Add A4x GCS checkpointing recipes

Author: akiki-liang0

training/a4x/llama3_70b/nemo-gke/nemo2602/checkpoint/128gpu-fp8mx-gcs/recipe/Chart.yaml

@@ -0,0 +1,21 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+apiVersion: v2
+name: a4x_jobset_workload
+description: a4x_jobset_workload
+type: application
+version: 0.1.0
+appVersion: "1.16.0"
+

training/a4x/llama3_70b/nemo-gke/nemo2602/checkpoint/128gpu-fp8mx-gcs/recipe/README.md

@@ -0,0 +1,264 @@
+<!-- mdformat global-off -->
+# Pretrain llama3-1-70b workloads on a4x GKE Node pools with Nvidia NeMo Framework
+
+This recipe outlines the steps for running a llama3-1-70b pretraining
+workload on [a4x GKE Node pools](https://cloud.google.com/kubernetes-engine) by using the
+[NVIDIA NeMo framework](https://github.com/NVIDIA/nemo).
+
+## Orchestration and deployment tools
+
+For this recipe, the following setup is used:
+
+- Orchestration - [Google Kubernetes Engine (GKE)](https://cloud.google.com/kubernetes-engine)
+- Pretraining job configuration and deployment - A Helm chart is used to
+  configure and deploy the [Kubernetes Jobset](https://kubernetes.io/blog/2025/03/23/introducing-jobset) resource which manages the execution of the
+  [NeMo pretraining workload](https://github.com/NVIDIA/nemo).
+
+## Test environment
+
+This recipe has been optimized for and tested with the following configuration:
+
+- GKE cluster
+Please follow Cluster Toolkit [instructions](https://github.com/GoogleCloudPlatform/cluster-toolkit/tree/main/examples/gke-a4x)
+to create your a4x GKE cluster.
+
+## Training dataset
+
+This recipe uses a mock pretraining dataset provided by the NeMo framework.
+
+## Docker container image
+
+This recipe uses the following docker images:
+
+- `nvcr.io/nvidia/nemo:26.02.01`
+- `us-docker.pkg.dev/gce-ai-infra/gpudirect-gib/nccl-plugin-gib-diagnostic-arm64:v1.1.0`
+
+## Run the recipe
+
+From your client workstation, complete the following steps:
+
+### Configure environment settings
+
+Set the environment variables to match your environment:
+
+ ```bash
+ export PROJECT_ID=<PROJECT_ID>
+ export CLUSTER_REGION=<CLUSTER_REGION>
+ export CLUSTER_NAME=<CLUSTER_NAME>
+ export GCS_BUCKET=<GCS_BUCKET> # Note: path should not be prefixed with gs://
+ export KUEUE_NAME=<KUEUE_NAME>
+ ```
+
+Replace the following values:
+
+ - `<PROJECT_ID>`: your Google Cloud project ID.
+ - `<CLUSTER_REGION>`: the region where your cluster is located.
+ - `<CLUSTER_NAME>`: the name of your GKE cluster.
+ - `<GCS_BUCKET>`: the name of your Cloud Storage bucket. Don't include the `gs://` prefix.
+ - `<KUEUE_NAME>`: the name of the Kueue local queue. The default queue created by the cluster toolkit is `a4x`. Make sure to verify the name of the local queue in your cluster.
+
+Set the default project:
+
+ ```bash
+ gcloud config set project $PROJECT_ID
+ ```
+
+### Get the recipe
+
+Clone the `gpu-recipes` repository and set a reference to the recipe folder.
+
+```
+git clone https://github.com/ai-hypercomputer/gpu-recipes.git
+cd gpu-recipes
+export REPO_ROOT=`git rev-parse --show-toplevel`
+export RECIPE_ROOT=$REPO_ROOT/training/a4x/llama3_70b/nemo-gke/nemo2602/128gpus-fp8mx-gcs/recipe
+cd $RECIPE_ROOT
+```
+
+### Get cluster credentials
+
+```
+gcloud container clusters get-credentials $CLUSTER_NAME --region $CLUSTER_REGION
+```
+### Setup GCS bucket PVCs
