ce8e104402b665c4d4162622907 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_195.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-195 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-195 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-195-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_195.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_206.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_206.yaml new file mode 100644 index 0000000000000000000000000000000000000000..11af3f52f16e6633fd483a3a8400ea69c3b9b2e5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_206.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-206 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-206 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-206-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_206.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_211.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_211.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e8149a35f29e14cdb60c15bde476d40a0ddfb6f7 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_211.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-211 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-211 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-211-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_211.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_224.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_224.yaml new file mode 100644 index 0000000000000000000000000000000000000000..de210d3c7d377eb7d4bf065af6d8686b4c7cc5e2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_224.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-224 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-224 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-224-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_224.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_237.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_237.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f282dbe1808891806cd233e76ab8aecc26e90c67 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_237.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-237 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-237 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-237-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_237.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_241.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_241.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4a29f3e4ace6a29d91bdea206cf3d51e8925ce2c --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_241.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-241 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-241 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-241-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_241.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_246.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_246.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e4a643040fac8516677112a19413033abdadf1bb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_246.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-246 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-246 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-246-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_246.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_249.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_249.yaml new file mode 100644 index 0000000000000000000000000000000000000000..58c1be699cf43876a518c95d24c0e27bbedf3de0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_249.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-249 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-249 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-249-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_249.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_255.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_255.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7342b9b1e6a11c47dc2351455b82250e51e3fe2b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_255.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-255 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-255 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-255-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_255.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_258.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_258.yaml new file mode 100644 index 0000000000000000000000000000000000000000..34df870d662cca208693bdb448c0d1545ba1d34b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_258.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-258 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-258 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-258-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_258.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_26.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_26.yaml new file mode 100644 index 0000000000000000000000000000000000000000..65f0c3e8082049a095e06301ea164973e1fe4adc --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_26.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-26 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-26 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-26-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_26.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_276.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_276.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1cecc67138cf738d4244b32f598c846579fc157f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_276.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-276 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-276 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-276-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_276.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_30.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_30.yaml new file mode 100644 index 0000000000000000000000000000000000000000..06fb1c1211bac0d27f6e4a28b4d5ed4174f2beb6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_30.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-30 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-30 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-30-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_30.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_31.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_31.yaml new file mode 100644 index 0000000000000000000000000000000000000000..842ad5c652880adc774273234bd2618e2a7d6741 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_31.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-31 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-31 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-31-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_31.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_311.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_311.yaml new file mode 100644 index 0000000000000000000000000000000000000000..56d85d8b35b8da4fbd5ca00a8f27a29e155e38d6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_311.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-311 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-311 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-311-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_311.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_314.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_314.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fe2e1320118b52dfa88e8534451e20895985bbe6 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_314.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-314 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-314 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-314-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_314.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_327.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_327.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9af3110d81e9a28cd4b4d58adba7adfe6ea8d4b9 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_327.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-327 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-327 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-327-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_327.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_352.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_352.yaml new file mode 100644 index 0000000000000000000000000000000000000000..835a91f05dd199c4f61215bde6f15515a8ad2164 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_352.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-352 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-352 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-352-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_352.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_357.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_357.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3ea1cca605ec3d5b045cd019cb933dc915d97bc5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_357.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-357 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-357 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-357-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_357.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_358.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_358.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cc868373fbe01231783bf502fa9bf0fffc47ed94 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_358.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-358 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-358 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-358-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_358.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_383.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_383.yaml new file mode 100644 index 0000000000000000000000000000000000000000..db5995fe11d4d5b9b9c29442f2f6c0e8086ec338 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_383.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-383 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-383 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-383-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_383.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_389.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_389.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1ff3c932c4a9cde87f95e5c285f4d84da140bd5c --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_389.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-389 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-389 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-389-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_389.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_391.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_391.yaml new file mode 100644 index 0000000000000000000000000000000000000000..06b8e3a0a749c9a6f5258c713b38546d9dee2bbc --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_391.