Merge branch 'main' into ocr

small fixed
adding mistral ocr (#4069 )
2026-09-27 07:31:58 +00:00 · 2025-05-23 13:31:37 -04:00 · 2025-05-23 17:27:59 +00:00 · 2025-05-23 16:03:27 +00:00 · 2025-05-23 15:42:46 +00:00 · 2025-05-23 15:24:06 +00:00
504 changed files with 124073 additions and 58645 deletions
@@ -238,7 +238,7 @@ def _get_notebook_python_version(notebook_path: str) -> str:

            # Look for the python version specification pattern
            re_match = re.search(
-                "python version = (\d+\.\d+)", markdown, flags=re.IGNORECASE
+                r"python version = (\d+\.\d+)", markdown, flags=re.IGNORECASE
            )
            if re_match:
                # get the version number
@@ -1,5 +1,3 @@
-notebooks/official/vizier/gapic-vizier-multi-objective-optimization.ipynb
-notebooks/official/pipelines/lightweight_functions_component_io_kfp.ipynb
 notebooks/official/ml_metadata/sdk-metric-parameter-tracking-for-locally-trained-models.ipynb
 notebooks/official/custom/custom-tabular-bq-managed-dataset.ipynb
 .cloud-build/tests/python_version_test.ipynb
@@ -0,0 +1,10 @@
+version: 2
+updates:
+
+  # Ignore model garden dockerfiles:
+  - package-ecosystem: "npm"
+    directory: "/community-content/vertex_model_garden"
+    schedule:
+      interval: "monthly"
+    ignore:
+      - dependency-name: "*"
@@ -4,7 +4,7 @@
 # 2. To lint specific notebooks:
 # docker run -v ${PWD}:/setup/app gcr.io/python-docs-samples-tests/notebook_linter:latest notebooks/1.ipynb notebooks/2.ipynb

-FROM python:3.12
+FROM python:3.13

 WORKDIR setup

@@ -2,9 +2,9 @@ git+https://github.com/tensorflow/docs
 ipython
 jupyter
 nbconvert
-black==24.4.2
-pyupgrade==3.15.2
-isort==5.13.2
-flake8==7.0.0
-nbqa==1.8.5
+black==25.1.0
+pyupgrade==3.19.1
+isort==6.0.1
+flake8==7.2.0
+nbqa==1.9.1

@@ -58,7 +58,7 @@ done
 # Only check notebooks in test folders modified in this pull request.
 # Note: Use process substitution to persist the data in the array
 if [ ${#notebooks[@]} -eq 0 ]; then
-    echo "Checking for changed notebooked using git"
+    echo "Checking for changed notebooks using git"
    while read -r file || [ -n "$line" ]; do
        notebooks+=("$file")
    done < <(git diff --name-only main... | grep '\.ipynb$')
@@ -1,37 +1,176 @@
-# Google Cloud Vertex AI Samples
+# ![Google Cloud](https://avatars.githubusercontent.com/u/2810941?s=60&v=4) Google Cloud Vertex AI Samples

-[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)
-
-Welcome to the Google Cloud [Vertex AI](https://cloud.google.com/vertex-ai/docs/) sample repository.
+This repository contains notebooks, code samples, sample apps, and other resources that demonstrate how to use, develop and manage machine learning and generative AI workflows using Google Cloud Vertex AI.

 ## Overview

-The repository contains [notebooks](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/master/notebooks) and [community content](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/master/community-content) that demonstrate how to develop and manage ML workflows using Google Cloud Vertex AI.
+[Vertex AI](https://cloud.google.com/vertex-ai) is a fully-managed, unified AI development platform for building and using generative AI. This repository is designed to help you get started with Vertex AI. Whether you're new to Vertex AI or an experienced ML practitioner, you'll find valuable resources here.
+
+For more Vertex AI Generative AI notebook samples, please visit the Vertex AI [Generative AI](https://github.com/GoogleCloudPlatform/generative-ai) GitHub repository.
+
+## Explore, learn and contribute
+
+You can explore, learn, and contribute to this repository to unleash the full potential of machine learning on Vertex AI! 
+
+### Explore and learn
+
+Explore this repository, follow the links in the header section of each of the notebooks to -
+
+![Colab](https://cloud.google.com/ml-engine/images/colab-logo-32px.png)  Open and run the notebook in [Colab](https://colab.google/)\
+![Colab Enterprise](https://cloud.google.com/ml-engine/images/colab-enterprise-logo-32px.png)  Open and run the notebook in [Colab Enterprise](https://cloud.google.com/colab/docs/introduction)\
+![Workbench](https://lh3.googleusercontent.com/UiNooY4LUgW_oTvpsNhPpQzsstV5W8F7rYgxgGBD85cWJoLmrOzhVs_ksK_vgx40SHs7jCqkTkCk=e14-rj-sc0xffffff-h130-w32)  Open and run the notebook in [Vertex AI Workbench](https://cloud.google.com/vertex-ai/docs/workbench/introduction)\
+![Github](https://cloud.google.com/ml-engine/images/github-logo-32px.png)  View the notebook on Github
+
+### Contribute
+
+See the [Contributing Guide](https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/master/CONTRIBUTING.md).
+
+## Get started
+
+To get started using Vertex AI, you must have a Google Cloud project. 
+
+- If you don't have a Google Cloud project, you can learn and build on GCP for free using [Free Trail](https://cloud.google.com/free).
+- Once you have a Google Cloud project, you can learn more about [setting up a project and a development environment](https://cloud.google.com/vertex-ai/docs/start/cloud-environment).
+

 ## Repository structure

 ```bash
-├── community-content - Sample code and tutorials contributed by the community
 ├── notebooks
-│   ├── community - Notebooks contributed by the community
 │   ├── official - Notebooks demonstrating use of each Vertex AI service
 │   │   ├── automl
 │   │   ├── custom
 │   │   ├── ...
+│   ├── community - Notebooks contributed by the community
+│   │   ├── model_garden
+│   │   ├── ...
+├── community-content - Sample code and tutorials contributed by the community
+
 ```
+## Examples

-## Contributing
+<!-- markdownlint-disable MD033 -->
+<table>

-Contributions welcome! See the [Contributing Guide](https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/master/CONTRIBUTING.md).
+  <tr>
+    <th style="text-align: center;">Category</th>
+    <th style="text-align: center;">Product</th>
+    <th style="text-align: center;">Description</th>
+  </tr>
+  <tr>
+    <td>Model</td>
+    <td>
+      <a href="notebooks/community/model_garden"><code>Model Garden/</code></a>
+    </td>
+    <td>
+      Curated collection of first-party, open-source, and third-party models available on Vertex AI including Gemini, Gemma, Llama 3, Claude 3 and many more.
+    </td>
+  </tr>
+  <tr>
+    <td>Data</td>
+    <td>
+      <a href="notebooks/official/feature_store"><code>Feature Store/</code></a>
+    </td>
+    <td>
+      Set up and manage online serving using Vertex AI Feature Store.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/datasets"><code>datasets/</code></a>
+    </td>
+    <td>
+      Use BigQuery and Data Labeling service with Vertex AI.
+    </td>
+  </tr>
+  <tr>
+    <td>Model development</td>
+    <td>
+      <a href="notebooks/official/automl"><code>automl/</code></a>
+    </td>
+    <td>
+      Train and make predictions on AutoML models
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/custom"><code>custom/</code></a>
+    </td>
+    <td>
+      Create, deploy and serve custom  models on Vertex AI
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/ray_on_vertex_ai"><code>ray_on_vertex_ai/</code></a>
+    </td>
+    <td>
+      Use Colab Enterprise and Vertex AI SDK for Python to connect to the Ray Cluster.
+    </td>
+  </tr>
+  <tr>
+    <td>Deploy and use</td>
+    <td>
+      <a href="notebooks/official/prediction"><code>prediction/</code></a>
+    </td>
+    <td>
+      Build, train and deploy models using prebuilt containers for custom training and prediction.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/model_registry"><code>model_registry/</code></a>
+    </td>
+    <td>
+      Use Model Registry to create and register a model.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/explainable_ai"><code>Explainable AI/</code></a>
+    </td>
+    <td>
+      Use Vertex Explainable AI's feature-based and example-based explanations to explain how or why a model produced a specific prediction.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/ml_metadata"><code>ml_metadata/</code></a>
+    </td>
+    <td>
+      Record the metadata and artifacts and query that metadata to help analyze, debug, and audit the performance of your ML system.
+    </td>
+  </tr>
+  <tr>
+    <td>Tools</td>
+    <td>
+      <a href="notebooks/official/pipelines"><code>Pipelines/</code></a>
+    </td>
+    <td>
+      Use `Vertex AI Pipelines` and `Google Cloud Pipeline Components` to build, tune, or deploy a custom model.
+    </td>
+  </tr>
+</table>
+<!-- markdownlint-enable MD033 -->

-## Getting help

-Please use the [issues page](https://github.com/GoogleCloudPlatform/vertex-ai-samples/issues) to provide feedback or submit a bug report.
+## Get help
+
+Please use the [Issues page](https://github.com/GoogleCloudPlatform/vertex-ai-samples/issues) to provide feedback or submit a bug report.

 ## Disclaimer

 This is not an officially supported Google product. The code in this repository is for demonstrative purposes only.

-## Feedback

-Please feel free to fill out our [survey](https://bit.ly/vertex-ai-samples-survey) to give us feedback on the repo and its content.
+## References
+- [Vertex AI Jupyter Notebook tutorials](https://cloud.google.com/vertex-ai/docs/tutorials/jupyter-notebooks)
+- Vertex AI [Generative AI](https://github.com/GoogleCloudPlatform/generative-ai) GitHub repository
+- [Vertex AI documentaton](https://cloud.google.com/vertex-ai/docs)
+  
@@ -20,6 +20,7 @@
 /vertex_model_garden/model_oss/movinet @KCFindstr
 /vertex_model_garden/model_oss/data_converter @KCFindstr
 /vertex_model_garden/model_oss/peft @weigary
+/vertex_model_garden/model_oss/peft/templates @rayandasoriya
 /vertex_model_garden/model_oss/lm-evaluation-harness @kathyyu-google
 /vertex_model_garden/model_oss/tfvision @dstnluong-google
 /vertex_model_garden/model_oss/fvlm @minwoo33park
@@ -1,16 +1,40 @@
-FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime
+# Stage 1: Build Environment
+FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime AS builder

+# Install necessary tools and dependencies
 RUN apt-get update && \
    apt-get install -y curl gnupg && \
    echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] http://packages.cloud.google.com/apt cloud-sdk main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
-    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg  add - && \
+    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \
    apt-get update -y && \
-    apt-get install google-cloud-sdk -y
+    apt-get install -y google-cloud-sdk

+# Copy application code
 COPY . /trainer

+# Set working directory
 WORKDIR /trainer

-RUN pip install -r requirements.txt
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt

