Add files using upload-large-folder tool

.gitattributes

@@ -57,3 +57,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+en_hn_south/final.json filter=lfs diff=lfs merge=lfs -text

en_hn_south/final.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf8f3c13bf07efae57befd0f1fc2ff78455dace4759b764c3312f5aeda806bd1
+size 67119608

sangraha/tok_sangraha/tok_jobs/tokenizer_0.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-0
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_0.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_1.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-1
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_1.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_10.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-10
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_10.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_11.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-11
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_11.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_12.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-12
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_12.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_13.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-13
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_13.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_14.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-14
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_14.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_15.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-15
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_15.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_16.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-16
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_16.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_17.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-17
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_17.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_18.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-18
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_18.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_19.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-19
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_19.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_2.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-2
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_2.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_20.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-20
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_20.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_21.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-21
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_21.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_22.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-22
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_22.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_23.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-23
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_23.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_24.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-24
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_24.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_25.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-25
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_25.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_26.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-26
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_26.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_27.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-27
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_27.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_28.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-28
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_28.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_29.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-29
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_29.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_3.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-3
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_3.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_30.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-30
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_30.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_31.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-31
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_31.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_32.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-32
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_32.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_33.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-33
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_33.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_34.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-34
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_34.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_35.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-35
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_35.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_36.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-36
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_36.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_37.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-37
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_37.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_38.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-38
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_38.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_39.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-39
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_39.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_4.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-4
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_4.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_40.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-40
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_40.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_41.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-41
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_41.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_42.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-42
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_42.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_43.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-43
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_43.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_44.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-44
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_44.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_45.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-45
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_45.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_5.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-5
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_5.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_6.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-6
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_6.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_7.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-7
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_7.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_8.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-8
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_8.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs/tokenizer_9.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-9
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_9.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs1/tokenizer_10.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-10
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_10.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+                    

sangraha/tok_sangraha/tok_jobs1/tokenizer_11.yaml

@@ -0,0 +1,122 @@
+apiVersion: kubeflow.org/v2beta1
+kind: MPIJob
+metadata:
+  name: peacock-tokenizer-sangraha-11
+spec:
+  slotsPerWorker: 8
+  runPolicy:
+    backoffLimit: 0
+    cleanPodPolicy: Running
+  mpiReplicaSpecs:
+    Launcher:
+      replicas: 1
+      template:
+        spec:
+          hostIPC: true
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-launcher
+              imagePullPolicy: IfNotPresent
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  sleep 10;
+                  export SYNAPSE_VERSION="1.17.0";
+                  export WORKER_DIR="/mnt/weka/peacock/training/llama3b";
+                  export MODEL_PATH="${WORKER_DIR}/Megatron-DeepSpeed";
+                  export MEGATRON_DEEPSPEED_ROOT=$MODEL_PATH;
+                  export MODEL_REQ="${MODEL_PATH}/megatron/core/requirements.txt";
+
+                  export TOKENIZER_CMD=/mnt/weka/peacock/peacock-data/idc/datasets/sangraha//tok_sangraha/tok_files/tokenizer_11.sh
+
+                  export HL_LLAMA_VER=2;
+                  export HL_LLAMA_MODEL_SIZE=3;
+
+                  export HL_DATA_DIR_ROOT="/mnt/weka/peacock/peacock-data/idc/hineng/hn_eng_bn/merged_all";
+                  export HL_DATA_FILE_PREFIX="all";
+                  export HL_DATASET_NAME=slimpajama-subset;
+                  export HL_DP=16;
+                  export HL_PP=2;
+                  export HL_TP=4;
+                  export HL_TRAIN_ITERS=240000;
+                  export HL_SAVE_INTERVAL=1000;
+
+                  #export HL_EXIT_INTERVAL=100;
+                  #export HL_SAVE=0;
+
+                  export HL_MICRO_BATCH=8;
+
+                  export HL_TOKENIZER_TYPE=SentencePieceTokenizer;
+                  export HL_TOKENIZER_DIR="/mnt/weka/peacock/peacock-data/tokenization/trained-tokenizer/enhiben_50k";
+
+                  export RESULT_DESCRIPTION="${HL_DATASET_NAME}-TI${HL_TRAIN_ITERS}-TP${HL_TP}PP${HL_PP}DP${HL_DP}-MBS${HL_MICRO_BATCH}";
+                  export DEBUG=True;
+
+                  set -exo pipefail; 
+
+
+                  $WORKER_DIR/launch/setup.sh;
+
+                  mpirun  --npernode 1 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x MODEL_REQ=$MODEL_REQ \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x MODEL_PATH=$MODEL_PATH \
+                    -x MODEL_REFERENCES_ROOT=$MODEL_REFERENCES_ROOT \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION \
+                    $WORKER_DIR/launch/setup.sh;
+
+                  cd $MODEL_PATH;
+
+                  mpirun  -npernode 8 \
+                    --tag-output \
+                    --allow-run-as-root \
+                    --prefix $MPI_ROOT \
+                    -x WORKER_DIR=$WORKER_DIR \
+                    -x SYNAPSE_VERSION=$SYNAPSE_VERSION $TOKENIZER_CMD;
+    Worker:
+      replicas: 1
+      template:
+        spec:
+          runtimeClassName: habana
+          nodeSelector:
+            role: worker
+          volumes:
+            - name: pvc-wekafs-peacock
+              persistentVolumeClaim:
+                claimName: pvc-wekafs-peacock
+          hostIPC: true
+          containers:
+            - image: vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
+              name: llama2-3b-pretrain-worker
+              imagePullPolicy: IfNotPresent
+              securityContext:
+                capabilities:
+                  add:
+                    - SYSLOG
+              resources:
+                limits:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+                requests:
+                  habana.ai/gaudi: 8
+                  hugepages-2Mi: 300Gi
+                  memory: 600Gi
+                  cpu: 120
+              volumeMounts:
+                - name: pvc-wekafs-peacock
+                  mountPath: /mnt/weka/peacock
+