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-391 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-391 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-391-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_391.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_399.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_399.yaml new file mode 100644 index 0000000000000000000000000000000000000000..35aecc2852f0489e04b780d29aa69f2897d1be49 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_399.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-399 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-399 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-399-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_399.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_4.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_4.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5e2653cfcc513ca49e55c8a80fc0619d47f2fce8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_4.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-4 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-4 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-4-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_4.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_405.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_405.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ca2dd70f32cd38c2e7a8b1d231c0f2c72fc89f5d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_405.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-405 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-405 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-405-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_405.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_407.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_407.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dd16773dd7727890bb36a27c8fb432fa60d2eaa2 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_407.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-407 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-407 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-407-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_407.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_42.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_42.yaml new file mode 100644 index 0000000000000000000000000000000000000000..878cd43fe61108bd2e7d8358379bf65aab7a0c25 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_42.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-42 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-42 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-42-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_42.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_423.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_423.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c29e2beba1db5a75eaa8a225d6a57493aa13ac76 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_423.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-423 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-423 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-423-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_423.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_430.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_430.yaml new file mode 100644 index 0000000000000000000000000000000000000000..6133c6d5f3fbe84fbfd47da4bee68270db78d570 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_430.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-430 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-430 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-430-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_430.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_449.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_449.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2103752190ce79740db76c3bdec8bfa2839ff4b1 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_449.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-449 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-449 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-449-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_449.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_456.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_456.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9b7c76653cb093ba9aec6bab4a4a176c6d416f36 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_456.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-456 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-456 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-456-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_456.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_468.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_468.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ee66b9e19345cbe2b957ddf7e49172b5b522f454 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_468.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-468 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-468 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-468-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_468.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_472.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_472.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8d2bef2bff5a0f22c961f91b23d8c3ee8c65d20b --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_472.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-472 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-472 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-472-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_472.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_506.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_506.yaml new file mode 100644 index 0000000000000000000000000000000000000000..223469b0402a23ade3ac87bfaf5c0607ad4767cb --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_506.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-506 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-506 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-506-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_506.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_512.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_512.yaml new file mode 100644 index 0000000000000000000000000000000000000000..dae8cff3e6aa2cdc6dda1c3f38e362bfb6025e73 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_512.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-512 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-512 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-512-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_512.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_555.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_555.yaml new file mode 100644 index 0000000000000000000000000000000000000000..173b3b43e0adc06f1403107adc9dad9ec3968f72 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_555.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-555 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-555 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-555-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_555.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_578.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_578.yaml new file mode 100644 index 0000000000000000000000000000000000000000..359ccb5326305d915358755ee7ed841b9608492e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_578.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-578 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-578 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-578-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_578.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_583.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_583.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e93aa4f3fa9344e0445d37b47c303e37cc81c401 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_583.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-583 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-583 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-583-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_583.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_600.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_600.yaml new file mode 100644 index 0000000000000000000000000000000000000000..cd2d9c82683ea93e9c9fcd5e0fad27be8957a84e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_600.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-600 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-600 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-600-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_600.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_645.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_645.yaml new file mode 100644 index 0000000000000000000000000000000000000000..66c8a639bc3a88d4782f4b2c165f04eb938843fc --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_645.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-645 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-645 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-645-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_645.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_66.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_66.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2e5fc534a418c98ca452901ace454ca1b128d661 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_66.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-66 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-66 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-66-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_66.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_67.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_67.yaml new file mode 100644 index 0000000000000000000000000000000000000000..49de57ab99b3e5a65c376a0fb15dc63e897f6c99 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_67.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-67 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-67 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-67-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_67.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_687.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_687.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5da6ef6d161f14b5a492427eb148e96338b2725e --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_687.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-687 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-687 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-687-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_687.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_69.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_69.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d2122c320da5d874debc2e03c6615bbd6eee29ee --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_69.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-69 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-69 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-69-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_69.