-ENTRYPOINT ["python", "-m", "task"]
+# Stage 2: Runtime Environment
+FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime
+
+# Install Google Cloud SDK
+RUN apt-get update && \
+    apt-get install -y curl gnupg && \
+    echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] http://packages.cloud.google.com/apt cloud-sdk main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
+    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \
+    apt-get update -y && \
+    apt-get install -y google-cloud-sdk && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Copy from the builder stage
+COPY --from=builder /trainer /trainer
+
+# Set working directory
+WORKDIR /trainer
+
+# Set the entry point
+ENTRYPOINT ["python", "-m", "task"]
@@ -1,3 +1,3 @@
-torch==1.13.1
+torch==2.2.0
 torchvision==0.9.1
 tensorboard==2.5.0
@@ -1,3 +1,3 @@
-torch==1.13.1
+torch==2.2.0
 torchvision==0.9.1
 tensorboard==2.5.0
@@ -1,4 +1,4 @@
 google-cloud-bigquery==2.20.0
-tensorflow==2.7.2
+tensorflow==2.12.1
 pillow==10.3.0
 tf-agents==0.8.0
@@ -1,4 +1,4 @@
 google-cloud-pubsub==2.5.0
 pillow==10.3.0
 tf-agents==0.8.0
-tensorflow==2.7.2
+tensorflow==2.12.1
@@ -1,5 +1,5 @@
 dataclasses==0.6
 google-cloud-aiplatform==1.8.1
-tensorflow==2.7.2
+tensorflow==2.12.1
 pillow==10.3.0
 tf-agents==0.8.0
@@ -1 +1 @@
-tensorflow==2.7.2
+tensorflow==2.12.1
@@ -1 +1 @@
-tensorflow==2.7.2
+tensorflow==2.12.1
@@ -0,0 +1,15 @@
+# Vertex AI custom prediction routines samples
+
+## Overview
+Vertex Custom Prediction Routines(CPR) simplify the process of building custom containers 
+and make local model testing easy. Here are the sameple codes for different libraries.
+
+
+### Objectives
+The objective is to provide various samples for Vertex Custom Prediction Routine(CPR).
+
+
+### Supporting libraries
+* torch
+* sklearn
+* xgboost
@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import load_breast_cancer
+from sklearn.linear_model import RidgeClassifier
+
+class LinearRegressionPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = RidgeClassifier()
+        X, y = load_breast_cancer(return_X_y=True)
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}
@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import make_blobs
+from sklearn.linear_model import LinearRegression 
+
+class LinearRegressionPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = LogisticRegression()
+        X, y = make_blobs(n_samples=100, centers=2, n_features=2, random_state=1)
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict_proba(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}
@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.linear_model import SGDClassifier
+
+class SGDClassifierPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = SGDClassifier(max_iter=5)
+        X = [[0., 0.], [1., 1.]]
+        y = [0, 1]
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}
@@ -0,0 +1,34 @@
+import os
+import torch
+
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from torchvision.models import detection, resnet50, ResNet50_Weights
+from typing import Dict, List
+
+class ResNetPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists("model.pth.tar"):
+            self.model = detection.fasterrcnn_resnet50_fpn(pretrained=True)
+            stat_dic = torch.load("model.pth.tar")
+            self.model.load_state_dict(stat_dic['state_dict'])
+        else:
+            weights = ResNet50_Weights.DEFAULT
+            self.model = resnet50(weights=weights)
+        self.model.eval()
+
+    def preprocess(self, prediction_input: dict) -> torch.Tensor:
+        instances = prediction_input["instances"]
+        return torch.Tensor(instances)
+
+    @torch.inference_mode()
+    def predict(self, instances: torch.Tensor) -> List[str]:
+        return self._model(instances)
+
+    def postprocess(self, prediction_results: List[str]) -> Dict:
+        return {"predictions": prediction_results}
@@ -0,0 +1,73 @@
+import ast
+import json
+import os
+import pickle
+import torch
+
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from transformers import AutoModelForQuestionAnswering
+from typing import Dict, List
+
+class TorchTransformersPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        
+        if os.path.isfile("setup_config.json"):
+            with open("setup_config.json") as setup_config_file:
+                self.setup_config = json.load(setup_config_file)
+
+        if os.path.exists("model.pt"):
+            self.model = AutoModelForQuestionAnswering.from_pretrained("model.pt")
+            self.model.eval()
+        else:
+            raise ValueError("One of the following model files must be provided: model.pt.")
+
+    def preprocess(self, prediction_input: dict) -> torch.Tensor:
+        max_length = self.setup_config["max_length"]
+        instances = prediction_input["instances"]
+        question_context = ast.literal_eval(instances)
+        question = question_context["question"]
+        context = question_context["context"]
+        inputs = self.tokenizer.encode_plus(
+            question,
+            context,
+            max_length=int(max_length),
+            pad_to_max_length=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        input_ids = inputs["input_ids"]
+        attention_mask = inputs["attention_mask"]
+        return torch.Tensor(input_ids, attention_mask)
+    
+    @torch.inference_mode()
+    def predict(self, instances: torch.Tensor) -> List[str]:
+        input_ids, attention_mask = instances
+        outputs = self._model(input_ids, attention_mask)
+        answer_start_scores = outputs.start_logits
+        answer_end_scores = outputs.end_logits
+        
+        num_rows, num_cols = answer_start_scores.shape
+        inferences = []
+        for i in range(num_rows):
+            answer_start_scores_one_seq = answer_start_scores[i].unsqueeze(0)
+            answer_start = torch.argmax(answer_start_scores_one_seq)
+            answer_end_scores_one_seq = answer_end_scores[i].unsqueeze(0)
+            answer_end = torch.argmax(answer_end_scores_one_seq) + 1
+            prediction = self.tokenizer.convert_tokens_to_string(
+                self.tokenizer.convert_ids_to_tokens(
+                    input_ids[i].tolist()[answer_start:answer_end]
+                )
+            )
+            inferences.append(prediction)
+        return inferences
+
+    def postprocess(self, prediction_results: List[str]) -> Dict:
+        return {"predictions": prediction_results}
+     
+
@@ -0,0 +1,37 @@
+import os
+import numpy as np
+import pickle
+import xgboost as xgb
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import make_blobs
+from xgboost import XGBClassifier
+
+
+class ClassifierPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            booster = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            X, y = make_blobs(n_samples=100, centers=2, n_features=2, random_state=1)
+            model = XGBClassifier()
+            model.fit(X, y)
+            booster = model.get_booster()
+        self._booster = booster
+
+    def preprocess(self, prediction_input: dict) -> xgb.DMatrix:
+        instances = prediction_input["instances"]
+        return xgb.DMatrix(instances)
+
+    def predict(self, instances: xgb.DMatrix) -> np.ndarray:
+        return self._booster.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}
@@ -0,0 +1,41 @@
+import os
+import numpy as np
+import pandas as pd
+import pickle
+import xgboost as xgb
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+
+class XGBRankerPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            booster = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+            self._booster = booster
+        else:
+            N = 500
+            dates = pd.date_range(start='2023-01-01', end='2023-01-12', periods=N)
+            X = pd.DataFrame(np.random.randn(N, 5), columns=list('ABCDE'), index=dates)
+            y = pd.Series(np.random.randint(0, 10, size=N), index=dates, name='label')
+            group = X.groupby(dates + pd.offsets.MonthEnd(0)).size()
+            sample_weight = pd.Series(np.arange(len(group)), index=group.index)
+            model = xgb.XGBRanker(objective='rank:pairwise', max_depth=3, learning_rate=0.1, booster='gbtree', tree_method='hist', n_jobs=4, n_estimators=50, enable_categorical=False, random_state=42)
+            model.fit(X=X, y=y, group=group, sample_weight=sample_weight, verbose=True)
+            booster = model.get_booster()
+            self._booster = booster
+
+    def preprocess(self, prediction_input: dict) -> xgb.DMatrix:
+        instances = prediction_input["instances"]
+        return xgb.DMatrix(instances)
+
+    def predict(self, instances: xgb.DMatrix) -> np.ndarray:
+        return self._booster.predict(instances, output_margin=False, ntree_limit=0)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}
@@ -1,227 +0,0 @@
-# Benchmark report on fine tuning the OpenLLaMA 7B model on Google Cloud Vertex Model Garden
-
-Gary Wei, Software Engineer, Google Cloud
-Dustin Luong, Software Engineer, Google Cloud
-Changyu Zhu, Software Engineer, Google Cloud
-Genquan Duan, Software Engineer, Google Cloud
-
-## Introduction
-
-Fine-tuning of LLMs can be non-trivial to find an optimal configuration of
-machine types, training parameters, and other hyperparameters that achieves a
-good balance between cost efficiency and model performance. To facilitate users
-in conducting tuning experiments, this report benchmarks OpenLLaMA 7B
-fine-tuning on Google Cloud Vertex Model Garden, demonstrating both efficiency
-and effectiveness. The observations are general and can be applied to other LLM
-models.
-
-We benchmarked fine tuning algorithms [LoRA](https://arxiv.org/abs/2106.09685)
-and [QLoRA](https://arxiv.org/abs/2305.14314) supported by
-[huggingface PEFT libraries](https://github.com/huggingface/peft). LoRA, short
-for Low-Rank Adaptation of Large Language Models, is an improved fine tuning
-method where instead of fine tuning all the weights that constitute the weight
-matrix of the pre-trained large language model, two smaller matrices that
-approximate this larger matrix are fine-tuned. QLoRA is an even more
-memory-efficient version of LoRA, where the pretrained model is loaded to GPU
-memory as quantized 4-bit weights, while preserving similar effectiveness to
-LoRA. We also provide simple scripts and parameter settings to reproduce the
-results reported in this report.
-
-In general, there are many factors that affect the performance of fine-tuning
-experiments, such as hardware settings, parameters, cost, and accuracy. It is
-impractical to obtain benchmarks for all possible combinations of these factors.
-Instead, we focus on tuning a subset of related parameters and evaluating their
-impact on a set of chosen metrics. The evaluation metrics are GPU memory usage,
-percentage of parameters tuned, tuning speed, cost, and accuracy. The tuning
-parameters are batch size, lora rank, maximum sequence length, and maximum
-training steps.
-
-## Key takeaways
-
-   **Use QLoRA to minimize the peak GPU requirements**: The QLoRA can
-    significantly reduce the peak GPU memory usage by ~75% compared to LoRA. For
-    OpenLLaMA7b, the peak memory is ~28G for LoRA and ~7G for QLoRA.
-   **Use LoRA to maximize the tuning speed and minimize the tuning cost**: LoRA
-    is ~66% faster than QLoRA in fine tuning speed. LoRA/QLoRA tuning cost is
-    low generally, while LoRA is even ~40% cheaper than QLoRA with the same
-    parameters. Suggest to use QLoRA for limited GPU memories, and LoRA for
-    limited training budgets. For OpenLLaMA7b, the tuning speed for LoRA/QLoRA
-    ~5 samples / 3 samples per second, and the tuning cost for LoRA/QLoRA in 500
-    steps is ~$1/$1.7 on `a2-highgpu-1g` with 1 A100 40G GPU. The tuning cost
-    for QLoRA in 500 steps is $6.75 on n1-standard-8 with 1 V100 GPU, while LoRA
-    could not run because of OOM.
-   **Use QLoRA to tune models with large sequence lengths**. For OpenLLaMA7b,
-    the max sequence length for QLoRA can be 2048 when consuming 16.3G GPU,
-    while the max sequence length for LoRA is 512 when consuming 28.2G GPU, and
-    encounter OOM when max sequence length is 1024.
-   **Both LoRA and QLoRA give similar accuracy improvement after fine tuning.**
-    For OpenLLaMA7b, both LoRA/QLoRA can improve the average accuracy by ~4%
-    evaluating on 3 typical tasks (ARC challenge, HellaSwag and TruthfulQA),
-    after training 1875 steps on dataset
-    [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco).
-   **Use a big batch size if GPU memory is not a constraint**. For OpenLLaMA7b
-    with other default parameters, we suggest using a batch size as 24 for
-    QLoRA, but 2 for LoRA when tuning with 1 A100 40G. We also suggest using a
-    batch size as 8 for QLoRA when tuning with 1 V100. Tuning with LoRA and
-    batch size as 1 got OOM and we don't recommend tuning LoRA with 1 V100.
-
-## Benchmark Details
-
-### Experiment Setup
-
-The benchmark dataset is
-[timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco).
-The training dataset is directly downloaded from hugging face to the VM, before
-every experiment.
-
-The default tuning parameters during benchmark are:
-
-   Host VM: a2-highgpu-1g
-   Accelerator type: 1 A100 40G
-   batch size: 2
-   lora_rank: 16
-   max_seq_length: 512
-   precision_mode: float16
-   max_train_steps: 500
-
-For simplicity, we set the precision mode to `float16` when tuning LoRA models,
-and set the precision to `4bit` for QLoRA.
-
-Sample script to start fine tuning dockers in a VM on GCP.
-
-```shell
-IMAGE_TAG=us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-peft-train:latest
-docker run  --runtime=nvidia  -e NVIDIA_VISIBLE_DEVICES=0 \
--rm --name "test_gpu"  -it --pull=always ${IMAGE_TAG} \
--task=instruct-lora \
--pretrained_model_id=openlm-research/open_llama_7b \
--dataset_name="timdettmers/openassistant-guanaco"  \
--instruct_column_in_dataset="text" \
--precision_mode="float16" \
--output_dir=<OUTPUT DIR>  \
--lora_rank=2 \
--max_sequence_length=512 \
--learning_rate=2e-4 \
--max_steps=50
-```
-
-### GPU Memory
-
-In this benchmark, we investigated the impact of batch size, lora rank, and
-maximum sequence length on GPU memory, and then made recommendations on the
-maximum batch size for different GPUs.
-
-#### Peak GPU memory by batch size (GB)
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-peak-gpu-vs-batch-size.png" width="600">
-
-   The QLoRA can significantly reduce the peak GPU memory usage by ~75%
-    compared to LoRA. The peak GPU memory is ~28G for LoRA and ~7G for QLoRA
-    when batch size is 2.
-   QLoRA can support much larger batch sizes than LoRA
-    -   We can use a batch size as 32 for QLoRA, but only 2 for LoRA on 1 A100
-        40G.
-    -   We can use a batch size of 8 for QLoRA on 1 V100 GPU. LoRA will fail
-        with OOM even with a batch size of 1.
-
-#### Peak GPU memory by LoRA rank (GB)
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-peak-gpu-vs-lora-rank.png" width="600">
-
-   Peak GPU memories are quite similar for different LoRA ranks for both
-    LoRA/QLoRA.
-   The peak GPU memory increasing percentages are very small generally when
-    LoRA rank increases.
-    -   The peak GPU memory increases from 28G with LoRA rank 4 to 29.09G with
-        LoRA rank 64, and the increasing percentage is only ~3.9%.
-
-#### Peak GPU memory by max sequence length for LoRA/QLoRA (GB)
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-peak-gpu-vs-max-seq-length.png" width="600">
-
-   The peak GPU increases quickly when max sequence length increases for both
-    LoRA/QLoRA, and the increasing rate of LoRA is much faster than QLoRA.
-    -   For LoRA tuning, the GPU memory increased from 20.5G (max sequence
-        length=256) to 28.2G (max sequence length=512), an increase of ~37%.
-    -   For QLoRA tuning, the GPU memory increased from 6.94G (max sequence
-        length=256) to 7.57G (max sequence length=512), an increase of ~9%.
-   The max sequence length for QLoRA can be 2048 when consuming 16.3G GPU,
-    while the max sequence length for LoRA is 512 when consuming 28.2G GPU, and
-    encounter OOM when max sequence length is 1024.
-
-### Fine Tuning Parameters
-
-This section shows the number/percentage of trainable parameters, and the sizes
-of the fine tuned models. LoRA and QLoRA differ only in how they represent the
-precision of their parameters. The total number of parameters and the number of
-trainable parameters are the same for both methods.
-
-| LoRA Rank | Finetuned parameters | Total parameters | Trainable Parameter Percentage | Fine tuned model size (MB) |
-| --------- | -------------------- | ---------------- | ------------------------------ | -------------------------- |
-| 8         | 2.00E+07             | 6.76E+09         | 0.3%                           | 76.4                       |
-| 16        | 4.00E+07             | 6.78E+09         | 0.6%                           | 152.65                     |
-| 32        | 8.00E+07             | 6.82E+09         | 1.2%                           | 305.15                     |
-| 64        | 1.60E+08             | 6.90E+09         | 2.3%                           | 610.15                     |
-
-
-LoRA/QLoRA tunes quite a small fraction (only 0.3% with LoRA rank=8) of all
-parameters, and the tuned models are very small (only 76.4MB with LoRA rank=8).
-
-### Fine Tuning Speed And Costs
-
-The fine-tuning speed and cost are affected by various factors, such as the
-GPUs, LoRA ranks, and max sequence lengths.
-
-   LoRA is ~66% faster than QLoRA in fine tuning speed. The tuning speed for
-    LoRA/QLoRA ~5 samples / 3 samples per second on 1 A100 40G GPU
-   Higher LoRA ranks, slower tuning speed for both LoRA/QLoRA.
-    -   LoRA tuning speed reduces from ~5 samples per second with LoRA rank as 8
-        to ~4 samples per second with LoRA rank as 64, slowed down by 20%.
-    -   QLoRA tuning speed reduces from ~3 samples per second with LoRA rank as
-        8 to ~2.5 samples per second with LoRA rank as 64, slowed down by 17%.
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-tune-speed-vs-lora-rank.png" width="600">
-
-   Longer sequence lengths, slower tuning speed.
-    -   LoRA tuning speed reduces from ~5.56 samples per second with max
-        sequence length as 256 to ~4.84 samples per second with max sequence
-        length as 512 slowed down by 13%.
-    -   LoRA tuning speed reduces from ~2.95 samples per second with max
-        sequence length as 256 to ~2.88 samples per second with max sequence
-        length as 512 slowed down by ~2.4%.
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-tune-speed-lora-qlora.png" width="600">
-
-   LoRA/QLoRA tuning cost is low generally, while LoRA is even ~40% cheaper
-    than QLoRA with the same parameters.
-    -   The LoRA/QLoRA fine tuning cost for 500 steps is ~$1/$1.7 on 1 A100 40G.
-    -   The tuning cost for QLoRA in 500 steps is $6.75 on n1-standard-8 with 1
-        V100 GPU, while LoRA could not run because of OOM.
-
-<img src="images/openllama_7b_fine_tune_benchmark_report/openllama-7b-tune-cost-lora-qlora.png" width="600">
-
-### Accuracy
-
-We fine tuned Open Llama 7B model with
-[timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco),
-and report accuracy similar to the
-[HuggingFace leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
-using
-[Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness).
-[HuggingFace leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
-mainly compares models on ARC, HellaSwag, MMLU, and TruthfulQA. The authors did
-not publish OpenLLaMA 7B on MMLU
-([link](https://huggingface.co/openlm-research/open_llama_7b)). Therefore, we
-only benchmark accuracies on ARC, HellaSwag, and TruthfulQA.
-
-|                                                              | Mean | ARC  | HellaSwag | TruthfulQA | Tuning Parameters                                            |
-| ------------------------------------------------------------ | ---- | ---- | --------- | ---------- | ------------------------------------------------------------ |
-| OpenLLaMA7B ([Original Report](https://huggingface.co/openlm-research/open_llama_7b)) | 0.49 | 0.41 | 0.73      | 0.34       | n/a                                                          |
-| OpenLLaMA7B ([Re-run with lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)) | 0.51 | 0.47 | 0.72      | 0.35       | n/a                                                          |
-| OpenLLaMA7B+LoRA                                             | 0.56 | 0.48 | 0.74      | 0.45       | LoRA Rank=16; Max Sequence Length=512;Learning Rate=1e-4; Train steps=1875 |
-| OpenLLaMA7B+QLoRA                                            | 0.53 | 0.45 | 0.73      | 0.42       | LoRA Rank=16; Max Sequence Length=512; Learning Rate=1e-4; Train steps=1875 |
-
-   The base OpenLLaMA7B model gets better performance (2%) when using the
-    [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness).
-   LoRA/QLoRA can improve the performance by ~2-4% when trained for 1875 steps
-    with learning rate 1e-4.
@@ -1,6 +1,6 @@
 --find-links https://download.pytorch.org/whl/torch_stable.html

-torch==2.0.1+cu118
+torch==2.2.0
 numpy==1.26.1
 absl_py==2.0.0
 accelerate==0.24.0
@@ -16,7 +16,7 @@ Pillow==10.3.0
 rawpy==0.18.1
 scipy==1.11.3
 scikit-image==0.22.0
-scikit-learn==1.3.2
+scikit-learn==1.5.0
 tensorboard==2.15.0
 tensorboardX==2.6.2.2
 tqdm==4.66.3
@@ -6,8 +6,10 @@ import io
 import json
 import os
 import subprocess
+import time
 from typing import Any, Dict, Sequence

+from google import auth
 from google.cloud import storage
 import matplotlib.pyplot as plt
 import numpy as np
@@ -85,7 +87,9 @@ def get_job_name_with_datetime(prefix: str) -> str:
  Returns:
    A job name.
  """
-  return prefix + datetime.datetime.now().strftime("_%Y%m%d_%H%M%S")
+  now = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+  job_name = f"{prefix}-{now}".replace("_", "-")
+  return job_name


 def create_job_name(prefix: str) -> str:
@@ -99,7 +103,7 @@ def create_job_name(prefix: str) -> str:
  """
  user = os.environ.get("USER")
  now = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
-  job_name = f"{prefix}-{user}-{now}"
+  job_name = f"{prefix}-{user}-{now}".replace("_", "-")
  return job_name


@@ -228,7 +232,23 @@ def download_image(url: str) -> str:
    base64 encoded image.
  """
  response = requests.get(url)
-  return Image.open(io.BytesIO(response.content))
+  return Image.open(io.BytesIO(response.content))  # pytype: disable=bad-return-type  # pillow-102-upgrade
+
+
+def resize_image(image: Any, new_width: int = 1000) -> Any:
+  """Resizes an image to a certain width.
+
+  Args:
+    image: The image which has to be resized.
+    new_width: New width of the image.
+
+  Returns:
+    New resized image.
+  """
+  width, height = image.size
+  new_height = int(height * new_width / width)
+  new_img = image.resize((new_width, new_height))
+  return new_img


 def load_img(path: str) -> Any:
@@ -307,6 +327,150 @@ def get_prediction_instances(test_filepath: str, new_width: int = -1) -> Any:
  return instances


+def vqa_predict(
+    endpoint: Any,
+    question_prompts: Sequence[str],
+    image: Any,
+    language_code: str = "en",
+    new_width: int = 1000,
+    use_dedicated_endpoint: bool = False,
+) -> Sequence[str]:
+  """Predicts the answer to a question about an image using an Endpoint."""
+  # Resize and convert image to base64 string.
+  resized_image = resize_image(image, new_width)
+  resized_image_base64 = image_to_base64(resized_image)
+
+  instances = []
+  if question_prompts:
+    # Format question prompt
+    question_prompt_format = "answer {} {}\n"
+    for question_prompt in question_prompts:
+      if question_prompt:
+        instances.append({
+            "prompt": question_prompt_format.format(
+                language_code, question_prompt
+            ),
+            "image": resized_image_base64,
+        })
+  else:
+    instances.append({
+        "image": resized_image_base64,
+    })
+
+  response = endpoint.predict(
+      instances=instances, use_dedicated_endpoint=use_dedicated_endpoint
+  )
+  return [pred.get("response") for pred in response.predictions]
+
+
+def caption_predict(
+    endpoint: Any,
+    language_code: str,
+    image: Any,
+    caption_prompt: bool = False,
+    new_width: int = 1000,
+    use_dedicated_endpoint: bool = False,
+) -> str:
+  """Predicts a caption for a given image using an Endpoint."""
+  # Resize and convert image to base64 string.
+  resized_image = resize_image(image, new_width)
+  resized_image_base64 = image_to_base64(resized_image)
+
+  instance = {"image": resized_image_base64}
+
+  if caption_prompt:
+    # Format caption prompt
+    caption_prompt_format = "caption {}\n"
+    instance["prompt"] = caption_prompt_format.format(language_code)
+
+  instances = [instance]
+  response = endpoint.predict(
+      instances=instances, use_dedicated_endpoint=use_dedicated_endpoint
+  )
+  return response.predictions[0].get("response")
+
+
+def ocr_predict(
+    endpoint: Any,
+    ocr_prompt: str,
+    image: Any,
+    new_width: int = 1000,
+    use_dedicated_endpoint: bool = False,
+) -> str:
+  """Extracts text from a given image using an Endpoint."""
+  # Resize and convert image to base64 string.
+  resized_image = resize_image(image, new_width)
+  resized_image_base64 = image_to_base64(resized_image)
+
+  instance = {"image": resized_image_base64}
+  if ocr_prompt:
+    instance["prompt"] = ocr_prompt
+  instances = [instance]
+
+  response = endpoint.predict(
+      instances=instances, use_dedicated_endpoint=use_dedicated_endpoint
+  )
+  return response.predictions[0].get("response")
+
+
+def detect_predict(
+    endpoint: Any,
+    detect_prompt: str,
+    image: Any,
+    new_width: int = 1000,
+    use_dedicated_endpoint: bool = False,
+) -> str:
+  """Predicts the answer to a question about an image using an Endpoint."""
+  # Resize and convert image to base64 string.
+  resized_image = resize_image(image, new_width)
+  resized_image_base64 = image_to_base64(resized_image)
+
+  instance = {"image": resized_image_base64}
+  if detect_prompt:
+    instance["prompt"] = detect_prompt
+  instances = [instance]
+
+  response = endpoint.predict(
+      instances=instances, use_dedicated_endpoint=use_dedicated_endpoint
+  )
+  return response.predictions[0].get("response")
+
+
+def copy_model_artifacts(
+    model_id: str,
+    model_source: str,
+    model_destination: str,
+) -> None:
+  """Copies model artifacts from model_source to model_destination.
+
+  model_source and model_destination should be GCS path.
+
+  Args:
+    model_id: The model id.
+    model_source: The source of the model artifact.
+    model_destination: The destination of the model artifact.
+  """
+  if not model_source.startswith(GCS_URI_PREFIX):
+    raise ValueError(
+        f"{model_source} is not a GCS path starting with {GCS_URI_PREFIX}."
+    )
+  if not model_destination.startswith(GCS_URI_PREFIX):
+    raise ValueError(
+        f"{model_destination} is not a GCS path starting with {GCS_URI_PREFIX}."
+    )
+  model_source = f"{model_source}/{model_id}"
+  model_destination = f"{model_destination}/{model_id}"
+  print("Copying model artifact from ", model_source, " to ", model_destination)
+  subprocess.check_output([
+      "gcloud",
+      "storage",
+      "cp",
+      "-r",
+      model_source,
+      model_destination,
+  ])
+
+
 def get_quota(project_id: str, region: str, resource_id: str) -> int:
  """Returns the quota for a resource in a region.

@@ -357,40 +521,76 @@ def get_quota(project_id: str, region: str, resource_id: str) -> int:
  return -1


-def get_resource_id(accelerator_type: str, is_for_training: bool) -> str:
+def get_resource_id(
+    accelerator_type: str,
+    is_for_training: bool,
+    is_restricted_image: bool = False,
+    is_dynamic_workload_scheduler: bool = False,
+) -> str:
  """Returns the resource id for a given accelerator type and the use case.

  Args:
    accelerator_type: The accelerator type.
    is_for_training: Whether the resource is used for training. Set false for
      serving use case.
+    is_restricted_image: Whether the image is hosted in `vertex-ai-restricted`.
+    is_dynamic_workload_scheduler: Whether the resource is used with Dynamic
+      Workload Scheduler.