+1) Create GCS buckets for dataload and checkpointing.
+
+Ensure the buckets have `--uniform-bucket-level-access` and `--enable-hirearchical-namespace`
+
+```
+export CKPT_BUCKET_NAME=a4x-storage-ckpt
+export DL_BUCKET_NAME=a4x-storage-data
+
+gcloud storage buckets create "gs://${CKPT_BUCKET_NAME}" --location=$CLUSTER_REGION --uniform-bucket-level-access --enable-hierarchical-namespace
+
+gcloud storage buckets create "gs://${DL_BUCKET_NAME}" --location=$CLUSTER_REGION --uniform-bucket-level-access --enable-hierarchical-namespace
+```
+
+2) Upload training dataset to dataload bucket
+For commands, see: [Upload opjects from a file system](https://docs.cloud.google.com/storage/docs/uploading-objects)
+
+3) Create PersistentVolumes for the buckets and claim these volumes
+
+Replace `<DL_BUCKET_NAME>` and `<CKPT_BUCKET_NAME>` in gcs_pv.yamk and run the following command. 
+Ensure file cache is enabled for optimal performance.
+
+```
+kubectl apply -f ./gcs_pv.yaml
+```
+
+4) Add volume claims to values.yaml
+
+```
+...
+gcsVolumes: true
+psVolumes: false
+pvcMounts:
+- claimName: "a4x-storage-ckpt-pvc"
+  mountPath: "/gcsckpt"
+- claimName: "a4x-storage-data-pvc"
+  mountPath: "/gcsdata"
+...
+```
+
+Using PVC optimizes performance, but you can also mount the GCS bucket using gcsMounts option.
+
+### Configure and submit a pretraining job
+
+#### Using 32 node (128 gpus) fp8 precision with GCS dataload and checkpointing
+To execute the job with dataloading and checkpoint saving, run the following:
+
+```bash
+cd $RECIPE_ROOT
+export WORKLOAD_NAME=$USER-a4x-llama3-1-70b-32node
+export DATASET_TYPE=<DATASET_TYPE>
+export DATASET_PATHS=<DATASET_PATHS>
+export INDEX_MAPPING_DIR=<INDEX_MAPPING_DIR>
+export CKPT_SAVE_DIR=<CKPT_SAVE_DIR>
+export CKPT_SAVE_INTERVAL=<CKPT_SAVE_INTERVAL>
+helm install $WORKLOAD_NAME . -f values.yaml \
+--set-file workload_launcher=launcher.sh \
+--set-file workload_config=llama3-1-70b-fp8cs-gbs2048-gpus128.py \
+--set workload.image=nvcr.io/nvidia/nemo:26.02.01 \
+--set volumes.gcsMounts[0].bucketName=${GCS_BUCKET} \
+--set volumes.gcsMounts[0].mountPath=/job-logs \
+--set workload.envs[0].value=/job-logs/$WORKLOAD_NAME \
+--set workload.envs[5].value=${DATASET_TYPE} \
+--set workload.envs[6].value="/gcsdata/${DATASET_PATHS}" \
+--set workload.envs[7].value=/gcsdata/${INDEX_MAPPING_DIR} \
+--set workload.envs[8].value=/gcsckpt/${CKPT_SAVE_DIR} \
+--set workload.envs[9].value=/gcsckpt/${CKPT_SAVE_INTERVAL} \
+--set queue=${KUEUE_NAME}
+```
+
+Replace the following values:
+ - `<DATASET_TYPE>`: The type of dataset used (see [Megatron-Bridge data arguments](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/r0.3.0/scripts/performance#data-arguments))
+ - `<DATASET_PATHS>`: Paths to your dataset (for rp2 dataset)
+ - `<INDEX_MAPPING_DIR>`: Index mapping dir (for rp2 dataset)
+ - `<CKPT_SAVE_DIR>`: The directory where you wish you save checkpoints
+ - `<CKPT_SAVE_INTERVAL>`: Save checkpoint every CKPT_SAVE_INTERVAL train steps
+
+Note: Edit `recipe.dataset.num_workers = 8` in `run_script.py` to change the number of dataloading workers.
+
+This recipe uses Nvidia NeMo 26.02.01 container, which uses Megatron-Bridge for checkpointing and dataloading. See [checkpointing arguments](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/r0.3.0/scripts/performance#checkpointing-arguments) and [data arguments](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/r0.3.0/scripts/performance#data-arguments) for explanations on dataset types.