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_699.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_699.yaml new file mode 100644 index 0000000000000000000000000000000000000000..85b05e3bc5b977c7322236468e32a901dbc4524f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_699.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-699 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-699 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-699-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_699.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_701.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_701.yaml new file mode 100644 index 0000000000000000000000000000000000000000..326b4fa82e7199ba99ef2909e940407307bb4b20 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_701.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-701 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-701 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-701-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_701.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_745.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_745.yaml new file mode 100644 index 0000000000000000000000000000000000000000..ab11a9087382cec5c33c543b2974f8b3d0b7b3e0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_745.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-745 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-745 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-745-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_745.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_750.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_750.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5cdab8d51212c69ec028b58c0a61b3f2c7b69a55 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_750.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-750 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-750 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-750-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_750.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_763.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_763.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c27d94b1d61efa22e88afde7e1e1341669045b91 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_763.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-763 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-763 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-763-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_763.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_764.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_764.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b42ab5092f7d3e76410a574bca583a95cbfc7f68 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_764.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-764 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-764 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-764-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_764.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_768.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_768.yaml new file mode 100644 index 0000000000000000000000000000000000000000..66cefe5dd20e6e2e9f906400f077a8fe4e1b003d --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_768.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-768 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-768 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-768-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_768.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_776.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_776.yaml new file mode 100644 index 0000000000000000000000000000000000000000..78440083982a894d7b60e3efd260a837311548d8 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_776.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-776 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-776 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-776-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_776.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_78.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_78.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4ddb1a57c0e3f1d331b6bbf371e9cd149c950ac5 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_78.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-78 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-78 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-78-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_78.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_780.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_780.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9fc5561aa5682230083d8769c7ba03a9e978cd12 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_780.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-780 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-780 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-780-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_780.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_787.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_787.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5f7ae6d96f24c98427d43323f786d1f4d1b2f503 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_787.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-787 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-787 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-787-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_787.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_79.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_79.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d1caddc1d88655a445a0133773d6539845323e82 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_79.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-79 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-79 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-79-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_79.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_794.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_794.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9cde65f7b3e2477a60bef3779b38d589dd588e29 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_794.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-794 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-794 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-794-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_794.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_80.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_80.yaml new file mode 100644 index 0000000000000000000000000000000000000000..0014b74ec2c49cb65b6016e519ee3e142b466c3c --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_80.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-80 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-80 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-80-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_80.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_800.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_800.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3f0bd1df3e59e64b3d0bc01040da5ba0ac95f63f --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_800.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-800 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-800 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-800-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_800.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_810.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_810.yaml new file mode 100644 index 0000000000000000000000000000000000000000..428d20ec0a66fb5282c1344685c90865e2363e41 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_810.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-810 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-810 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-810-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_810.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_815.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_815.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8a227aafaee3369936444923b6a890a59b2c8ac0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_815.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-815 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-815 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-815-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_815.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_822.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_822.yaml new file mode 100644 index 0000000000000000000000000000000000000000..839101fb025a4021252434f0815a71daf3b506df --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_822.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-822 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-822 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-822-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_822.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 $TOKENIZER_CMD; + + + Worker: + replicas: 1 + template: + spec: + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + tolerations: + - key: "habana.ai/gaudi" + operator: "Exists" + effect: "NoSchedule" + - key: "k8s/namespace" + operator: "Equal" + value: "peacock" + effect: "NoSchedule" + hostIPC: true + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-llama-worker + imagePullPolicy: IfNotPresent + securityContext: + capabilities: + add: + - SYSLOG + resources: + limits: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + requests: + habana.ai/gaudi: 8 + hugepages-2Mi: 300Gi + memory: 700Gi + cpu: 150 + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock diff --git a/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_89.yaml b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_89.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c33f78a648655b7aa4deccdd1d181e5cf0defbe0 --- /dev/null +++ b/hn_eng_bn/tok_en-bn-hn/tok_jobs/tokenizer_89.yaml @@ -0,0 +1,104 @@ +apiVersion: kubeflow.org/v2beta1 +kind: MPIJob +metadata: + name: peacock-tokenizer-en-bn-hn-89 + namespace: peacock + labels: + app: peacock-tokenizer-en-bn-hn-89 +spec: + slotsPerWorker: 8 + runPolicy: + backoffLimit: 1 + cleanPodPolicy: Running + mpiReplicaSpecs: + Launcher: + replicas: 1 + template: + spec: + hostIPC: true + volumes: + - name: work-dir + persistentVolumeClaim: + claimName: peacock-fs-pvc + containers: + - image: vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest + name: peacock-tokenizer-en-bn-hn-89-launcher + imagePullPolicy: IfNotPresent + volumeMounts: + - name: work-dir + mountPath: /mnt/weka/peacock + command: ["/bin/bash", "-c"] + args: + - >- + + export SYNAPSE_VERSION="1.15.1"; + export WORKER_DIR="/mnt/weka/peacock/experiments/llama"; + + export MEGATRON_SETUP_CMD="$WORKER_DIR/launch/setup.sh" + export TOKENIZER_CMD=/mnt/weka/peacock/idc/datasets/hn_eng_bn/tok_en-bn-hn/tok_files/tokenizer_89.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