  Returns:
    The resource id.
  """
-  training_accelerator_map = {
-      "NVIDIA_TESLA_V100": "custom_model_training_nvidia_v100_gpus",
-      "NVIDIA_L4": "custom_model_training_nvidia_l4_gpus",
-      "NVIDIA_TESLA_A100": "custom_model_training_nvidia_a100_gpus",
-      "NVIDIA_TESLA_T4": "custom_model_training_nvidia_t4_gpus",
-      "TPU_V5e": "custom_model_training_tpu_v5e",
-      "TPU_V3": "custom_model_training_tpu_v3",
+  accelerator_suffix_map = {
+      "NVIDIA_TESLA_V100": "nvidia_v100_gpus",
+      "NVIDIA_TESLA_P100": "nvidia_p100_gpus",
+      "NVIDIA_L4": "nvidia_l4_gpus",
+      "NVIDIA_TESLA_A100": "nvidia_a100_gpus",
+      "NVIDIA_A100_80GB": "nvidia_a100_80gb_gpus",
+      "NVIDIA_H100_80GB": "nvidia_h100_gpus",
+      "NVIDIA_H100_MEGA_80GB": "nvidia_h100_mega_gpus",
+      "NVIDIA_TESLA_T4": "nvidia_t4_gpus",
+      "TPU_V5e": "tpu_v5e",
+      "TPU_V3": "tpu_v3",
+  }
+  default_training_accelerator_map = {
+      key: f"custom_model_training_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
+  }
+  dws_training_accelerator_map = {
+      key: f"custom_model_training_preemptible_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
+  }
+  restricted_image_training_accelerator_map = {
+      "NVIDIA_A100_80GB": "restricted_image_training_nvidia_a100_80gb_gpus",
  }
  serving_accelerator_map = {
-      "NVIDIA_TESLA_V100": "custom_model_serving_nvidia_v100_gpus",
-      "NVIDIA_L4": "custom_model_serving_nvidia_l4_gpus",
-      "NVIDIA_TESLA_A100": "custom_model_serving_nvidia_a100_gpus",
-      "NVIDIA_TESLA_T4": "custom_model_serving_nvidia_t4_gpus",
-      "TPU_V5e": "custom_model_serving_tpu_v5e",
+      key: f"custom_model_serving_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
  }
+
  if is_for_training:
+    if is_restricted_image and is_dynamic_workload_scheduler:
+      raise ValueError(
+          "Dynamic Workload Scheduler does not work for restricted image"
+          " training."
+      )
+    training_accelerator_map = (
+        restricted_image_training_accelerator_map
+        if is_restricted_image
+        else default_training_accelerator_map
+    )
    if accelerator_type in training_accelerator_map:
-      return training_accelerator_map[accelerator_type]
+      if is_dynamic_workload_scheduler:
+        return dws_training_accelerator_map[accelerator_type]
+      else:
+        return training_accelerator_map[accelerator_type]
    else:
      raise ValueError(
          f"Could not find accelerator type: {accelerator_type} for training."
      )
  else:
+    if is_dynamic_workload_scheduler:
+      raise ValueError("Dynamic Workload Scheduler does not work for serving.")
    if accelerator_type in serving_accelerator_map:
      return serving_accelerator_map[accelerator_type]
    else:
@@ -405,9 +605,16 @@ def check_quota(
    accelerator_type: str,
    accelerator_count: int,
    is_for_training: bool,
+    is_restricted_image: bool = False,
+    is_dynamic_workload_scheduler: bool = False,
 ):
  """Checks if the project and the region has the required quota."""
-  resource_id = get_resource_id(accelerator_type, is_for_training)
+  resource_id = get_resource_id(
+      accelerator_type,
+      is_for_training=is_for_training,
+      is_restricted_image=is_restricted_image,
+      is_dynamic_workload_scheduler=is_dynamic_workload_scheduler,
+  )
  quota = get_quota(project_id, region, resource_id)
  quota_request_instruction = (
      "Either use "
@@ -427,3 +634,76 @@ def check_quota(
        f"Quota not enough for {resource_id} in {region}: {quota} <"
        f" {accelerator_count}. {quota_request_instruction}"
    )
+
+
+def get_deploy_source() -> str:
+  """Gets deploy_source string based on running environment."""
+  vertex_product = os.environ.get("VERTEX_PRODUCT", "")
+  match vertex_product:
+    case "COLAB_ENTERPRISE":
+      return "notebook_colab_enterprise"
+    case "WORKBENCH_INSTANCE":
+      return "notebook_workbench"
+    case _:
+      # Legacy workbench, legacy colab, or other custom environments.
+      return "notebook_environment_unspecified"
+
+
+def _is_operation_done(op_name: str, region: str) -> bool:
+  """Checks if the operation is done.
+
+  Args:
+    op_name: The name of the operation to poll.
+    region: The region of the operation.
+
+  Returns:
+    True if the operation is done, False otherwise.
+
+  Raises:
+    ValueError: If the operation failed.
+  """
+  creds, _ = auth.default()
+  auth_req = auth.transport.requests.Request()
+  creds.refresh(auth_req)
+  headers = {
+      "Authorization": f"Bearer {creds.token}",
+  }
+  url = f"https://{region}-aiplatform.googleapis.com/ui/{op_name}"
+  response = requests.get(url, headers=headers)
+  operation_data = response.json()
+  if "error" in operation_data:
+    raise ValueError(f"Operation failed: {operation_data['error']}")
+  return operation_data.get("done", False)
+
+
+def poll_and_wait(
+    op_name: str, region: str, total_wait: int, interval: int = 60
+) -> None:
+  """Polls the operation and waits for it to complete.
+
+  Args:
+    op_name: The name of the operation to poll.
+    region: The region of the operation.
+    total_wait: The total wait time in seconds.
+    interval: The interval between each poll in seconds.
+
+  Raises:
+    TimeoutError: If the operation times out.
+  """
+  start_time = time.time()
+  while True:
+    if _is_operation_done(op_name, region):
+      break
+    time_elapsed = time.time() - start_time
+    if time_elapsed > total_wait:
+      raise TimeoutError(
+          f"Operation timed out after {int(time_elapsed)} seconds."
+      )
+    print(
+        "\rStill waiting for operation... Elapsed time in seconds:"
+        f" {int(time_elapsed):<6}",
+        end="",
+        flush=True,
+    )
+    time.sleep(interval)
+
@@ -0,0 +1,570 @@
+"""Functions for dataset validation.
+
+This tool is used to validate the dataset against the given template.
+"""
+
+import json
+import multiprocessing
+import os
+import subprocess
+from typing import Any, Callable, Dict, Tuple, Union
+from absl import logging
+import accelerate
+import datasets
+import transformers
+
+GCS_URI_PREFIX = "gs://"
+GCSFUSE_URI_PREFIX = "/gcs/"
+LOCAL_BASE_MODEL_DIR = "/tmp/base_model_dir"
+LOCAL_TEMPLATE_DIR = "/tmp/template_dir"
+_TEMPLATE_DIRNAME = "templates"
+_VERTEX_AI_SAMPLES_GITHUB_REPO_NAME = "vertex-ai-samples"
+_VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR = (
+    "community-content/vertex_model_garden/model_oss/peft/train/vmg/templates"
+)
+_MODELS_REQUIRING_PAD_TOKEN = ("llama", "falcon", "mistral", "mixtral")
+_MODELS_REQUIRING_EOS_TOEKN = ("gemma-2b", "gemma-7b")
+_DESCRIPTION_KEY = "description"
+_SOURCE_KEY = "source"
+_PROMPT_INPUT_KEY = "prompt_input"
+_PROMPT_NO_INPUT_KEY = "prompt_no_input"
+_RESPONSE_SEPARATOR = "response_separator"
+_INSTRUCTION_SEPARATOR = "instruction_separator"
+_CHAT_TEMPLATE_KEY = "chat_template"
+_KNOWN_KEYS = (
+    _DESCRIPTION_KEY,
+    _SOURCE_KEY,
+    _PROMPT_INPUT_KEY,
+    _PROMPT_NO_INPUT_KEY,
+    _RESPONSE_SEPARATOR,
+    _INSTRUCTION_SEPARATOR,
+    _CHAT_TEMPLATE_KEY,
+)
+
+
+def is_gcs_path(input_path: str) -> bool:
+  """Checks if the input path is a Google Cloud Storage (GCS) path.
+
+  Args:
+      input_path: The input path to be checked.
+
+  Returns:
+      True if the input path is a GCS path, False otherwise.
+  """
+  return input_path is not None and input_path.startswith(GCS_URI_PREFIX)
+
+
+def force_gcs_fuse_path(gcs_uri: str) -> str:
+  """Converts gs:// uris to their /gcs/ equivalents. No-op for other uris.
+
+  Args:
+    gcs_uri: The GCS URI to convert.
+
+  Returns:
+    The converted GCS URI.
+  """
+  if is_gcs_path(gcs_uri):
+    return GCSFUSE_URI_PREFIX + gcs_uri[len(GCS_URI_PREFIX) :]
+  else:
+    return gcs_uri
+
+
+def download_gcs_uri_to_local(
+    gcs_uri: str,
+    destination_dir: str = LOCAL_BASE_MODEL_DIR,
+    check_path_exists: bool = True,
+) -> str:
+  """Downloads GCS URI to local.
+
+  If GCS URI is a directory, gs://some/folder is downloaded to
+  /destination_dir/folder. If GCS URI is a file, gs://some/file is downloaded to
+  /destination_dir/file.
+
+  Args:
+    gcs_uri: GCS URI to download.
+    destination_dir: Local directory directory.
+    check_path_exists: Whether to check if the path exists.
+
+  Returns:
+    Local path to target folder/file.
+  """
+  target = os.path.join(
+      destination_dir,
+      os.path.basename(os.path.normpath(gcs_uri)),
+  )
+  if check_path_exists and os.path.exists(target):
+    logging.info("File %s already exists.", target)
+    return target
+  if accelerate.PartialState().is_local_main_process:
+    logging.info(
+        "Downloading file(s) from %s to %s...", gcs_uri, destination_dir
+    )
+    if not os.path.exists(destination_dir):
+      os.mkdir(destination_dir)
+    subprocess.check_output([
+        "gsutil",
+        "-m",
+        "cp",
+        "-r",
+        gcs_uri,
+        destination_dir,
+    ])
+    logging.info("Downloaded file(s) from %s to %s.", gcs_uri, destination_dir)
+  # Make sure ALL processes process to next step after data downloading is done.
+  # It matters for the main process to wait for other processes as well.
+  accelerate.PartialState().wait_for_everyone()
+  return target
+
+
+def get_template(template_path: str) -> Dict[str, str]:
+  """Gets the template dictionary given the file path.
+
+  Args:
+    template_path: Path to the template file.
+
+  Returns:
+    A dictionary of the template.
+
+  Raises:
+    ValueError: If the template file does not exist or contains unknown keys.
+  """
+  if is_gcs_path(template_path):
+    template_path = force_gcs_fuse_path(template_path)
+  elif not os.path.isfile(template_path):
+    template_path = os.path.join(
+        os.path.dirname(__file__),
+        _TEMPLATE_DIRNAME,
+        template_path + ".json",
+    )
+  if not os.path.isfile(template_path):
+    raise ValueError(f"Template file {template_path} does not exist.")
+  with open(template_path, "r") as f:
+    template_json: dict[str, str] = json.load(f)
+    for key in template_json:
+      if key not in _KNOWN_KEYS:
+        raise ValueError(f"Unknown key {key} in template {template_path}.")
+    return template_json
+
+
+def get_response_separator(template_json: Dict[str, str]) -> Union[str, None]:
+  return template_json.get(_RESPONSE_SEPARATOR, None)
+
+
+def get_instruction_separator(
+    template_json: Dict[str, str],
+) -> Union[str, None]:
+  return template_json.get(_INSTRUCTION_SEPARATOR, None)
+
+
+def _format_template_fn(
+    template: str,
+    input_column: str,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> Callable[[Dict[str, str]], Dict[str, str]]:
+  """Formats a dataset example according to a template.
+
+  Args:
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    input_column: The input column in the dataset to be used or updated by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    A function that formats data according to the template.
+  """
+  template_json = get_template(template)
+
+  if _CHAT_TEMPLATE_KEY not in template_json:
+
+    def format_fn(example: Dict[str, str]) -> Dict[str, str]:
+      format_dict = {key: value for key, value in example.items()}
+      if format_dict.get(input_column):
+        format_str = template_json[_PROMPT_INPUT_KEY]
+      elif _PROMPT_NO_INPUT_KEY in template_json:
+        format_str = template_json[_PROMPT_NO_INPUT_KEY]
+      else:
+        raise KeyError(
+            f"The template {os.path.basename(template)} does not contain"
+            f" {_PROMPT_INPUT_KEY} or {_PROMPT_NO_INPUT_KEY} key."
+        )
+      try:
+        return {input_column: format_str.format(**format_dict)}
+      except KeyError as e:
+        raise KeyError(
+            f"The template {os.path.basename(template)} contains a key {e} in"
+            f" {_PROMPT_INPUT_KEY} or {_PROMPT_NO_INPUT_KEY} that does not"
+            " exist in the dataset example. The dataset example looks like"
+            f" {format_dict}."
+        ) from e
+
+    return format_fn
+  elif (
+      _PROMPT_INPUT_KEY in template_json
+      or _PROMPT_NO_INPUT_KEY in template_json
+  ):
+    raise ValueError(
+        f"chat_template templates do not support {_PROMPT_INPUT_KEY} or"
+        f" {_PROMPT_NO_INPUT_KEY} templates."
+    )
+  else:
+    if tokenizer is None:
+      raise ValueError("A tokenizer is required for chat_template templates.")
+    # Assign HuggingFace jinja template.
+    tokenizer.chat_template = template_json[_CHAT_TEMPLATE_KEY]
+
+    def format_fn(example: Dict[str, str]) -> Dict[str, str]:
+      try:
+        return {
+            input_column: tokenizer.apply_chat_template(
+                example[input_column],
+                tokenize=False,
+                add_generation_prompt=False,
+            )
+        }
+      except KeyError as e:
+        raise KeyError(
+            f"The template {os.path.basename(template)} contains a key {e} in"
+            f" {_CHAT_TEMPLATE_KEY} that does not exist in the dataset example."
+        ) from e
+
+    return format_fn
+
+
+def _get_split_string(
+    split: str,
+    dataset_percent: int | None = None,
+    dataset_k_rows: int | None = None,
+) -> str:
+  """Gets the formatted split string for the dataset.
+
+  This is used to format the split string as per
+  https://huggingface.co/docs/datasets/v2.21.0/loading#slice-splits. Also, this
+  function will only be used to load the partial dataset for validating the
+  dataset against the template.
+
+  Args:
+    split: Split of the dataset.
+    dataset_percent: The percentage of the dataset to load.
+    dataset_k_rows: The top k sequences to load from the dataset.
+
+  Returns:
+    A formatted split string.
+  """
+  # Validate the dataset_percent and dataset_k_rows values.
+  if dataset_percent and dataset_k_rows:
+    raise ValueError(
+        "You can set either validate_percentage_of_dataset or"
+        " validate_k_rows_of_dataset, but not both."
+    )
+
+  if dataset_percent:
+    logging.info("Loading %d percent of the dataset...", dataset_percent)
+    return f"{split}[:{dataset_percent}%]"
+
+  if dataset_k_rows:
+    logging.info("Loading top %d rows of the dataset...", dataset_k_rows)
+    return f"{split}[:{dataset_k_rows}]"
+
+  return split
+
+
+def _github_template_path(template: str) -> str:
+  """Generates the path to the template in the Vertex AI Samples GitHub repo.
+
+  Args:
+    template: Name of the template.
+
+  Returns:
+    The path to the template in the Vertex AI Samples GitHub repo.
+  """
+  # vertex-ai-samples directory may lie under separate directory depending on
+  # the scratch_dir parameter in the notebook execution environment.
+  vertex_ai_samples_abs_path = os.getcwd().split(
+      _VERTEX_AI_SAMPLES_GITHUB_REPO_NAME
+  )[0]
+  return os.path.join(
+      vertex_ai_samples_abs_path,
+      _VERTEX_AI_SAMPLES_GITHUB_REPO_NAME,
+      _VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR,
+      template + ".json",
+  )
+
+
+def _get_dataset(
+    dataset_name: str,
+    split: str,
+    num_proc: int | None = None,
+) -> datasets.DatasetDict:
+  """Gets a dataset.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    num_proc: Number of processors to use.
+
+  Returns:
+    A dataset.
+  """
+  dataset_name = force_gcs_fuse_path(dataset_name)
+  if os.path.isfile(dataset_name):
+    # Custom dataset.
+    return datasets.load_dataset(
+        "json",
+        data_files=[dataset_name],
+        split=split,
+        num_proc=num_proc,
+    )
+  # HF dataset.
+  return datasets.load_dataset(dataset_name, split=split, num_proc=num_proc)
+
+
+def should_add_pad_token(model_id: str) -> bool:
+  """Returns whether the model requires adding a special pad token.
+
+  Args:
+    model_id: The name of the model.
+
+  Returns:
+    True if the model requires adding a special pad token, False otherwise.
+  """
+  return any(s.lower() in model_id.lower() for s in _MODELS_REQUIRING_PAD_TOKEN)
+
+
+def should_add_eos_token(model_id: str) -> bool:
+  """Returns whether the model requires adding a special eos token.
+
+  Args:
+    model_id: The name of the model.
+
+  Returns:
+    True if the model requires adding a special eos token, False otherwise.
+  """
+  return any(m in model_id for m in _MODELS_REQUIRING_EOS_TOEKN)
+
+
+def load_tokenizer(
+    pretrained_model_id: str,
+    padding_side: str | None = None,
+    access_token: str | None = None,
+) -> transformers.AutoTokenizer:
+  """Loads tokenizer based on `pretrained_model_id`.
+
+  Args:
+    pretrained_model_id: The name of the pretrained model.
+    padding_side: The side to pad the input on.
+    access_token: The access token to use for the tokenizer.
+
+  Returns:
+    The tokenizer.
+  """
+  tokenizer_kwargs = {}
+  if should_add_eos_token(pretrained_model_id):
+    tokenizer_kwargs["add_eos_token"] = True
+  if padding_side:
+    tokenizer_kwargs["padding_side"] = padding_side
+
+  with accelerate.PartialState().local_main_process_first():
+    tokenizer = transformers.AutoTokenizer.from_pretrained(
+        pretrained_model_id,
+        trust_remote_code=False,
+        use_fast=True,
+        token=access_token,
+        **tokenizer_kwargs,
+    )
+
+  if should_add_pad_token(pretrained_model_id):
+    tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+
+  return tokenizer
+
+
+def get_filtered_dataset(
+    dataset: Any,
+    input_column: str,
+    max_seq_length: int,
+    tokenizer: transformers.PreTrainedTokenizer,
+) -> Any:
+  """Returns the dataset by removing examples that are longer than max_seq_length.
+
+  Args:
+    dataset: The dataset to filter.
+    input_column: The input column in the dataset to be used.
+    max_seq_length: The maximum sequence length.
+    tokenizer: The tokenizer.
+  """
+  actual_dataset_length = len(dataset)
+  filtered_dataset = dataset.filter(
+      lambda x: len(tokenizer(x[input_column])["input_ids"]) <= max_seq_length
+  )
+  filtered_dataset_length = len(filtered_dataset)
+  if actual_dataset_length != filtered_dataset_length:
+    examples_removed_percent = (
+        (actual_dataset_length - filtered_dataset_length)
+        * 100
+        / actual_dataset_length
+    )
+    logging.info(
+        "(%.2f%%) of examples token length is <= max-seq-length(%d); (%.2f%%) >"
+        " max-seq-length. Filtering out %d example(s) which are longer than"
+        " max-seq-length.",
+        100 - examples_removed_percent,
+        max_seq_length,
+        examples_removed_percent,
+        actual_dataset_length - filtered_dataset_length,
+    )
+
+  return filtered_dataset
+
+
+def format_dataset(
+    dataset: datasets.Dataset,
+    input_column: str,
+    template: str = None,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> datasets.Dataset:
+  """Takes a raw dataset and formats it using a template and tokenizer.
+
+  Args:
+    dataset: The raw (unprocessed) dataset to format.
+    input_column: The input column in the dataset to be used or updaded by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    A dataset compatible with the template.
+  """
+  return dataset.map(
+      _format_template_fn(
+          template,
+          input_column=input_column,
+          tokenizer=tokenizer,
+      )
+  )
+
+
+def load_dataset_with_template(
+    dataset_name: str,
+    split: str,
+    input_column: str,
+    template: str = None,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> Tuple[Any, Any]:
+  """Loads dataset with templates.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    input_column: The input column in the dataset to be used or updaded by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    The raw dataset and the dataset compatible with the template.
+  """
+  raw = _get_dataset(dataset_name, split=split)
+  if template:
+    templated = format_dataset(raw, input_column, template, tokenizer)
+  else:
+    templated = None
+
+  return raw, templated
+
+
+def validate_dataset_with_template(
+    dataset_name: str,
+    split: str,
+    input_column: str,
+    template: str,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+    max_seq_length: int | None = None,
+    use_multiprocessing: bool = False,
+    validate_percentage_of_dataset: int | None = None,
+    validate_k_rows_of_dataset: int | None = None,
+) -> Any:
+  """Validates dataset with templates.
+
+  This function will be used to load the dataset and validate it against the
+  template. In case of validation, we also allow the users to load the dataset
+  partially by allowing them to read x% or top k rows of the dataset. To
+  validate the dataset, the template file must be available in the GCS bucket
+  and the dataset must be available either in the GCS bucket or Hugging Face.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    input_column: The input column in the dataset to be used or updaded by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    tokenizer: The tokenizer to use for chat_template templates.
+    max_seq_length: The maximum sequence length.
+    use_multiprocessing: If True, it will use multiprocessing to load the
+      dataset.
+    validate_percentage_of_dataset: The percentage of the dataset to load.
+    validate_k_rows_of_dataset: The top k sequences to load from the dataset.
+
+  Returns:
+    None if the validation is successful, otherwise returns the error message.
+  """
+  if not template:
+    raise ValueError("template is required for validate_dataset.")
+
+  if not dataset_name:
+    raise ValueError("dataset_name is empty.")
+
+  if not split:
+    raise ValueError("split is empty.")
+
+  split = _get_split_string(
+      split,
+      validate_percentage_of_dataset,
+      validate_k_rows_of_dataset,
+  )
+
+  num_proc = multiprocessing.cpu_count() if use_multiprocessing else 1
+
+  # gcsfuse cannot be used from the notebook runtime env. Hence, we have
+  # to download dataset and template from gcs to local.
+  if is_gcs_path(dataset_name):
+    dataset_name = download_gcs_uri_to_local(dataset_name, LOCAL_BASE_MODEL_DIR)
+
+  if is_gcs_path(template):
+    template_path = download_gcs_uri_to_local(template, LOCAL_TEMPLATE_DIR)
+  elif os.path.isfile(_github_template_path(template)):
+    template_path = _github_template_path(template)
+  else:
+    raise ValueError(
+        f"Template file {template} does not exist. To validate the"
+        " dataset, please provide a valid GCS path for the template or a valid"
+        " template name from"
+        f" https://github.com/GoogleCloudPlatform/{_VERTEX_AI_SAMPLES_GITHUB_REPO_NAME}/tree/main/{_VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR}."
+    )
+
+  dataset = format_dataset(
+      _get_dataset(dataset_name, split, num_proc),
+      input_column,
+      template_path,
+      tokenizer,
+  )
+
+  if tokenizer is not None:
+    get_filtered_dataset(
+        dataset=dataset,
+        input_column=input_column,
+        max_seq_length=max_seq_length,
+        tokenizer=tokenizer,
+    )
+  print(
+      "Dataset {} is compatible with the {} template.".format(
+          os.path.basename(dataset_name), os.path.basename(template)
+      )
+  )
@@ -1,142 +0,0 @@
-"""Causal language modeling with LoRA models."""
-
-# pylint: disable=g-importing-member
-
-from datasets import load_dataset
-from peft import get_peft_model
-from peft import LoraConfig
-import torch
-from torch import nn
-import transformers
-from transformers import AutoModelForCausalLM
-from transformers import AutoTokenizer
-from transformers import BitsAndBytesConfig
-from transformers import TrainingArguments
-from util import constants
-
-
-def finetune_causal_language_modeling(
-    pretrained_model_id: str,
-    dataset_name: str,
-    output_dir: str,
-    precision_mode: str = None,
-    lora_rank: int = 16,
-    lora_alpha: int = 32,
-    lora_dropout: float = 0.05,
-    warmup_steps: int = 10,
-    max_steps: int = 10,
-    learning_rate: float = 2e-4,
-    local_pretrained_model_id: str = None,
-) -> None:
-  """Finetunes causal language modelings."""
-  if precision_mode == constants.PRECISION_MODE_32:
-    model = AutoModelForCausalLM.from_pretrained(
-        local_pretrained_model_id
-        if local_pretrained_model_id
-        else pretrained_model_id,
-        torch_dtype=torch.float32,
-        device_map="auto",
-    )
-  elif precision_mode == constants.PRECISION_MODE_16:
-    model = AutoModelForCausalLM.from_pretrained(
-        local_pretrained_model_id
-        if local_pretrained_model_id
-        else pretrained_model_id,
-        torch_dtype=torch.bfloat16,
-        device_map="auto",
-    )
-  elif precision_mode == constants.PRECISION_MODE_8:
-    quantization_config = BitsAndBytesConfig(
-        load_in_8bit=True, int8_threshold=0
-    )
-    model = AutoModelForCausalLM.from_pretrained(
-        local_pretrained_model_id
-        if local_pretrained_model_id
-        else pretrained_model_id,
-        torch_dtype=torch.float16,
-        device_map="auto",
-        quantization_config=quantization_config,
-    )
-  else:
-    quantization_config = BitsAndBytesConfig(
-        load_in_4bit=True,
-        bnb_4bit_quant_type="nf4",
-        bnb_4bit_compute_dtype=torch.bfloat16,
-    )
-    model = AutoModelForCausalLM.from_pretrained(
-        local_pretrained_model_id
-        if local_pretrained_model_id
-        else pretrained_model_id,
-        device_map="auto",
-        torch_dtype=torch.bfloat16,
-        quantization_config=quantization_config,
-    )
-
-  tokenizer = AutoTokenizer.from_pretrained(
-      local_pretrained_model_id
-      if local_pretrained_model_id
-      else pretrained_model_id
-  )
-  if "llama" in pretrained_model_id:
-    tokenizer.pad_token = "[PAD]"
-
-  for param in model.parameters():
-    # Freezes the model - train adapters later.
-    param.requires_grad = False
-    if param.ndim == 1:
-      # Casts the small parameters (e.g. layernorm) to fp32 for stability.
-      param.data = param.data.to(torch.float32)
-
-  # Reduces the number of stored activations.
-  model.gradient_checkpointing_enable()
-  model.enable_input_require_grads()
-
-  class CastOutputToFloat(nn.Sequential):
-
-    def forward(self, x):
-      return super().forward(x).to(torch.float32)
-
-  model.lm_head = CastOutputToFloat(model.lm_head)
-
-  config = LoraConfig(
-      r=lora_rank,
-      lora_alpha=lora_alpha,
-      target_modules=["q_proj", "v_proj"],
-      lora_dropout=lora_dropout,
-      bias="none",
-      task_type="CAUSAL_LM",
-  )
-
-  model = get_peft_model(model, config)
-  model.print_trainable_parameters()
-
-  data = load_dataset(dataset_name)
-  data = data.map(
-      lambda samples: tokenizer(samples["quote"]),
-      batched=True,
-  )
-
-  trainer = transformers.Trainer(
-      model=model,
-      train_dataset=data["train"],
-      args=TrainingArguments(
-          per_device_train_batch_size=4,
-          gradient_accumulation_steps=4,
-          warmup_steps=warmup_steps,
-          max_steps=max_steps,
-          learning_rate=learning_rate,
-          fp16=True,
-          logging_steps=1,
-          output_dir=output_dir,
-          ddp_find_unused_parameters=False,
-      ),
-      data_collator=transformers.DataCollatorForLanguageModeling(
-          tokenizer,
-          mlm=False,
-      ),
-  )
-  # Silence the warnings. Please re-enable for inference!
-  model.config.use_cache = False
-  trainer.train()
-
-  model.save_pretrained(output_dir)
@@ -0,0 +1,28 @@
+# Base on pytorch-cuda image.
+FROM pytorch/pytorch:2.0.0-cuda11.7-cudnn8-devel
+
+# Install tools.
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update
+RUN apt-get install -y --no-install-recommends apt-utils
+RUN apt-get install -y --no-install-recommends curl
+RUN apt-get install -y --no-install-recommends wget
+RUN apt-get install -y --no-install-recommends git
+
+# Install libraries.
+ENV PIP_ROOT_USER_ACTION=ignore
+RUN python3 -m pip install --upgrade pip
+RUN pip install tokenizers==0.13.3
+RUN pip install accelerate==0.21.0
+RUN pip install sentencepiece==0.1.99
+RUN pip install datasets==2.14.4
+RUN pip install protobuf==4.24.1
+
+# Install transformers
+RUN git clone https://github.com/huggingface/transformers.git
+WORKDIR transformers
+# Pin the commit to add-code-llama 08/25/2023
+RUN git reset --hard 015f8e110d270a0ad42de4ae5b98198d69eb1964
+RUN pip install -e .
+
+ENTRYPOINT ["python","src/transformers/models/llama/convert_llama_weights_to_hf.py"]
@@ -0,0 +1,22 @@
+# Dockerfile for Language Model Conversion.
+#
+# To build:
+# docker build -f model_oss/peft/dockerfile/conversion.Dockerfile . -t ${YOUR_IMAGE_TAG}
+#
+# To push to gcr:
+# docker tag ${YOUR_IMAGE_TAG} gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+# docker push gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+
+FROM tensorflow/build:2.14-python3.8
+
+RUN git clone https://github.com/facebookresearch/llama-recipes.git && \
+    cd llama-recipes && \
+    pip install -r requirements.txt && \
+    pip freeze | grep transformers && \
+    git clone https://github.com/huggingface/transformers.git && \
+    cd transformers && \
+    pip install protobuf
+
+WORKDIR /llama-recipes/transformers
+
+ENTRYPOINT ["python","src/transformers/models/llama/convert_llama_weights_to_hf.py"]
@@ -7,39 +7,40 @@
 # docker tag ${YOUR_IMAGE_TAG} gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
 # docker push gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}