+
+You can also edit values.yaml directly.
+```
+- name: DATASET_TYPE
+  value: "rp2"
+- name: DATASET_PATHS
+  value: "/gcsdata/${DATASET_PATHS}"
+- name: INDEX_MAPPING_DIR
+  value: "/gcsdata/${INDEX_MAPPING_DIR}"
+- name: CKPT_SAVE_DIR
+  value: "/gcsckpt/${CKPT_SAVE_DIR}"
+- name: CKPT_SAVE_INTERVAL
+  value: "10"
+```
+
+To load a checkpoint, add the following:
+```bash
+cd $RECIPE_ROOT
+export WORKLOAD_NAME=$USER-a4x-llama3-1-70b-32node
+export CKPT_LOAD_DIR=<CKPT_LOAD_DIR>
+export CKPT_LOAD_STEP=<CKPT_LOAD_STEP>
+helm install $WORKLOAD_NAME . -f values.yaml \
+--set-file workload_launcher=launcher.sh \
+--set-file workload_config=llama3-1-70b-fp8cs-gbs2048-gpus128.py \
+--set workload.image=nvcr.io/nvidia/nemo:26.02.01 \
+--set volumes.gcsMounts[0].bucketName=${GCS_BUCKET} \
+--set volumes.gcsMounts[0].mountPath=/job-logs \
+--set workload.envs[0].value=/job-logs/$WORKLOAD_NAME \
+--set workload.envs[10].value=/gcsckpt/${CKPT_LOAD_DIR} \
+--set workload.envs[11].value="/gcsckpt/${CKPT_LOAD_STEP}" \
+--set queue=${KUEUE_NAME}
+```
+
+Or edit values.yaml directly
+```
+- name: CKPT_LOAD_DIR
+  value: "/gcsckpt/${CKPT_LOAD_DIR}"
+- name: CKPT_LOAD_STEP
+  value: "10"
+```
+
+**Examples**
+
+-   To set the number of training steps to 100, run the following command from your client:
+
+```bash
+cd $RECIPE_ROOT
+export WORKLOAD_NAME=$USER-a4x-llama3-1-70b-32node
+helm install $WORKLOAD_NAME . -f values.yaml \
+--set-file workload_launcher=launcher.sh \
+--set-file workload_config=llama3-1-70b-fp8cs-gbs2048-gpus128.py \
+--set workload.image=nvcr.io/nvidia/nemo:26.02.01 \
+--set volumes.gcsMounts[0].bucketName=${GCS_BUCKET} \
+--set volumes.gcsMounts[0].mountPath=/job-logs \
+--set workload.envs[0].value=/job-logs/$WORKLOAD_NAME \
+--set workload.envs[5].value=${DATASET_TYPE} \
+--set workload.envs[6].value="/gcsdata/${DATASET_PATHS}" \
+--set workload.envs[7].value=/gcsdata/${INDEX_MAPPING_DIR} \
+--set workload.envs[8].value=/gcsckpt/${CKPT_SAVE_DIR} \
+--set workload.envs[9].value=/gcsckpt/${CKPT_SAVE_INTERVAL} \
+--set queue=${KUEUE_NAME} \
+--set workload.arguments[0]="trainer.max_steps=100"
+```
+
+
+### Monitor the job
+
+To check the status of pods in your job, run the following command:
+
+```
+kubectl get pods | grep $USER-a4x-llama3-1-70b-32node
+```
+
+Replace the following:
+
+- JOB_NAME_PREFIX - your job name prefix. For example $USER-a4x-llama3-1-70b-32node.
+
+To get the logs for one of the pods, run the following command:
+
+```
+kubectl logs POD_NAME
+```
+
+Information about the training job's progress, including crucial details such as
+loss, step count, and step time, is generated by the rank 0 process.
+This process runs on the pod whose name begins with
+`JOB_NAME_PREFIX-workload-0-0`.
+For example: `$USER-a4x-llama3-1-70b-32node-workload-0-0-s9zrv`.
+
+Logs will display checkpoint save directory:
+```
+28_worker0/0   successfully saved checkpoint from iteration      40 to /gcsckpt/ckpt/asq-llama70b-ckpt-128gpu-gcs-2026-04-03-03-51-29 [ t 1/2, p 1/4 ] 
+```
+
+### Uninstall the Helm release
+
+You can delete the job and other resources created by the Helm chart. To
+uninstall Helm, run the following command from your client:
+
+```bash
+helm uninstall $USER-a4x-llama3-1-70b-32node
+```