-FROM pytorch/torchserve:0.7.0-gpu
+FROM pytorch/torchserve:0.11.0-gpu

 USER root

-ENV infer_port=7080
-ENV mng_port=7081
-ENV model_name="peft_serving"
+ENV INFER_PORT=7080
+ENV MNG_PORT=7081
+ENV MODEL="peft_serving"
 ENV PATH="/home/model-server/:${PATH}"

-RUN apt-get update && apt-get install -y --no-install-recommends \
+RUN apt-get update && apt-get -y upgrade && apt-get install -y --no-install-recommends \
        curl \
        wget \
        vim \
        git \
        git-lfs
 RUN git lfs install
+RUN apt-get autoremove -y

 # Install libraries.
 ENV PIP_ROOT_USER_ACTION=ignore
 RUN python3 -m pip install --upgrade pip
-RUN pip install --upgrade torch==2.0.1
+RUN pip install --upgrade torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118
 RUN pip install torchvision==0.15.2
 RUN pip install tokenizers==0.13.3
 RUN pip install accelerate==0.21.0
 RUN pip install sentencepiece==0.1.99
 RUN pip install grpcio-status==1.33.2
 RUN pip install protobuf==3.19.6
-RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/peft.git
+RUN pip install peft==0.5.0
 RUN pip install datasets==2.14.4
-RUN pip install triton==2.0.0.dev20221120
+RUN pip install triton==3.0.0
 RUN pip install xformers==0.0.20
-RUN pip install google-cloud-storage==2.7.0
-RUN pip install absl-py==1.4.0
+RUN pip install google-cloud-storage
+RUN pip install absl-py
 RUN pip install scipy==1.10.1
 RUN pip install evaluate==0.4.0
 RUN pip install scikit-learn==1.2.2
@@ -47,52 +48,43 @@ RUN pip install loralib==0.1.1
 RUN pip install bitsandbytes==0.39.0
 RUN pip install trl==0.4.4
 RUN pip install einops==0.6.1
-
-# Install diffusers from source.
-RUN git clone --depth 1 --branch v0.16.1 https://github.com/huggingface/diffusers.git
-WORKDIR diffusers
-RUN pip install -e .
-WORKDIR /home/model-server
-
-# Install transformers from source.
-RUN git clone --depth 1 --branch v4.31.0 https://github.com/huggingface/transformers.git
-# The patch is used to change the transformers loading model behavior:
-# 1) For models on Huggingface hub: if the model has multiple shards, each shard
-#    will be downloaded separately and get deleted after loading to GPU.
-# 2) For models on local disk: if a model bin file is actually a text file
-#    recording a GCS path, the model file will be downloaded and get deleted
-#    after loading to GPU.
-COPY model_oss/peft/hf_transformers_lazy_download.patch /home/model-server/hf_transformers_lazy_download.patch
-WORKDIR transformers
-RUN git apply /home/model-server/hf_transformers_lazy_download.patch
-RUN pip install -e .
-WORKDIR /home/model-server
+RUN pip install optimum==1.13.2
+RUN pip install auto-gptq==0.4.2
+RUN pip install https://github.com/casper-hansen/AutoAWQ/releases/download/v0.1.7/autoawq-0.1.7+cu118-cp39-cp39-linux_x86_64.whl
+RUN pip install diffusers==0.27.2
+RUN pip install tiktoken==0.6.0
+RUn pip install git+https://github.com/huggingface/transformers.git@76fa17c1663a0efeca7208c20579833365584889
+RUN pip install pynvml==11.4.0
+RUN pip install -i https://test.pypi.org/simple/ bitsandbytes

 # Copy license.
+WORKDIR /home/model-server
 RUN wget https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/main/LICENSE

 # Copy model artifacts.
 COPY model_oss/peft/handler.py /home/model-server/handler.py
 COPY model_oss/peft/config.properties /home/model-server/config.properties
 COPY model_oss/util/ /home/model-server/util/
+COPY model_oss/util/pytorch_startup_prober.sh /model_garden/scripts/pytorch_startup_prober.sh
 ENV PYTHONPATH /home/model-server/

 # Expose ports.
-EXPOSE ${infer_port}
-EXPOSE ${mng_port}
+EXPOSE ${INFER_PORT}
+EXPOSE ${MNG_PORT}

 # Set environments.
 ENV TASK "causal-language-modeling-lora"
-ENV MODEL_ID "openlm-research/open_llama_7b"
+ENV BASE_MODEL_ID ""
+ENV MODEL_ID ""
 ENV PRECISION_LOADING_MODE "float16"
 ENV FINETUNED_LORA_MODEL_PATH ""
-
+ENV TRUST_REMOTE_CODE ""

 # Archive model artifacts and dependencies.
 # Do not set --model-file and --serialized-file because model and checkpoint
 # will be dynamically loaded in handler.py.
 RUN torch-model-archiver \
--model-name=${model_name} \
+--model-name=${MODEL} \
 --version=1.0 \
 --handler=/home/model-server/handler.py \
 --runtime=python3 \
@@ -103,5 +95,5 @@ RUN torch-model-archiver \
 # Run Torchserve HTTP serve to respond to prediction requests.
 CMD ["torchserve", "--start", \
     "--ts-config", "/home/model-server/config.properties", \
-     "--models", "${model_name}=${model_name}.mar", \
+     "--models", "${MODEL}=${MODEL}.mar", \
     "--model-store", "/home/model-server/model-store"]
@@ -1,111 +0,0 @@
-# Dockerfile for PEFT Training.
-#
-# To build:
-# docker build -f model_oss/peft/dockerfile/train.Dockerfile . -t ${YOUR_IMAGE_TAG}
-#
-# To push to gcr:
-# docker tag ${YOUR_IMAGE_TAG} gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
-# docker push gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
-
-# Builds GPU docker image of PyTorch
-# Uses multi-staged approach to reduce size
-# Stage 1
-# Use base conda image to reduce time
-FROM continuumio/miniconda3:latest AS compile-image
-# Specify py version
-ENV PYTHON_VERSION=3.8
-# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
-RUN apt-get update && \
-    apt-get install -y curl git wget software-properties-common git-lfs && \
-    apt-get clean && \
-    rm -rf /var/lib/apt/lists*
-
-# Install audio-related libraries
-RUN apt-get update && \
-    apt install -y ffmpeg
-
-RUN apt install -y libsndfile1-dev
-RUN git lfs install
-
-# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
-RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
-RUN python3 -m pip install --no-cache-dir --upgrade pip
-
-# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
-# We don't install pytorch here yet since CUDA isn't available
-# instead we use the direct torch wheel
-ENV PATH /opt/conda/envs/peft/bin:$PATH
-# Activate our bash shell
-RUN chsh -s /bin/bash
-SHELL ["/bin/bash", "-c"]
-# Activate the conda env and install transformers + accelerate from source
-RUN source activate peft
-RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/transformers
-RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate
-RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/peft#egg=peft[test]
-RUN python3 -m pip install --no-cache-dir bitsandbytes
-
-# Stage 2
-FROM nvidia/cuda:11.2.2-cudnn8-devel-ubuntu20.04 AS build-image
-COPY --from=compile-image /opt/conda /opt/conda
-ENV PATH /opt/conda/bin:$PATH
-
-# Install apt libs
-RUN apt-get update && \
-    apt-get install -y curl git wget vim && \
-    apt-get clean && \
-    rm -rf /var/lib/apt/lists*
-
-# Copy license.
-RUN wget https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/main/LICENSE
-
-RUN echo "source activate peft" >> ~/.profile
-
-# Install libraries.
-RUN pip install --upgrade torch==2.0.1
-RUN pip install torchvision==0.15.2
-RUN pip install git+https://github.com/huggingface/transformers@de9255de27abfcae4a1f816b904915f0b1e23cd9
-RUN pip install transformers -U
-RUN pip install accelerate==0.21.0
-RUN pip install sentencepiece==0.1.99
-RUN pip install grpcio-status==1.33.2
-RUN pip install protobuf==3.19.6
-RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/peft.git
-RUN pip install datasets==2.9.0
-RUN pip install triton==2.0.0.dev20221120
-RUN pip install xformers==0.0.20
-RUN pip install Jinja2==3.1.2
-RUN pip install ftfy==6.1.1
-RUN pip install cloudml-hypertune==0.1.0.dev6
-RUN pip install tensorboard==2.12.0
-RUN pip install scipy==1.10.1
-RUN pip install evaluate==0.4.0
-RUN pip install scikit-learn==1.2.2
-RUN pip install loralib==0.1.1
-RUN pip install bitsandbytes==0.39.0
-RUN pip install trl==0.4.4
-RUN pip install einops==0.6.1
-RUN pip install google-cloud-storage==2.7.0
-
-RUN git clone --depth 1 --branch v0.16.1 https://github.com/huggingface/diffusers.git
-WORKDIR diffusers
-RUN pip install -e .
-
-# Switch to diffusers examples folder.
-WORKDIR examples
-
-# NOTE: use 'sed' to modify train_text_to_image_lora.py to
-#       fix the bug for accelerator.
-RUN sed -i \
-"s#logging_dir=logging_dir#project_dir=logging_dir#g" \
-text_to_image/train_text_to_image_lora.py
-
-# Config accelerate.
-RUN mkdir -p ./vertex_vision_model_garden_peft/
-COPY model_oss/peft/train.sh ./vertex_vision_model_garden_peft/train.sh
-COPY model_oss/peft/*.py     ./vertex_vision_model_garden_peft/
-COPY model_oss/util /diffusers/examples/util
-ENV PYTHONPATH /diffusers/examples/
-
-# Generate accelerate config at the beginning of docker run.
-ENTRYPOINT ["python3", "vertex_vision_model_garden_peft/main.py"]
@@ -72,15 +72,39 @@ class PeftHandler(BaseHandler):
        "PRECISION_LOADING_MODE", constants.PRECISION_MODE_16
    )
    self.task = os.environ.get("TASK", CAUSAL_LANGUAGE_MODELING_LORA)
-    self.base_model_id = os.environ.get("BASE_MODEL_ID", None)
-    self.model_id = self.base_model_id
-    if not self.base_model_id:
-      self.model_id = os.environ.get("MODEL_ID", "")
+    trust_remote_code = os.environ.get("TRUST_REMOTE_CODE", None)
+    if trust_remote_code == "false":
+      self.trust_remote_code = False
+    else:
+      self.trust_remote_code = True
+
+    # If present, the path of the model in the container.
+    aip_storage_dir = os.environ.get("AIP_STORAGE_DIR", None)
+
+    # If present, the URI of the model in a google owned GCS bucket.
+    aip_storage_uri = os.environ.get("AIP_STORAGE_URI", None)
+
+    model_id = os.environ.get("MODEL_ID", None)
+    base_model_id = os.environ.get("BASE_MODEL_ID", None)
+
+    self.model_id = None
+    if aip_storage_dir:
+      self.model_id = aip_storage_dir
+      logging.info(f"Loaded base model from AIP_STORAGE_DIR: {self.model_id}.")
+    elif aip_storage_uri:
+      self.model_id = aip_storage_uri
+      logging.info(f"Loaded base model from AIP_STORAGE_URI: {self.model_id}.")
+    elif model_id:
+      self.model_id = model_id
+      logging.info(f"Loaded base model from MODEL_ID: {self.model_id}.")
+    elif base_model_id:
+      # Note: BASE_MODEL_ID has been unified with MODEL_ID.
+      # MODEL_ID should be used whenever possible.
+      self.model_id = base_model_id
+      logging.info(f"Loaded base model from BASE_MODEL_ID: {self.model_id}.")
+
    self.quantization = os.environ.get("QUANTIZATION", None)
-    logging.info(f"Load base model id from MODEL_ID:{self.model_id}.")
-    if not self.model_id:
-      self.model_id = os.environ.get("AIP_STORAGE_URI", "")
-      logging.info(f"Load base model id from AIP_STORAGE_URI: {self.model_id}.")
+
    if not self.model_id:
      raise ValueError("Base model id is must be set.")
    if fileutils.is_gcs_path(self.model_id):
@@ -101,8 +125,7 @@ class PeftHandler(BaseHandler):

    logging.info(
        f"Using task:{self.task}, base model:{self.model_id}, lora model:"
-        f" {self.finetuned_lora_model_path}, and precision"
-        f" {self.precision_mode}."
+        f" {self.finetuned_lora_model_path}, precision {self.precision_mode}."
    )

    self.pipeline = None
@@ -145,11 +168,18 @@ class PeftHandler(BaseHandler):
    elif (
        self.task == CAUSAL_LANGUAGE_MODELING_LORA or self.task == INSTRUCT_LORA
    ):
-      tokenizer = AutoTokenizer.from_pretrained(self.model_id)
+      tokenizer = AutoTokenizer.from_pretrained(
+          self.model_id,
+          trust_remote_code=self.trust_remote_code,
+      )
+
      logging.debug("Initialized the tokenizer.")
      if self.task == CAUSAL_LANGUAGE_MODELING_LORA:
        if self.quantization == constants.AWQ:
-          model = AutoAWQForCausalLM.from_quantized(self.model_id)
+          model = AutoAWQForCausalLM.from_quantized(
+              self.model_id,
+              trust_remote_code=self.trust_remote_code,
+          )
        elif self.quantization == constants.GPTQ or not self.quantization:
          if self.precision_mode == constants.PRECISION_MODE_32:
            model = AutoModelForCausalLM.from_pretrained(
@@ -157,6 +187,7 @@ class PeftHandler(BaseHandler):
                return_dict=True,
                torch_dtype=torch.float32,
                device_map="auto",
+                trust_remote_code=self.trust_remote_code,
            )
          elif self.precision_mode == constants.PRECISION_MODE_16B:
            model = AutoModelForCausalLM.from_pretrained(
@@ -164,6 +195,7 @@ class PeftHandler(BaseHandler):
                return_dict=True,
                torch_dtype=torch.bfloat16,
                device_map="auto",
+                trust_remote_code=self.trust_remote_code,
            )
          elif self.precision_mode == constants.PRECISION_MODE_16:
            model = AutoModelForCausalLM.from_pretrained(
@@ -171,6 +203,7 @@ class PeftHandler(BaseHandler):
                return_dict=True,
                torch_dtype=torch.float16,
                device_map="auto",
+                trust_remote_code=self.trust_remote_code,
            )
          elif self.precision_mode == constants.PRECISION_MODE_8:
            quantization_config = BitsAndBytesConfig(
@@ -182,6 +215,7 @@ class PeftHandler(BaseHandler):
                torch_dtype=torch.float16,
                device_map="auto",
                quantization_config=quantization_config,
+                trust_remote_code=self.trust_remote_code,
            )
          else:
            quantization_config = BitsAndBytesConfig(
@@ -195,6 +229,7 @@ class PeftHandler(BaseHandler):
                device_map="auto",
                torch_dtype=torch.bfloat16,
                quantization_config=quantization_config,
+                trust_remote_code=self.trust_remote_code,
            )
        else:
          raise ValueError(f"Invalid QUANTIZATION value: {self.quantization}")
@@ -203,14 +238,14 @@ class PeftHandler(BaseHandler):
          model = AutoModelForCausalLM.from_pretrained(
              self.model_id,
              torch_dtype=torch.bfloat16,
-              trust_remote_code=True,
+              trust_remote_code=self.trust_remote_code,
              device_map="auto",
          )
        except:  # pylint: disable=bare-except
          model = AutoModelForCausalLM.from_pretrained(
              self.model_id,
              torch_dtype=torch.bfloat16,
-              trust_remote_code=True,
+              trust_remote_code=self.trust_remote_code,
              device_map="auto",
          )
      logging.debug("Initialized the base model.")
@@ -329,4 +364,4 @@ class PeftHandler(BaseHandler):
    return f"Prompt:\n{prompt.strip()}\nOutput:\n{output}"


-# pylint: enable=logging-fstring-interpolation
+# pylint: enable=logging-fstring-interpolation
@@ -1,131 +0,0 @@
-diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
-index 45459ed..32527f4 100644
--- a/src/transformers/modeling_utils.py
-+++ b/src/transformers/modeling_utils.py
-@@ -32,6 +32,8 @@ import torch
- from packaging import version
- from torch import Tensor, nn
- from torch.nn import CrossEntropyLoss
-+from huggingface_hub import hf_hub_download
-+from google.cloud import storage
-
- from .activations import get_activation
- from .configuration_utils import PretrainedConfig
-@@ -442,6 +444,29 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
-     """
-     Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
-     """
-+    delete_download = False
-+    tmp_dir = "/tmp/model"
-+    os.makedirs(tmp_dir, exist_ok=True)
-+    if isinstance(checkpoint_file, dict):
-+        # Download model file from huggingface
-+        print(f"==> Download model from HF: {checkpoint_file}")
-+        checkpoint_file = hf_hub_download(
-+            local_dir=tmp_dir, local_dir_use_symlinks=False, force_download=True, resume_download=True, **checkpoint_file)
-+        delete_download = True
-+    else:
-+        with open(checkpoint_file, "rb") as f:
-+            is_gcs_file = (f.read(2) == b"gs")
-+        if is_gcs_file:
-+            # Download model file from GCS
-+            with open(checkpoint_file, "r") as f:
-+                gcs_file = f.read()
-+            checkpoint_file = os.path.join(tmp_dir, gcs_file.split("/")[-1])
-+            print(f"==> Download model from GCS: {gcs_file} to: {checkpoint_file}")
-+            client = storage.Client()
-+            with open(checkpoint_file, 'wb') as f:
-+                client.download_blob_to_file(gcs_file, f)
-+            delete_download = True
-+
-     if checkpoint_file.endswith(".safetensors") and is_safetensors_available():
-         # Check format of the archive
-         with safe_open(checkpoint_file, framework="pt") as f:
-@@ -455,9 +480,9 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
-             raise NotImplementedError(
-                 f"Conversion from a {metadata['format']} safetensors archive to PyTorch is not implemented yet."
-             )
-        return safe_load_file(checkpoint_file)
-+        state_dict = safe_load_file(checkpoint_file)
-     try:
-        return torch.load(checkpoint_file, map_location="cpu")
-+        state_dict = torch.load(checkpoint_file, map_location="cpu")
-     except Exception as e:
-         try:
-             with open(checkpoint_file) as f:
-@@ -478,6 +503,10 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
-                 f"at '{checkpoint_file}'. "
-                 "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True."
-             )
-+    if delete_download:
-+        print(f"==> Delete downloaded model: {checkpoint_file}")
-+        os.remove(checkpoint_file)
-+    return state_dict
-
-
- def set_initialized_submodules(model, state_dict_keys):
-@@ -3179,7 +3208,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
-             return mismatched_keys
-
-         if resolved_archive_file is not None:
-            folder = os.path.sep.join(resolved_archive_file[0].split(os.path.sep)[:-1])
-+            if isinstance(resolved_archive_file, str):
-+                folder = os.path.sep.join(resolved_archive_file[0].split(os.path.sep)[:-1])
-+            else:
-+                folder = None
-         else:
-             folder = None
-         if device_map is not None and is_safetensors:
-diff --git a/src/transformers/utils/hub.py b/src/transformers/utils/hub.py
-index ffed743..4b15770 100644
--- a/src/transformers/utils/hub.py
-+++ b/src/transformers/utils/hub.py
-@@ -414,20 +414,34 @@ def cached_file(
-     user_agent = http_user_agent(user_agent)
-     try:
-         # Load from URL or cache if already cached
-        resolved_file = hf_hub_download(
-            path_or_repo_id,
-            filename,
-            subfolder=None if len(subfolder) == 0 else subfolder,
-            repo_type=repo_type,
-            revision=revision,
-            cache_dir=cache_dir,
-            user_agent=user_agent,
-            force_download=force_download,
-            proxies=proxies,
-            resume_download=resume_download,
-            use_auth_token=use_auth_token,
-            local_files_only=local_files_only,
-        )
-+        if filename.endswith(".bin"):
-+            # NOTE: To save disk we do not download bin file eagerly. Do not support safetensors.
-+            resolved_file = dict(
-+                repo_id=path_or_repo_id,
-+                filename=filename,
-+                subfolder=None if len(subfolder) == 0 else subfolder,
-+                repo_type=repo_type,
-+                revision=revision,
-+                user_agent=user_agent,
-+                proxies=proxies,
-+                use_auth_token=use_auth_token,
-+            )
-+            print(f"--> Apply lazy download to bin file: {resolved_file}")
-+        else:
-+            resolved_file = hf_hub_download(
-+                path_or_repo_id,
-+                filename,
-+                subfolder=None if len(subfolder) == 0 else subfolder,
-+                repo_type=repo_type,
-+                revision=revision,
-+                cache_dir=cache_dir,
-+                user_agent=user_agent,
-+                force_download=force_download,
-+                proxies=proxies,
-+                resume_download=resume_download,
-+                use_auth_token=use_auth_token,
-+                local_files_only=local_files_only,
-+            )
-
-     except RepositoryNotFoundError:
-         raise EnvironmentError(
@@ -1,97 +0,0 @@
-"""Instruct/Chat with LoRA models."""
-
-# pylint: disable=g-importing-member
-from datasets import load_dataset
-from peft import LoraConfig
-import torch
-from transformers import AutoModelForCausalLM
-from transformers import AutoTokenizer
-from transformers import BitsAndBytesConfig
-from transformers import TrainingArguments
-from trl import SFTTrainer
-
-
-def finetune_instruct(
-    pretrained_model_id: str,
-    dataset_name: str,
-    output_dir: str,
-    lora_rank: int = 64,
-    lora_alpha: int = 16,
-    lora_dropout: float = 0.1,
-    warmup_ratio: int = 0.03,
-    max_steps: int = 10,
-    max_seq_length: int = 512,
-    learning_rate: float = 2e-4,
-) -> None:
-  """Finetunes instruct."""
-  dataset = load_dataset(dataset_name, split="train")
-
-  bnb_config = BitsAndBytesConfig(
-      load_in_4bit=True,
-      bnb_4bit_quant_type="nf4",
-      bnb_4bit_compute_dtype=torch.float16,
-  )
-
-  model = AutoModelForCausalLM.from_pretrained(
-      pretrained_model_id,
-      quantization_config=bnb_config,
-      trust_remote_code=True,
-  )
-  model.config.use_cache = False
-
-  tokenizer = AutoTokenizer.from_pretrained(
-      pretrained_model_id, trust_remote_code=True
-  )
-  tokenizer.pad_token = tokenizer.eos_token
-
-  peft_config = LoraConfig(
-      lora_alpha=lora_alpha,
-      lora_dropout=lora_dropout,
-      r=lora_rank,
-      bias="none",
-      task_type="CAUSAL_LM",
-      target_modules=[
-          "query_key_value",
-          "dense",
-          "dense_h_to_4h",
-          "dense_4h_to_h",
-      ],
-  )
-
-  per_device_train_batch_size = 4
-  gradient_accumulation_steps = 4
-  optim = "paged_adamw_32bit"
-  save_steps = 10
-  logging_steps = 10
-  max_grad_norm = 0.3
-  lr_scheduler_type = "constant"
-
-  training_arguments = TrainingArguments(
-      output_dir=output_dir,
-      per_device_train_batch_size=per_device_train_batch_size,
-      gradient_accumulation_steps=gradient_accumulation_steps,
-      optim=optim,
-      save_steps=save_steps,
-      logging_steps=logging_steps,
-      learning_rate=learning_rate,
-      fp16=True,
-      max_grad_norm=max_grad_norm,
-      max_steps=max_steps,
-      warmup_ratio=warmup_ratio,
-      group_by_length=True,
-      lr_scheduler_type=lr_scheduler_type,
-  )
-
-  trainer = SFTTrainer(
-      model=model,
-      train_dataset=dataset,
-      peft_config=peft_config,
-      dataset_text_field="text",
-      max_seq_length=max_seq_length,
-      tokenizer=tokenizer,
-      args=training_arguments,
-  )
-  for name, module in trainer.model.named_modules():
-    if "norm" in name:
-      module = module.to(torch.float32)
-  trainer.train()
@@ -1,177 +0,0 @@
-"""Main function to start PEFT finetuning."""
-import subprocess
-
-from absl import app
-from absl import flags
-from absl import logging
-
-from peft import causal_language_modeling_lora
-from peft import instruct_lora
-from peft import sequence_classification_lora
-from util import constants
-from util import fileutils
-
-_TASK = flags.DEFINE_string(
-    'task',
-    constants.CAUSAL_LANGUAGE_MODELING_LORA,
-    'The supported PEFT tasks.',
-)
-
-_PRETRAINED_MODEL_ID = flags.DEFINE_string(
-    'pretrained_model_id',
-    None,
-    'The pretrained model id. Supported models can be causal language modeling'
-    ' models from https://github.com/huggingface/peft/tree/main.',
-    required=True,
-)
-
-_DATASET_NAME = flags.DEFINE_string(
-    'dataset_name',
-    None,
-    'The dataset name in huggingface.',
-    required=True,
-)
-
-_OUTPUT_DIR = flags.DEFINE_string(
-    'output_dir',
-    None,
-    'The output directory.',
-    required=True,
-)
-
-_PRECISION_MODE = flags.DEFINE_string(
-    'precision_mode',
-    constants.PRECISION_MODE_16,
-    'Supported finetuning precision_modes are `{}` and `{}`.'.format(
-        constants.PRECISION_MODE_8, constants.PRECISION_MODE_16
-    ),
-)
-
-_LORA_RANK = flags.DEFINE_integer(
-    'lora_rank',
-    16,
-    'The rank of the update matrices, expressed in int. Lower rank results in'
-    ' smaller update matrices with fewer trainable parameters, referring to'
-    ' https://huggingface.co/docs/peft/conceptual_guides/lora.',
-)
-
-_LORA_ALPHA = flags.DEFINE_integer(
-    'lora_alpha',
-    32,
-    'LoRA scaling factor, referring to'
-    ' https://huggingface.co/docs/peft/conceptual_guides/lora.',
-)
-
-_LORA_DROPOUT = flags.DEFINE_float(
-    'lora_dropout',
-    0.05,
-    'dropout probability of the LoRA layers, referring to'
-    ' https://huggingface.co/docs/peft/task_guides/token-classification-lora.',
-)
-
-_WARMUP_STEPS = flags.DEFINE_integer(
-    'warmup_steps',
-    10,
-    'Number of steps for the warmup in the learning rate scheduler.',
-)
-
-_WARMUP_RATIO = flags.DEFINE_float(
-    'warmup_ratio',
-    0.03,
-    'The warmup ratio in the learning rate scheduler.',
-)
-
-_MAX_STEPS = flags.DEFINE_integer(
-    'max_steps',
-    10,
-    'Total number of training steps.',
-)
-
-_MAX_SEQ_LENGTH = flags.DEFINE_integer(
-    'max_seq_length',
-    512,
-    'The maximum sequence length.',
-)
-
-_NUM_EPOCHS = flags.DEFINE_integer(
-    'num_epochs',
-    20,
-    'The number of training epochs.',
-)
-
-_BATCH_SIZE = flags.DEFINE_integer(
-    'batch_size',
-    32,
-    'The batch size.',
-)
-
-_LEARNING_RATE = flags.DEFINE_float(
-    'learning_rate',
-    2e-4,
-    'The learning rate after the potential warmup period.',
-)
-
-
-def main(_) -> None:
-  task = _TASK.value
-  pretrained_model_id = _PRETRAINED_MODEL_ID.value
-  local_pretrained_model_id = None
-  if pretrained_model_id.startswith(constants.GCS_URI_PREFIX):
-    logging.info(
-        'Start to copy pretrained models locally: %s.', pretrained_model_id
-    )
-    fileutils.download_gcs_dir_to_local(
-        pretrained_model_id, constants.LOCAL_BASE_MODEL_DIR
-    )
-    local_pretrained_model_id = constants.LOCAL_BASE_MODEL_DIR
-    logging.info(
-        'Finished copying pretrained models locally to: %s.',
-        local_pretrained_model_id,
-    )
-  if task == constants.TEXT_TO_IMAGE_LORA:
-    subprocess.run(['/bin/bash', 'train.sh'], check=True)
-  elif task == constants.SEQUENCE_CLASSIFICATION_LORA:
-    sequence_classification_lora.finetune_sequence_classification(
-        pretrained_model_id=pretrained_model_id,
-        dataset_name=_DATASET_NAME.value,
-        output_dir=_OUTPUT_DIR.value,
-        lora_rank=_LORA_RANK.value,
-        lora_alpha=_LORA_ALPHA.value,
-        lora_dropout=_LORA_DROPOUT.value,
-        num_epochs=_NUM_EPOCHS.value,
-        batch_size=_BATCH_SIZE.value,
-        learning_rate=_LEARNING_RATE.value,
-    )
-  elif task == constants.CAUSAL_LANGUAGE_MODELING_LORA:
-    causal_language_modeling_lora.finetune_causal_language_modeling(
-        pretrained_model_id=pretrained_model_id,
-        dataset_name=_DATASET_NAME.value,
-        output_dir=_OUTPUT_DIR.value,
-        precision_mode=_PRECISION_MODE.value,
-        lora_rank=_LORA_RANK.value,
-        lora_alpha=_LORA_ALPHA.value,
-        lora_dropout=_LORA_DROPOUT.value,
-        warmup_steps=_WARMUP_STEPS.value,
-        max_steps=_MAX_STEPS.value,
-        learning_rate=_LEARNING_RATE.value,
-        local_pretrained_model_id=local_pretrained_model_id,
-    )
-  elif task == constants.INSTRUCT_LORA:
-    instruct_lora.finetune_instruct(
-        pretrained_model_id=pretrained_model_id,
-        dataset_name=_DATASET_NAME.value,
-        output_dir=_OUTPUT_DIR.value,
-        lora_rank=_LORA_RANK.value,
-        lora_alpha=_LORA_ALPHA.value,
-        lora_dropout=_LORA_DROPOUT.value,
-        warmup_ratio=_WARMUP_RATIO.value,
-        max_steps=_MAX_STEPS.value,
-        max_seq_length=_MAX_SEQ_LENGTH.value,
-        learning_rate=_LEARNING_RATE.value,
-    )
-  else:
-    raise ValueError('The task {} is not supported.'.format(task))
-
-
-if __name__ == '__main__':
-  app.run(main)
@@ -1,133 +0,0 @@
-"""Sequence classification with LoRA models."""
-
-# pylint: disable=g-importing-member
-
-from datasets import load_dataset
-import evaluate
-from peft import get_peft_model
-from peft import LoraConfig
-import torch
-from torch.optim import AdamW
-from torch.utils.data import DataLoader
-from tqdm import tqdm
-from transformers import AutoModelForSequenceClassification
-from transformers import AutoTokenizer
-from transformers import get_linear_schedule_with_warmup
-
-
-def finetune_sequence_classification(
-    pretrained_model_id: str,
-    dataset_name: str,
-    output_dir: str,
-    lora_rank: int = 8,
-    lora_alpha: int = 16,
-    lora_dropout: float = 0.1,
-    num_epochs: int = 20,
-    batch_size: int = 32,
-    learning_rate: float = 3e-4,
-) -> None:
-  """Finetunes sequence classification."""
-  task = "mrpc"
-  device = "cuda"
-
-  peft_config = LoraConfig(
-      task_type="SEQ_CLS",
-      inference_mode=False,
-      r=lora_rank,
-      lora_alpha=lora_alpha,
-      lora_dropout=lora_dropout,
-  )
-  if any(k in pretrained_model_id for k in ("gpt", "opt", "bloom")):
-    padding_side = "left"
-  else:
-    padding_side = "right"
-
-  tokenizer = AutoTokenizer.from_pretrained(
-      pretrained_model_id, padding_side=padding_side
-  )
-  if getattr(tokenizer, "pad_token_id") is None:
-    tokenizer.pad_token_id = tokenizer.eos_token_id
-
-  datasets = load_dataset(dataset_name, task)
-  metric = evaluate.load(dataset_name, task)
-
-  def tokenize_function(examples):
-    # max_length=None => use the model max length (it's actually the default)
-    outputs = tokenizer(
-        examples["sentence1"],
-        examples["sentence2"],
-        truncation=True,
-        max_length=None,
-    )
-    return outputs
-
-  tokenized_datasets = datasets.map(
-      tokenize_function,
-      batched=True,
-      remove_columns=["idx", "sentence1", "sentence2"],
-  )
-
-  # We also rename the 'label' column to 'labels' which is the expected name for
-  # labels by the models of the transformers library.
-  tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
-
-  def collate_fn(examples):
-    return tokenizer.pad(examples, padding="longest", return_tensors="pt")
-
-  # Instantiate dataloaders.
-  train_dataloader = DataLoader(
-      tokenized_datasets["train"],
-      shuffle=True,
-      collate_fn=collate_fn,
-      batch_size=batch_size,
-  )
-  eval_dataloader = DataLoader(
-      tokenized_datasets["validation"],
-      shuffle=False,
-      collate_fn=collate_fn,
-      batch_size=batch_size,
-  )
-
-  model = AutoModelForSequenceClassification.from_pretrained(
-      pretrained_model_id, return_dict=True
-  )
-  model = get_peft_model(model, peft_config)
-  model.print_trainable_parameters()
-
-  optimizer = AdamW(params=model.parameters(), lr=learning_rate)
-
-  # Instantiate scheduler
-  lr_scheduler = get_linear_schedule_with_warmup(
-      optimizer=optimizer,
-      num_warmup_steps=0.06 * (len(train_dataloader) * num_epochs),
-      num_training_steps=(len(train_dataloader) * num_epochs),
-  )
-
-  model.to(device)
-  for epoch in range(num_epochs):
-    model.train()
-    for _, batch in enumerate(tqdm(train_dataloader)):
-      batch.to(device)
-      outputs = model(**batch)
-      loss = outputs.loss
-      loss.backward()
-      optimizer.step()
-      lr_scheduler.step()
-      optimizer.zero_grad()
-
-    model.eval()
-    for _, batch in enumerate(tqdm(eval_dataloader)):
-      batch.to(device)
-      with torch.no_grad():
-        outputs = model(**batch)
-      predictions = outputs.logits.argmax(dim=-1)
-      references = batch["labels"]
-      metric.add_batch(
-          predictions=predictions,
-          references=references,
-      )
-
-    eval_metric = metric.compute()
-    print(f"epoch {epoch}:", eval_metric)
-
-  model.save_pretrained(output_dir)
@@ -1,6 +0,0 @@
-#!/bin/bash
-
-# Setup accelerate config before running trainer.
-python -c "from accelerate.utils import write_basic_config; write_basic_config(mixed_precision='fp16')"
-
-accelerate launch "$@"
@@ -0,0 +1,16 @@
+# Dockerfile for axolotl training.
+#
+# To build:
+# docker build -f model_oss/peft/train/axolotol/dockerfile/train.Dockerfile . -t ${YOUR_IMAGE_TAG}
+#
+# To push to gcr:
+# docker tag ${YOUR_IMAGE_TAG} gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+# docker push gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+
+FROM winglian/axolotl:main-latest
+
+RUN mkdir -p ./vertex_vision_model_garden/
+
+COPY model_oss/peft/train/axolotl/*.py ./vertex_vision_model_garden/
+
+ENTRYPOINT ["python3", "./vertex_vision_model_garden/train_entrypoint.py"]
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+# Run copybara first:
+# cloud/ml/applications/vision/model_garden/copybara/run_copybara_local.sh
+# Run docker build:
+# cloud/ml/applications/vision/model_garden/model_oss/peft/train/axolotl/scripts/build_train_docker.sh
+
+set -x
+
+COPYBARA_DIR="/tmp/train_docker/"
+
+pushd "${COPYBARA_DIR}"
+
+PROJECT="cloud-nas-260507"
+IMAGE_TAG="gcr.io/${PROJECT}/axolotl-train:${USER}-test"
+
+docker build -f model_oss/peft/train/axolotl/dockerfile/train.Dockerfile . -t "${IMAGE_TAG}"
+docker push "${IMAGE_TAG}"
+
+popd
@@ -0,0 +1,88 @@
+"""Entrypoint for axolotl train docker."""
+
+import argparse
+import json
+import os
+import subprocess
+
+
+def _get_multi_node_flags(cluster_spec: str) -> list[str]:
+  """Returns the multi-node flags."""
+  print(f'CLUSTER_SPEC: {cluster_spec}')
+
+  cluster_data = json.loads(cluster_spec)
+
+  # Get primary node info
+  primary_node = cluster_data['cluster']['workerpool0'][0]
+  print(f'primary node: {primary_node}')
+  primary_node_addr, primary_node_port = primary_node.split(':')
+  print(f'primary node address: {primary_node_addr}')
+  print(f'primary node port: {primary_node_port}')
+
+  # Determine node rank of this machine
+  workerpool = cluster_data['task']['type']
+  if workerpool == 'workerpool0':
+    node_rank = 0
+  else:
+    node_rank = cluster_data['task']['index'] + 1
+  print(f'node rank: {node_rank}')
+
+  # Calculate total nodes
+  num_worker_nodes = len(cluster_data['cluster']['workerpool1'])
+  num_nodes = num_worker_nodes + 1  # Add 1 for the primary node
+  print(f'num nodes: {num_nodes}')
+
+  return [
+      f'--machine_rank={node_rank}',
+      f'--num_machines={num_nodes}',
+      f'--main_process_ip={primary_node_addr}',
+      f'--main_process_port={primary_node_port}',
+      '--max_restarts=0',
+      '--monitor_interval=120',
+      '--dynamo_backend=no',
+  ]
+
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--config_file')
+  parser.add_argument('--huggingface_access_token')
+  args, unknown = parser.parse_known_args()
+
+  accelerate_flags = []
+
+  if args.config_file:
+    accelerate_flags.append(f'--config_file={args.config_file}')
+
+  if cluster_spec := os.getenv('CLUSTER_SPEC', default=None):
+    print('========== Launch on cloud multi nodes ==========')
+    accelerate_flags.extend(_get_multi_node_flags(cluster_spec))
+
+  cmd = (
+      [
+          'accelerate',
+          'launch',
+      ]
+      + accelerate_flags
+      + [
+          '-m',
+          'axolotl.cli.train',
+      ]
+      + unknown
+  )
+  print(f'{cmd=}', flush=True)
+
+  env = os.environ.copy()
+
+  if args.huggingface_access_token:
+    env['HF_TOKEN'] = args.huggingface_access_token
+
+  subprocess.run(
+      cmd,
+      check=True,
+      env=env,
+  )
+
+
+if __name__ == '__main__':
+  main()
@@ -0,0 +1,95 @@
+"""Class that bundles docker related flags."""
+
+import getpass
+import os
+import pwd
+
+
+class CommandBuilder:
+  """Base class for building commands."""
+
+  def __init__(self):
+    self._defaults = []
+    self._env_vars = {}
+
+  def add_env_var(self, var: str, val: str) -> None:
+    """Add environment variable to the command.
+
+    Args:
+      var: environment variable name.
+      val: environment variable value.
+    """
+    self._env_vars[var] = val
+
+  def add_mount_map(self, host_path, docker_path):
+    pass
+
+
+class DockerCommandBuilder(CommandBuilder):
+  """Bundle docker related flags."""
+
+  def __init__(self, docker_uri: str, shm_size: str = '128gb'):
+    super().__init__()
+    self._docker_uri = [docker_uri]
+    self.privilege_mode = []
+    self.entrypoint = []
+
+    self._defaults = [
+        'docker',
+        'run',
+        '--gpus=all',
+        '--net=host',
+        '--rm',
+        f'--shm-size={shm_size}',
+    ]
+
+    self._mount_maps = []
+    user = getpass.getuser()
+    # username ends with `_google_com` is managed by ldap and does not have a
+    # corresponding entry in /etc/passwd or /etc/group file. We cannot enable
+    # non-root docker user with below method.
+    if not user.endswith('_google_com'):
+      uid = os.getuid()
+      gid = pwd.getpwuid(uid).pw_gid
+      self._defaults += [
+          f'--user={uid}:{gid}',
+          '--volume=/etc/group:/etc/group:ro',
+          '--volume=/etc/passwd:/etc/passwd:ro',
+      ]
+
+  def add_mount_map(self, host_path, docker_path):
+    self._mount_maps.append(f'--volume={host_path}:{docker_path}')
+
+  def add_privilege_mode(self):
+    self.privilege_mode = ['--privileged']
+
+  def add_entrypoint(self, entrypoint: list[str]):
+    self.entrypoint = entrypoint
+
+  def build_cmd(self) -> str:
+    return (
+        self._defaults
+        + [f'--env={var}={val}' for var, val in self._env_vars.items()]
+        + self._mount_maps
+        + self.privilege_mode
+        + self._docker_uri
+        + self.entrypoint
+    )
+
+
+class PythonCommandBuilder(CommandBuilder):
+  """Bundle Python test command related flags."""
+
+  def __init__(self):
+    super().__init__()
+    self._defaults = [
+        'python3',
+        './vertex_vision_model_garden_peft/train/vmg/train_entrypoint.py',
+    ]
+
+  def build_cmd(self) -> str:
+    os.environ.update(self._env_vars)
+    return self._defaults
+
+  def add_entrypoint(self, entrypoint: list[str]):
+    self._defaults = entrypoint
@@ -0,0 +1,471 @@
+"""Test util class."""
+
+import copy
+import dataclasses
+import datetime
+import inspect
+import os
+import signal
+import subprocess
+import sys
+from absl import flags
+from absl import logging
+from absl.testing import parameterized
+import command_builder
+import immutabledict
+import torch
+
+_DOCKER_URI = flags.DEFINE_string('docker_uri', None, 'docker image uri')
+
+_DRY_RUN = flags.DEFINE_bool('dry_run', False, 'dry-run the commands')
+
+_LOCAL_INPUT_DIR = flags.DEFINE_string(
+    'local_input_dir',
+    os.path.expanduser('~/test_input'),
+    'local directory for storing input data.',
+)
+
+_LOCAL_OUTPUT_DIR = flags.DEFINE_string(
+    'local_output_dir',
+    '/tmp',
+    'local directory for storing test output.',
+)
+
+
+_GCS_INPUT_DIR = flags.DEFINE_string(
+    'gcs_input_dir',
+    'gs://vmg-tuning-docker-test',
+    'GCS directory that stores model checkpoint, dataset and etc.',
+)
+
+_GCS_OUTPUT_DIR = flags.DEFINE_string(
+    'gcs_output_dir',
+    'gs://vmg-tuning-docker-test/output',
+    'GCS directory that stores test output.',
+)
+
+_GCS_TESTDATA_DIR = 'peft-train-image-test'
+
+_THROUGHPUT_TEST_EXCEPTIONS = immutabledict.immutabledict({
+    ('bm_deepspeed_zero3_8gpu_gemma-2-9b-it_4bit.txt', '12.0'): float('inf'),
+    ('bm_fsdp_8gpu_llama3.1-70b-hf_4bit.txt', '20.0'): float('inf'),
+    ('bm_deepspeed_zero2_8gpu_gemma-2-2b-it_bfloat16.txt', '12.0'): 20.0,
+    ('bm_deepspeed_zero3_8gpu_gemma-2-2b-it_4bit.txt', '4.0'): 20.0,
+    ('bm_deepspeed_zero3_8gpu_gemma-2-27b-it_4bit.txt', '4.0'): 20.0,
+})
+
+
+@dataclasses.dataclass
+class BenchmarkStats:
+  """Class to store the benchmark result.
+
+  Attributes:
+    peak_mem: peak memory in GB.
+    throughput: throughput in tokens/sec.
+  """
+
+  peak_mem: float
+  throughput: float
+
+
+class TestBase(parameterized.TestCase):
+  """Test base class that defines how to run commands."""
+
+  def setUp(self):
+    super().setUp()
+
+    # Create a copy of the environment variables
+    self.old_env_var = copy.deepcopy(os.environ)
+    if _DOCKER_URI.value:
+      self.command_builder = command_builder.DockerCommandBuilder(
+          _DOCKER_URI.value
+      )
+    else:
+      self.command_builder = command_builder.PythonCommandBuilder()
+
+    self.command_builder.add_mount_map(
+        os.path.expanduser('~'), os.path.expanduser('~')
+    )
+    self.command_builder.add_mount_map(
+        self.local_input_dir(), self.local_input_dir()
+    )
+
+    self.task_cmd_builder = None
+
+  def tearDown(self):
+    super().tearDown()
+    # Restore the original environment variables
+    os.environ.clear()
+    os.environ.update(self.old_env_var)
+
+  def cmd(self):
+    return self.command_builder.build_cmd() + self.task_cmd_builder.build_cmd()
+
+  def run_cmd(self) -> int:
+    return run_cmd(self.cmd(), output_file=None)
+
+  def gcs_output_dir(self):
+    return _GCS_OUTPUT_DIR.value
+
+  def local_input_dir(self):
+    """Returns local input dir in host/docker."""
+    return _LOCAL_INPUT_DIR.value
+
+  def local_output_dir(self):
+    """Returns local output dir in host/docker."""
+    return _LOCAL_OUTPUT_DIR.value
+
+  def get_testcase_name(self):
+    """Returns the function name at the calling site."""
+    # https://docs.python.org/3/library/inspect.html#inspect.FrameInfo
+    cur_frame = inspect.currentframe()
+    # https://stackoverflow.com/a/17366561
+    return cur_frame.f_back.f_code.co_name
+
+
+def get_timestamp():
+  return datetime.datetime.now(datetime.timezone.utc).strftime(
+      '%Y%m%d_%H%M%S%Z'
+  )
+
+
+def download_from_gcs(gcs_uri: str, local_dir: str):
+  if not os.path.exists(local_dir):
+    os.mkdir(local_dir)
+  subprocess.check_output([
+      'gcloud',
+      'storage',
+      'cp',
+      '-r',
+      gcs_uri,
+      local_dir,
+  ])
+
+
+def get_test_data_path(name: str, download: bool = True) -> str:
+  """Gets test data path.
+
+  Args:
+    name: name of the test data
+    download: if True, then download data from GCS and returns its local path.
+
+  Returns:
+    test data path.
+  """
+  if not download:
+    return os.path.join(_GCS_INPUT_DIR.value, name)
+
+  local_data = os.path.join(_LOCAL_INPUT_DIR.value, name)
+  if not os.path.exists(local_data):
+    # If `name` is a file in sub-folders, then create the sub-folders under
+    # `_LOCAL_INPUT_DIR`.
+    local_data_dir = os.path.dirname(local_data)
+    if not os.path.exists(local_data_dir):
+      os.makedirs(local_data_dir)
+
+    download_from_gcs(os.path.join(_GCS_INPUT_DIR.value, name), local_data_dir)
+
+  return local_data
+
+
+def run_cmd(cmd: list[str], output_file: str = None) -> int:
+  """Runs the command and returns the return code.
+
+  Args:
+    cmd: The command to run.
+    output_file: The file to write the output to.
+
+  Returns:
+    The return code of the command.
+  """
+  logging.info('running command: \n%s', ' \\\n'.join(cmd))
+  if _DRY_RUN.value:
+    return 0
+  stdout = sys.stdout if output_file is None else open(output_file, 'w')
+  p = subprocess.Popen(cmd, stdout=stdout, stderr=sys.stderr)
+  try:
+    unused_output, unused_error = p.communicate()
+    return_code = p.returncode
+  except KeyboardInterrupt:
+    p.send_signal(signal.SIGINT)
+    return_code = 0
+  finally:
+    if output_file is not None:
+      stdout.close()
+  return return_code
+
+
+def get_pretrained_model_name_or_path(model_id: str) -> str:
+  # If `model_id` contains `/`, it is assumed to be HF model or model from GCS.
+  if '/' in model_id:
+    return model_id
+
+  return get_test_data_path(model_id, download=True)
+
+
+def is_gpu_h100():
+  """Checks if the GPU is H100."""
+  return 'H100' in torch.cuda.get_device_name()
+
+
+def is_gpu_a100():
+  """Checks if the GPU is A100."""
+  return 'A100' in torch.cuda.get_device_name()
+
+
+def _get_formatted_string(max_seq_length: int) -> str:
+  """Returns the formatted string for max_seq_length.
+
+  Args:
+    max_seq_length: max sequence length to get the formatted string.
+
+  Returns:
+    formatted string for max_seq_length.
+  """
+  return f'{max_seq_length/1024.0:.1f}'
+
+
+def get_benchmark_results(
+    benchmark_file_path: str, max_seq_length: int
+) -> BenchmarkStats:
+  """Gets benchmark results from the benchmark file.
+
+  Args:
+    benchmark_file_path: path to the benchmark file.
+    max_seq_length: max sequence length to get the benchmark results.
+
+  Returns:
+    peak_mem: peak memory in GB.
+    throughput: throughput in tokens/sec.
+  """
+  formatted_max_seq_length = _get_formatted_string(max_seq_length)
+  peak_mem, throughput = None, None
+  with open(benchmark_file_path, 'r') as f:
+    for line in f:
+      if line.startswith(formatted_max_seq_length):
+        metrics = line.split('|')
+        try:
+          peak_mem = float(metrics[1].strip())
+        except ValueError:
+          pass
+        try:
+          throughput = float(metrics[2].strip())
+        except ValueError:
+          pass
+        break
+    else:
+      logging.error(
+          'No metrics found for max_seq_length %s in %s',
+          formatted_max_seq_length,
+          benchmark_file_path,
+      )
+  return BenchmarkStats(peak_mem, throughput)
+
+
+def print_benchmark_file(file_path: str) -> None:
+  """Prints the contents of the file.
+
+  Args:
+    file_path: path to the file.
+  """
+  with open(file_path, 'r') as f:
+    for line in f:
+      logging.info(line.strip())
+
+
+def print_benchmark_results(
+    benchmark_file_path: str, benchmark_type: str
+) -> None:
+  """Prints the benchmark results.
+
+  Args:
+    benchmark_file_path: path to the benchmark file.
+    benchmark_type: type of the benchmark.
+  """
+  benchmark_filename = os.path.basename(benchmark_file_path)
+  logging.info('--------------------------------------------------------------')
+  logging.info('%s benchmark for %s', benchmark_type, benchmark_filename)
+  logging.info('--------------------------------------------------------------')
+  print_benchmark_file(benchmark_file_path)
+
+
+def _calculate_percent_change(
+    actual_value: float, expected_value: float
+) -> float:
+  """Calculates the percent change between the actual and expected values.
+
+  Args:
+    actual_value: actual value to compare.
+    expected_value: expected value to compare.
+
+  Returns:
+    percent change between the actual and expected values.
+  """
+  return ((actual_value - expected_value) / expected_value) * 100.0
+
+
+def compare_benchmark_results(
+    expected_benchmark_file_path: str,
+    actual_benchmark_file_path: str,
+    allowed_threshold: float,
+    max_seq_length: int,
+) -> bool:
+  """Compares if the benchmark results are the similar.
+
+  Args:
+    expected_benchmark_file_path: path to the expected benchmark file.
+    actual_benchmark_file_path: path to the actual benchmark file.
+    allowed_threshold: allowed percent range of the benchmark results.
+    max_seq_length: max sequence length to get the benchmark results.
+
+  Returns:
+    True if the benchmark results are the similar, False otherwise.
+  """
+  benchmark_filename = os.path.basename(expected_benchmark_file_path)
+  expected_results = get_benchmark_results(
+      expected_benchmark_file_path, max_seq_length
+  )
+  expected_peak_mem, expected_throughput = (
+      expected_results.peak_mem,
+      expected_results.throughput,
+  )
+  actual_results = get_benchmark_results(
+      actual_benchmark_file_path, max_seq_length
+  )
+  actual_peak_mem, actual_throughput = (
+      actual_results.peak_mem,
+      actual_results.throughput,
+  )
+  formatted_max_seq_length = _get_formatted_string(max_seq_length)
+
+  # Case 1: both peak mem and throughput are None(ideally due to OOM)
+  if expected_peak_mem is None and actual_peak_mem is None:
+    logging.info(
+        'Both peak mem and throughput are None for max_seq_length %d.',
+        max_seq_length,
+    )
+    return True
+
+  check_oom_exception = _THROUGHPUT_TEST_EXCEPTIONS.get(
+      (benchmark_filename, formatted_max_seq_length), 0.0
+  ) == float('inf')
+  # Case 2: When something strated to fail recently, or something which failed
+  # before but is working now.
+  if expected_peak_mem is None and actual_peak_mem is not None:
+    if check_oom_exception:
+      return True
+    logging.error(
+        'One of the failing benchmarks in %s is passing now for max_seq_length'
+        ' %d. The expected peak mem and throughput are None, but the actual'
+        ' peak mem is %f and actual throughput is %f',
+        benchmark_filename,
+        max_seq_length,
+        actual_peak_mem,
+        actual_throughput,
+    )
+    return False
+  if actual_peak_mem is None and expected_peak_mem is not None:
+    if check_oom_exception:
+      return True
+    logging.error(
+        'One of the passing benchmarks in %s is failing now for max_seq_length'
+        ' %d. The actual peak mem and throughput are None, but the expected'
+        ' peak mem is %f and expected throughput is %f',
+        benchmark_filename,
+        max_seq_length,
+        expected_peak_mem,
+        expected_throughput,
+    )
+    return False
+  # Case 3: When both actual peak mem and throughput lies within the range
+  # of their respective expected values.
+  mem_percent_change = _calculate_percent_change(
+      actual_peak_mem, expected_peak_mem
+  )
+  throughput_percent_change = _calculate_percent_change(
+      actual_throughput, expected_throughput
+  )
+  allowed_threshold = _THROUGHPUT_TEST_EXCEPTIONS.get(
+      (benchmark_filename, formatted_max_seq_length), allowed_threshold
+  )
+
+  if abs(mem_percent_change) > allowed_threshold:
+    logging.error(
+        'The peak memory is changing by more than %f%% for max_seq_length %d.'
+        ' Expected: %f, Actual: %f',
+        allowed_threshold,
+        max_seq_length,
+        expected_peak_mem,
+        actual_peak_mem,
+    )
+    return False
+  if abs(throughput_percent_change) > allowed_threshold:
+    logging.error(
+        'The throughput is changing by more than %f%% for max_seq_length %d.'
+        ' Expected throughput: %f, Actual throughput: %f',
+        allowed_threshold,
+        max_seq_length,
+        expected_throughput,
+        actual_throughput,
+    )
+    return False
+
+  return True
+
+
+def check_benchmark_results(
+    actual_benchmark_file_path: str,
+    model_family: str,
+    allowed_threshold: float,
+    max_seq_length: int,
+) -> bool:
+  """Checks the benchmark result between the actual and expected benchmark files.
+
+  Args:
+    actual_benchmark_file_path: path to the actual benchmark file.
+    model_family: family of the model.
+    allowed_threshold: allowed range of the benchmark results in percent.
+    max_seq_length: max sequence length to get the benchmark results.
+
+  Returns:
+    True if the benchmark results are the similar, False otherwise.
+  """
+  benchmark_filename = os.path.basename(actual_benchmark_file_path)
+  get_test_data_path(_GCS_TESTDATA_DIR)
+  expected_benchmark_file_path = os.path.join(
+      _LOCAL_INPUT_DIR.value,
+      _GCS_TESTDATA_DIR,
+      model_family,
+      benchmark_filename,
+  )
+  print_benchmark_results(expected_benchmark_file_path, 'Expected')
+  print_benchmark_results(actual_benchmark_file_path, 'Actual')
+
+  return compare_benchmark_results(
+      expected_benchmark_file_path,
+      actual_benchmark_file_path,
+      allowed_threshold,
+      max_seq_length,
+  )
+
+
+def list_gcs_directories(bucket: str, directory: str) -> list[str]:
+  """Lists GCS files."""
+  output = subprocess.check_output([
+      'gcloud',
+      'storage',
+      'ls',
+      f'gs://{bucket}/{directory}',
+  ])
+  return output.decode('utf-8').splitlines()
+
+
+def delete_gcs_object(gcs_directory: str):
+  """Deletes GCS object."""
+  subprocess.check_output([
+      'gcloud',
+      'storage',
+      'rm',
+      '-r',
+      f'{gcs_directory}',
+  ])
@@ -0,0 +1,79 @@
+"""Get cluster info from environment variables."""
+
+import dataclasses
+import json
+import os
+
+from absl import logging
+
+
+@dataclasses.dataclass
+class ClusterInfo:
+  """Contains information about the cluster.
+
+  Attributes:
+    primary_node_addr: The address of the primary node.
+    primary_node_port: The port of the primary node.
+    node_rank: The rank of the node.
+    num_nodes: The number of nodes in the cluster.
+  """
+
+  primary_node_addr: str | None = None
+  primary_node_port: str | None = None
+  node_rank: int = 0
+  num_nodes: int = 1
+
+  # Allows unpacking operation like
+  # primary_node_addr, primary_node_port, _, _ = ClusterInfo()
+  # See https://stackoverflow.com/a/70753113
+  def __iter__(self):
+    return iter(dataclasses.astuple(self))
+
+
+def get_cluster_spec() -> ClusterInfo:
+  """Parses CLUSTER_SPEC environment variable and returns the cluster info.
+
+  Returns:
+    A ClusterInfo object.
+  """
+  cluster_spec = os.getenv('CLUSTER_SPEC', None)
+
+  # If CLUSTER_SPEC is not set, use individual vars to construct cluster info.
+  if not cluster_spec:
+    cluster_info = ClusterInfo(
+        primary_node_addr=os.getenv('MASTER_ADDR', None),
+        primary_node_port=os.getenv('MASTER_PORT', None),
+        node_rank=int(os.getenv('RANK', '0')),
+        num_nodes=int(os.getenv('NNODES', '1')),
+    )
+    return cluster_info
+
+  cluster_data = json.loads(cluster_spec)
+  # Get primary node info
+  primary_node = cluster_data['cluster']['workerpool0'][0]
+  logging.info('primary node: %s', primary_node)
+  primary_node_addr, primary_node_port = primary_node.split(':')
+  logging.info('primary node address: %s', primary_node_addr)
+  logging.info('primary node port: %s', primary_node_port)
+
+  # Determine node rank of this machine
+  workerpool = cluster_data['task']['type']
+  if workerpool == 'workerpool0':
+    node_rank = 0
+  elif workerpool == 'workerpool1':
+    # Add 1 for the primary node, since `index` is the index of workerpool1.
+    node_rank = cluster_data['task']['index'] + 1
+  else:
+    raise ValueError(
+        'Only workerpool0 and workerpool1 are supported. Unknown workerpool:'
+        f' {workerpool}'
+    )
+  logging.info('node rank: %s', node_rank)
+
+  # Calculate total nodes.
+  num_nodes = 1  # For the primary node.
+  if 'workerpool1' in cluster_data['cluster']:
+    num_nodes += len(cluster_data['cluster']['workerpool1'])
+  logging.info('num nodes: %s', num_nodes)
+
+  return ClusterInfo(primary_node_addr, primary_node_port, node_rank, num_nodes)
@@ -0,0 +1,24 @@
+"""Utility functions."""
+
+import logging
+import subprocess
+import sys
+import time
+
+
+def run_cmd(cmd: list[str]) -> float:
+  """Runs the command and logs the output.
+
+  Args:
+    cmd: The command to run.
+
+  Returns:
+    The time it took to run the command.
+  """
+  cmd_str = ' \\\n'.join(cmd)
+  logging.info('launching cmd: \n%s', cmd_str)
+  start_time = time.time()
+  subprocess.run(cmd, stdout=sys.stdout, stderr=sys.stdout, check=True)
+  elapsed_time = round(time.time() - start_time, 2)
+  logging.info('Command %s finished in %0.2f seconds.', cmd_str, elapsed_time)
+  return elapsed_time
@@ -0,0 +1,197 @@
+"""Calculate dataset statistics like token, example and character counts."""
+
+from collections.abc import Mapping, Sequence
+import dataclasses
+import json
+from typing import Any
+import datasets
+import numpy as np
+import transformers
+from util import dataset_validation_util
+
+_MAX_NUM_DATASET_SAMPLES = 6
+
+
+@dataclasses.dataclass
+class SupervisedTuningDatasetBucket:
+  """Represents a histogram bucket for tuning dataset distribution stats."""
+
+  count: float = 0
+  left: float = 0
+  right: float = 0
+
+
+@dataclasses.dataclass
+class SupervisedTuningDatasetDistribution:
+  """Represents a histogram with summary statistics for tuning dataset distribution stats."""
+
+  sum: int = 0
+  billable_sum: int = 0
+  min: float = 0
+  max: float = 0
+  mean: float = 0
+  median: float = 0
+  p5: float = 0
+  p95: float = 0
+  buckets: list[SupervisedTuningDatasetBucket] = dataclasses.field(
+      default_factory=list
+  )
+
+
+# Represents detailed tuning dataset statistics.
+@dataclasses.dataclass
+class SupervisedTuningDataStats:
+  """Represents detailed tuning dataset stats."""
+
+  tuning_dataset_example_count: int = 0
+  total_tuning_character_count: int = 0
+  total_billable_token_count: int = 0
+  tuning_step_count: int = 0
+  # Represents a histogram and some summary statistics of the number of input
+  # tokens across examples.
+  user_input_token_distribution: SupervisedTuningDatasetDistribution | None = (
+      None
+  )
+  # Represents a histogram and some summary statistics for the number of output
+  # tokens across examples.
+  user_output_token_distribution: SupervisedTuningDatasetDistribution | None = (
+      None
+  )
+  # Represents the number of "messages" (a single-turn conversation will have a
+  # single message) across examples.
+  user_message_per_example_distribution: (
+      SupervisedTuningDatasetDistribution | None
+  ) = None
+  user_dataset_examples: list[str] = dataclasses.field(default_factory=list)
+
+
+def get_dataset_stats(
+    *,
+    raw: Any,
+    templated: Any,
+    template: str,
+    tokenizer: transformers.PreTrainedTokenizer,
+    column: str,
+    effective_batch_size: int,
+) -> Mapping[str, Any]:
+  """Calculates dataset statistics for managed fine-tuning, e.g., total number of tokens."""
+  tokenized_dataset = templated.map(lambda x: tokenizer(x[column]))
+  inputs = tokenized_dataset["input_ids"]
+  tuning_dataset_example_count = int(len(inputs))
+  total_billable_token_count = int(np.sum([len(ex) for ex in inputs]))
+  total_tuning_character_count = int(
+      np.sum([len(ex[column]) for ex in templated])
+  )
+  tuning_step_count = (
+      tuning_dataset_example_count + effective_batch_size - 1
+  ) // effective_batch_size
+
+  # Assume that data is represented as ChatCompletions or Vertex Text-Bison
+  # formats to extract per-example input/output tokens.
+  user_inputs = []
+  user_outputs = []
+  user_input_messages_counts = []
+
+  for ex in raw:
+    if "messages" in ex:
+      messages = ex["messages"]
+      if messages:
+        # For ChatCompletions assume the last turn (i.e. the instruction
+        # response) is the expected output.
+        user_inputs.append({**ex, "messages": messages[:-1]})
+        user_outputs.append({**ex, "messages": messages[-1:]})
+        # Exclude everything but the last message for the number of input
+        # messages.
+        user_input_messages_counts.append(len(messages[:-1]))
+    elif "input_text" in ex:
+      # For Vertex Text-Bison, the `output_text` field is the expected output.
+      user_inputs.append({**ex, "output_text": ""})
+      user_outputs.append(
+          {**ex, "input_text": ex["output_text"], "output_text": ""}
+      )
+      # Vertex Text-Bison goes from input -> output; i.e. there is only a single
+      # input "message".
+      user_input_messages_counts.append(1)
+
+  def calc_histogram(
+      counts: Sequence[int],
+  ) -> SupervisedTuningDatasetDistribution:
+    mean = np.mean(counts)
+    median = np.median(counts).item()
+    max_count = np.max(counts).item()
+    min_count = np.min(counts).item()
+    count_sum = np.sum(counts).item()
+    p5 = np.percentile(counts, 0.05).item()
+    p95 = np.percentile(counts, 0.95).item()
+    hist, bin_edges = np.histogram(counts, bins=10)
+
+    return SupervisedTuningDatasetDistribution(
+        sum=count_sum,
+        billable_sum=count_sum,
+        min=min_count,
+        max=max_count,
+        mean=mean,
+        median=median,
+        p5=p5,
+        p95=p95,
+        buckets=[
+            SupervisedTuningDatasetBucket(
+                count=hist[i].item(),
+                left=bin_edges[i].item(),
+                right=bin_edges[i + 1].item(),
+            )
+            for i in range(len(hist))
+        ],
+    )
+
+  # Tokenize input and output messages separately to generate separate summary
+  # statistics about them.
+  user_input_token_distribution = None
+  if user_inputs:
+    user_input_dataset = dataset_validation_util.format_dataset(
+        datasets.Dataset.from_list(user_inputs), column, template, tokenizer
+    )
+    user_input_tokenized_dataset = user_input_dataset.map(
+        lambda x: tokenizer(x[column])
+    )
+    user_input_tokens = user_input_tokenized_dataset["input_ids"]
+    user_input_token_counts = np.array([len(ex) for ex in user_input_tokens])
+    user_input_token_distribution = calc_histogram(user_input_token_counts)
+
+  user_output_token_distribution = None
+  if user_outputs:
+    user_output_dataset = dataset_validation_util.format_dataset(
+        datasets.Dataset.from_list(user_outputs), column, template, tokenizer
+    )
+    user_output_tokenized_dataset = user_output_dataset.map(
+        lambda x: tokenizer(x[column])
+    )
+    user_output_tokens = user_output_tokenized_dataset["input_ids"]
+    user_output_token_counts = np.array([len(ex) for ex in user_output_tokens])
+    user_output_token_distribution = calc_histogram(user_output_token_counts)
+
+  user_messages_per_example_distribution = None
+  if user_input_messages_counts:
+    user_input_messages_counts = np.array(user_input_messages_counts)
+    user_messages_per_example_distribution = calc_histogram(
+        user_input_messages_counts
+    )
+
+  user_dataset_examples = [
+      json.dumps(ex)
+      for ex in raw.shuffle().select(
+          range(min(len(raw), _MAX_NUM_DATASET_SAMPLES))
+      )
+  ]
+
+  dataset_stats = SupervisedTuningDataStats(
+      tuning_dataset_example_count=tuning_dataset_example_count,
+      total_tuning_character_count=total_tuning_character_count,
+      total_billable_token_count=total_billable_token_count,
+      tuning_step_count=tuning_step_count,
+      user_input_token_distribution=user_input_token_distribution,
+      user_output_token_distribution=user_output_token_distribution,
+      user_message_per_example_distribution=user_messages_per_example_distribution,
+      user_dataset_examples=user_dataset_examples,
+  )
+  return dataclasses.asdict(dataset_stats)
@@ -0,0 +1,140 @@
+"""Util functions for reporting device (GPU, CPU) stats."""
+
+import dataclasses
+
+import psutil
+import pynvml
+import torch
+
+
+@dataclasses.dataclass
+class GpuStats:
+  """Holds information about GPU usage stats.
+
+  For memory related, see
+  https://pytorch.org/docs/stable/notes/cuda.html#cuda-memory-management
+  """
+
+  # device id
+  device_id: int
+  # memory reserved.
+  reserved: float
+  # memory occupied.
+  occupied: float
+  # memory reserved, but not used.
+  unused: float
+  # nvidia-smi usually reports more memory usages than pytorch (for driver,
+  # kernel and etc). `smi_diff` tracks this difference.
+  smi_diff: float
+  # Gpu utilization.
+  util: float
+
+  # Allows unpacking operation like
+  # device_id, reserved, occupied, unused, smi_diff, util = GpuStats(...)
+  # See https://stackoverflow.com/a/70753113
+  def __iter__(self):
+    return iter(dataclasses.astuple(self))
+
+
+def gpu_stats() -> GpuStats:
+  """Reports GPU memory usage and utilization."""
+  # See https://pytorch.org/docs/stable/notes/cuda.html#memory-management
+  bytes_per_gb = 1024.0**3
+  device = torch.cuda.current_device()
+  occupied = torch.cuda.memory_allocated(device) / bytes_per_gb
+  reserved = torch.cuda.memory_reserved(device) / bytes_per_gb
+  unused = reserved - occupied
+
+  def smi_mem(device):
+    try:
+      pynvml.nvmlInit()
+      handle = pynvml.nvmlDeviceGetHandleByIndex(device)
+      info = pynvml.nvmlDeviceGetMemoryInfo(handle)
+      return info.used / bytes_per_gb
+    except pynvml.NVMLError:
+      return 0.0
+
+  mem_used_smi = smi_mem(device)
+  smi_diff = mem_used_smi - reserved
+
+  util = torch.cuda.utilization(device)
+  return GpuStats(device, reserved, occupied, unused, smi_diff, util)
+
+
+def gpu_stats_str(stats: GpuStats | None = None) -> str:
+  if stats is None:
+    stats = gpu_stats()
+  device, reserved, occupied, unused, smi_diff, util = stats
+  return (
+      f"GPU ({device=}) memory: {reserved:.2f}({occupied=:.2f}, {unused=:.2f}),"
+      f" {smi_diff=:.2f} GB. Utilization: {util:.2f}%"
+  )
+
+
+@dataclasses.dataclass
+class CpuStats:
+  """Holds information about CPU usage stats."""
+
+  # Total CPU virtual memory i.e. virtual memory allocated + unallocated.
+  total_virtual_mem: float
+  # CPU virtual memory available for use.
+  unallocated_virtual_mem: float
+  # CPU virtual memory already used.
+  allocated_virtual_mem: float
+  # Total CPU swap memory i.e. swap memory allocated + unallocated.
+  total_swap_mem: float
+  # CPU swap memory available for use.
+  unallocated_swap_mem: float
+  # CPU swap memory already used.
+  allocated_swap_mem: float
+  # CPU utilization percentage.
+  utilization: float
+
+
+def cpu_stats() -> CpuStats:
+  """Reports CPU memory usage and utilization."""
+
+  # https://psutil.readthedocs.io/en/latest/#memory
+  gb = 1024.0**3
+  vmem = psutil.virtual_memory()
+  vmem_total = vmem.total / gb
+  vmem_available = vmem.available / gb
+  vmem_used = vmem_total - vmem_available
+  smem = psutil.swap_memory()
+  swap_total = smem.total / gb
+  swap_free = smem.free / gb
+  swap_used = smem.used / gb
+  # https://psutil.readthedocs.io/en/latest/#psutil.cpu_percent
+  cpu_util = psutil.cpu_percent(interval=1e-6)
+  return CpuStats(
+      total_virtual_mem=vmem_total,
+      unallocated_virtual_mem=vmem_available,
+      allocated_virtual_mem=vmem_used,
+      total_swap_mem=swap_total,
+      unallocated_swap_mem=swap_free,
+      allocated_swap_mem=swap_used,
+      utilization=cpu_util,
+  )
+
+
+def cpu_stats_str(stats: CpuStats | None = None) -> str:
+  """Returns a string representation of the CPU stats."""
+
+  if stats is None:
+    stats = cpu_stats()
+  total, occupied, unused = (
+      stats.total_virtual_mem,
+      stats.allocated_virtual_mem,
+      stats.unallocated_virtual_mem,
+  )
+  virtual_mem = (
+      f"CPU virtual memory: {total:.2f}({occupied=:.2f}, {unused=:.2f}) GB"
+  )
+  total, occupied, unused = (
+      stats.total_swap_mem,
+      stats.allocated_swap_mem,
+      stats.unallocated_swap_mem,
+  )
+  swap_mem = f"CPU swap memory: {total:.2f}({occupied=:.2f}, {unused=:.2f}) GB"
+  percent = stats.utilization
+  return f"{virtual_mem} {swap_mem} CPU Utilization: {percent:.2f}%"
@@ -0,0 +1,126 @@
+"""Different trainer callbacks for PEFT Trainer."""
+
+from collections.abc import MutableMapping
+import math
+import time
+
+from absl import logging
+import accelerate
+from transformers import TrainingArguments
+from transformers.trainer_callback import TrainerCallback
+from transformers.trainer_callback import TrainerControl
+from transformers.trainer_callback import TrainerState
+
+from util import device_stats
+
+
+class TrainerStatsCallback(TrainerCallback):
+  """Trainer callback to report trainer stats."""
+
+  def __init__(self, max_seq_length, filename=None):
+    self._max_seq_length = max_seq_length
+    self._filename = filename
+
+    self._partial_state = accelerate.PartialState()
+    self._start_time = float('nan')
+    self._prev_time = float('nan')
+    self._peak_mem = 0.0
+    self._avg_throughput = 0.0
+
+  def on_log(
+      self,
+      args: TrainingArguments,
+      state: TrainerState,
+      control: TrainerControl,
+      logs: MutableMapping[str, float] | None = None,
+      **kwargs,
+  ) -> None:
+    """Calculates perplexity from train loss.
+
+    Args:
+      args: Arguments passed to the trainer.
+      state: State of the trainer.
+      control: Control of the trainer.
+      logs: A dict of logs from the training loop.
+      **kwargs: Additional keyword arguments, not used in this callback.
+    """
+    del kwargs  # Unused.
+    if self._partial_state.is_main_process:
+      train_loss = logs.get('loss') if logs is not None else None
+      if train_loss is not None:
+        perplexity = round(float(math.exp(train_loss)), 4)
+        logs['perplexity'] = perplexity
+
+  def on_step_end(
+      self,
+      args: TrainingArguments,
+      state: TrainerState,
+      control: TrainerControl,
+      **kwargs,
+  ):
+    if self._partial_state.is_main_process:
+      if state.global_step == 1:
+        self._prev_time = time.time()
+        self._prev_num_token = state.num_input_tokens_seen
+        throughput = 0.0
+      else:
+        cur_time = time.time()
+        cur_num_token = state.num_input_tokens_seen
+        throughput = (cur_num_token - self._prev_num_token) / (
+            cur_time - self._prev_time
+        )
+        self._prev_time = cur_time
+        self._prev_num_token = cur_num_token
+        self._avg_throughput += (throughput - self._avg_throughput) / (
+            state.global_step - 1
+        )
+
+      gpu_stats = device_stats.gpu_stats()
+      self._peak_mem = max(
+          gpu_stats.reserved + gpu_stats.smi_diff, self._peak_mem
+      )
+      logging.info(
+          'on_step_end: Throughput: %.2f token/s. %s, %s',
+          throughput,
+          device_stats.gpu_stats_str(gpu_stats),
+          device_stats.cpu_stats_str(),
+      )
+
+  def on_train_begin(
+      self,
+      args: TrainingArguments,
+      state: TrainerState,
+      control: TrainerControl,
+      **kwargs,
+  ):
+    if self._partial_state.is_main_process:
+      self._start_time = time.time()
+      logging.info(
+          'on_train_begin: %s, %s',
+          device_stats.gpu_stats_str(),
+          device_stats.cpu_stats_str(),
+      )
+
+  def on_train_end(
+      self,
+      args: TrainingArguments,
+      state: TrainerState,
+      control: TrainerControl,
+      **kwargs,
+  ):
+    if self._partial_state.is_main_process:
+      train_time = time.time() - self._start_time
+      throughput = state.num_input_tokens_seen / train_time
+      logging.info(
+          'training time %.2f s, throughput (including overhead, e.g., ckpt'
+          ' saving): %.2f token/s, peak_mem: %.2f GB',
+          train_time,
+          throughput,
+          self._peak_mem,
+      )
+      if self._filename:
+        with open(self._filename, 'a') as out_f:
+          out_f.write(
+              f'{self._max_seq_length/1024.0:.1f} | {self._peak_mem:.2f} |'
+              f' {self._avg_throughput:.2f}\n'
+          )
@@ -0,0 +1,18 @@
+group:
+- vertex
+task: custom_loglikelihood
+dataset_path: json
+dataset_name: null
+output_type: loglikelihood
+training_split: null
+validation_split: null
+test_split: test
+doc_to_text: "Request: {{prompt}}\nResponse:"
+doc_to_target: " {{ground_truth}}"
+metric_list:
+- metric: perplexity
+  aggregation: perplexity
+  higher_is_better: false
+- metric: acc
+  aggregation: mean
+  higher_is_better: true
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: fp16
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: fp16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_config_file: /diffusers/examples/vertex_vision_model_garden_peft/zero2.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_config_file: /diffusers/examples/vertex_vision_model_garden_peft/zero2.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_config_file: /diffusers/examples/vertex_vision_model_garden_peft/zero3.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,17 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_config_file: /diffusers/examples/vertex_vision_model_garden_peft/zero3.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Gemma2DecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 16
+num_processes: 128
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 3
+num_processes: 24
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 4
+num_processes: 32
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: HYBRID_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 2
+num_processes: 16
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: HYBRID_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 3
+num_processes: 24
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: HYBRID_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 4
+num_processes: 32
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,28 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  fsdp_backward_prefetch: NO_PREFETCH
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+  fsdp_activation_checkpointing: false
+main_training_function: main
+mixed_precision: bf16
+machine_rank: 0
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
@@ -0,0 +1,24 @@
+{
+  "zero_optimization": {
+    "stage": 2,
+    "contiguous_gradients": false,
+    "overlap_comm": false
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "wall_clock_breakdown": false
+}
@@ -0,0 +1,31 @@
+{
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": false,
+    "contiguous_gradients": false,
+    "sub_group_size": 0,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 0,
+    "stage3_max_reuse_distance": 0,
+    "stage3_gather_16bit_weights_on_model_save": true
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "wall_clock_breakdown": false
+}
@@ -0,0 +1,43 @@
+# Doc about format of conda environment file
+# https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#create-env-file-manually
+name: merge
+channels:
+- nodefaults
+- conda-forge
+dependencies:
+- _libgcc_mutex=0.1=conda_forge
+- _openmp_mutex=4.5=2_gnu
+- bzip2=1.0.8=h4bc722e_7
+- ca-certificates=2024.7.4=hbcca054_0
+- ld_impl_linux-64=2.40=hf3520f5_7
+- libffi=3.4.2=h7f98852_5
+- libgcc-ng=14.1.0=h77fa898_0
+- libgomp=14.1.0=h77fa898_0
+- libnsl=2.0.1=hd590300_0
+- libsqlite=3.46.0=hde9e2c9_0
+- libuuid=2.38.1=h0b41bf4_0
+- libxcrypt=4.4.36=hd590300_1
+- libzlib=1.3.1=h4ab18f5_1
+- ncurses=6.5=h59595ed_0
+- openssl=3.3.1=h4bc722e_2
+- pip=24.2=pyhd8ed1ab_0
+- python=3.10.14=hd12c33a_0_cpython
+- readline=8.2=h8228510_1
+- setuptools=72.1.0=pyhd8ed1ab_0
+- tk=8.6.13=noxft_h4845f30_101
+- tzdata=2024a=h0c530f3_0
+- wheel=0.44.0=pyhd8ed1ab_0
+- xz=5.2.6=h166bdaf_0
+- pip:
+  - --extra-index-url https://download.pytorch.org/whl/cu121
+  - absl-py==2.1.0
+  - accelerate==0.34.2  # Needed for fp8
+  - datasets==2.19.2
+  - fbgemm-gpu==0.8.0+cu121  # Needed for fp8
+  - kfp==2.5.0
+  - peft==0.12.0
+  - protobuf==3.20.3
+  - pynvml==11.5.3
+  - torch==2.4.0+cu121  # Needed for fp8
+  - transformers==4.47.1
+  - trl==0.11.2
@@ -0,0 +1,32 @@
+# Doc about format of requirement file
+# https://pip.pypa.io/en/stable/reference/requirements-file-format
+
+--extra-index-url https://download.pytorch.org/whl/cu118
+--extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+
+# keep sorted
+accelerate==0.34.2
+auto_gptq==0.7.1+cu118
+autoawq==0.2.8
+bitsandbytes==0.43.2
+cloudml-hypertune==0.1.0.dev6
+datasets==2.20.0
+deepspeed==0.15.2
+diffusers==0.25.1
+evaluate==0.4.3
+fsspec==2024.3.1
+gcsfs==2024.3.1
+immutabledict==4.2.1
+ninja==1.11.1 # Needed to avoid `ninja 1.11.1.1 is not supported on this platform` error
+nltk==3.9.1
+optimum==1.17.1
+peft==0.12.0
+pynvml==11.5.3
+rouge_score==0.1.2
+torch==2.2.2+cu118
+torchvision==0.17.2+cu118
+transformers==4.47.1
+trl==0.11.2
+wandb==0.17.1
+ydata-profiling==4.7.0 # Upgrade the version from 4.6.0 to 4.7.0 to fix the old `pydantic` package error.
+psutil==6.0.0
@@ -0,0 +1,83 @@
+# Dockerfile for PEFT Training.
+#
+# To build:
+# docker build -f model_oss/peft/dockerfile/train.Dockerfile . -t ${YOUR_IMAGE_TAG}
+#
+# To push to gcr:
+# docker tag ${YOUR_IMAGE_TAG} gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+# docker push gcr.io/${YOUR_PROJECT}/${YOUR_IMAGE_TAG}
+
+# Picked from https://cloud.google.com/deep-learning-containers/docs/choosing-container#pytorch
+FROM us-docker.pkg.dev/deeplearning-platform-release/gcr.io/pytorch-cu121.2-2.py310:m123
+RUN apt-get update && \
+    apt-get upgrade -y && \
+    apt-get install -y curl git wget software-properties-common vim libaio-dev && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Copy license.
+RUN wget https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/main/LICENSE
+
+
+# Install libraries.
+ENV PIP_ROOT_USER_ACTION=ignore
+RUN pip install --upgrade pip
+
+# Remove packages that are not needed and are causing conflicts.
+# dataproc_jupyter_plugin was installed as a part of pytorch-cu121.2-2.py310
+# container which we don't need. It depends on ibis-framework and bigframes.
+# The package and its dependencies request lower versions of pyarrow/pydantic
+# than deepspeed/datasets. So, dataproc_jupyter_plugin conflicts with
+# deepspeed/datasets.
+RUN pip uninstall -y dataproc_jupyter_plugin ibis-framework bigframes
+
+# Prefer to install with requirement file as much as possible for reasons
+# described in b/355034754.
+COPY model_oss/peft/train/vmg/dockerfile/requirements.txt /tmp/requirements.txt
+RUN pip install -r /tmp/requirements.txt
+
+# flash-attn cannot be installed with the requirement file approach above
+# because of the `no-build-isolation` requirement.
+#
+# It is OK to install it after other packages FOR NOW because it only has
+# limited dependencies. And there's no concern about it overwriting previously
+# installed packages.
+# https://github.com/Dao-AILab/flash-attention/blob/v2.6.3/setup.py#L523
+RUN pip install flash-attn==2.6.3 --no-build-isolation
+
+# Install `diffusers` library as editable and in root folder (/) on purpose.
+RUN git clone --depth 1 --branch v0.25.1 https://github.com/huggingface/diffusers.git
+# Remove `diffusers` (NOTE that the dependency libraries are kept).
+RUN pip uninstall -y diffusers
+# Using `--no-deps` option to make sure previously installed packages are not
+# overwritten.
+RUN pip install --no-deps -e /diffusers
+
+# Make sure there's no inconsistent pip libraries.
+RUN pip check
+
+# Install merge related packages in a separate env.
+COPY model_oss/peft/train/vmg/dockerfile/merge_env.yaml /tmp/merge_env.yaml
+RUN conda env create -n merge --yes --file /tmp/merge_env.yaml
+RUN conda init
+
+# Switch to diffusers examples folder.
+WORKDIR /diffusers/examples
+
+RUN mkdir -p ./vertex_vision_model_garden_peft/
+COPY model_oss/peft/train/vmg/configs/* ./vertex_vision_model_garden_peft/
+COPY model_oss/peft/train/vmg/*.py ./vertex_vision_model_garden_peft/train/vmg/
+COPY model_oss/peft/train/vmg/templates /diffusers/examples/util/templates
+COPY model_oss/peft/train/util/*.py /diffusers/examples/util/
+COPY model_oss/util/* /diffusers/examples/util/
+COPY model_oss/notebook_util/dataset_validation_util.py /diffusers/examples/util
+COPY model_oss/peft/train/vmg/tests/*.py ./vertex_vision_model_garden_peft/tests/
+COPY model_oss/peft/train/test_utils/test_util.py ./vertex_vision_model_garden_peft/tests/
+COPY model_oss/peft/train/test_utils/command_builder.py ./vertex_vision_model_garden_peft/tests/
+
+RUN chmod a+rwX -R /diffusers/examples/
+ENV PYTHONPATH /diffusers/examples/
+# Must disable torch XLA, otherwise runtime uses CPU even if GPU exists.
+ENV USE_TORCH_XLA 0
+
+ENTRYPOINT ["python3", "./vertex_vision_model_garden_peft/train/vmg/train_entrypoint.py"]
@@ -0,0 +1,284 @@
+"""Library for running evaluations during training."""
+
+from collections.abc import Callable, Mapping, MutableMapping, Sequence
+import dataclasses
+import string
+from typing import Type
+
+from absl import logging
+import evaluate
+import numpy as np
+import torch
+import transformers
+
+from util import dataset_validation_util
+from util import constants
+
+
+_STRING_TRANSLATOR = str.maketrans("", "", string.punctuation)
+
+_GREATER_IS_BETTER_MAP = {
+    "loss": False,
+    "perplexity": False,
+    "bleu": True,
+    "google_bleu": True,
+    "rouge1": True,
+    "rouge2": True,
+    "rougeL": True,
+    "rougeLsum": True,
+}
+
+
+@dataclasses.dataclass(frozen=True)
+class EvalConfig:
+  """Configuration for running evaluations during training.
+
+  Attributes:
+    steps: The number of steps to run evaluation.
+    tasks: The list of tasks to run evaluation on.
+    per_device_batch_size: The per device batch size for evaluation.
+    limit: The maximum number of examples to evaluate.
+    metric_name: The name of the metric to compute.
+    tokenize_dataset: Whether to tokenize the dataset.
+    dataset_path: The path to the dataset.
+    split: The split of the dataset to evaluate.
+    template: The template to use for the dataset.
+    column: The column name of the dataset.
+    metric_for_best_model: The metric to use for loading the best model.
+  """
+
+  steps: int
+  per_device_batch_size: int
+  limit: float | None
+  metric_name: Sequence[str]
+  tokenize_dataset: bool
+  dataset_path: str = ""
+  split: str = "test"
+  template: str = ""
+  column: str = constants.DEFAULT_TRAIN_COLUMN
+  metric_for_best_model: str | None = None
+
+
+def create_trainer(
+    cls: Type[transformers.Trainer],
+    eval_config: EvalConfig | None,
+    tokenizer: transformers.PreTrainedTokenizerBase | None,
+    args: transformers.TrainingArguments,
+    **kwargs,
+) -> transformers.Trainer:
+  """Creates a trainer. If eval config is provided, injects evaluation loop.
+
+  Args:
+    cls: The trainer class.
+    eval_config: The evaluation config.
+    tokenizer: The tokenizer.
+    args: The training arguments.
+    **kwargs: The keyword arguments.
+
+  Returns:
+    A trainer.
+  """
+  if not eval_config:
+    return cls(args=args, **kwargs)
+
+  args.eval_strategy = "steps"
+  args.eval_steps = eval_config.steps
+  args.per_device_eval_batch_size = eval_config.per_device_batch_size
+  args.metric_for_best_model = eval_config.metric_for_best_model
+  args.greater_is_better = _GREATER_IS_BETTER_MAP.get(
+      eval_config.metric_for_best_model, None
+  )
+  args.save_strategy = (
+      transformers.trainer_utils.SaveStrategy.STEPS
+      if eval_config.metric_for_best_model is None
+      else transformers.trainer_utils.SaveStrategy.BEST
+  )
+
+  kwargs["tokenizer"] = tokenizer
+
+  try:
+    _, eval_dataset = dataset_validation_util.load_dataset_with_template(
+        dataset_name=eval_config.dataset_path,
+        split=eval_config.split,
+        input_column=eval_config.column,
+        template=eval_config.template,
+        tokenizer=tokenizer,
+    )
+    if eval_config.limit is not None:
+      if eval_config.limit >= 1:
+        limit = int(eval_config.limit)
+      else:
+        limit = int(eval_config.limit * len(eval_dataset))
+      eval_dataset = eval_dataset.select(range(limit))
+    if tokenizer is not None:
+      eval_dataset = dataset_validation_util.get_filtered_dataset(
+          dataset=eval_dataset,
+          input_column=eval_config.column,
+          max_seq_length=kwargs["max_seq_length"],
+          tokenizer=tokenizer,
+      )
+    if eval_config.tokenize_dataset:
+      eval_dataset = eval_dataset.map(
+          lambda samples: tokenizer(samples[eval_config.column])
+      )
+    kwargs["eval_dataset"] = eval_dataset
+  except (OSError, ValueError, IndexError) as e:
+    logging.warning(
+        "Failed to load eval dataset %s. Evaluation will be skipped.\n%s",
+        eval_config.dataset_path,
+        e,
+    )
+    del args.evaluation_strategy
+    del args.eval_steps
+    del args.per_device_eval_batch_size
+  return cls(args=args, **kwargs)
+
+
+def _cleanup_text(text: str) -> str:
+  """Cleans up the prediction and references text.
+
+  Args:
+    text: The text to clean up.
+
+  Returns:
+    Cleaned up text.
+  """
+  text = text.translate(_STRING_TRANSLATOR)
+  text = text.strip()
+  text = " ".join(text.split())
+  return text.lower()
+
+
+def create_compute_metrics(
+    tokenizer: transformers.PreTrainedTokenizerBase,
+    eval_metrics: Mapping[str, evaluate.EvaluationModule],
+) -> Callable[[transformers.EvalPrediction], MutableMapping[str, float]]:
+  """Creates a compute_metrics function using Hugging Face evaluate library.
+
+  Args:
+    tokenizer: The tokenizer for decoding predictions.
+    eval_metrics: The eval metrics to compute.
+
+  Returns:
+    Function that computes comprehensive metrics.
+  """
+
+  def _preprocess_data(
+      predictions: np.ndarray, labels: np.ndarray
+  ) -> tuple[Sequence[str], Sequence[str]]:
+    """Preprocesses predictions and lavels before evaluation.
+
+    Args:
+      predictions: The predictions to preprocess.
+      labels: The labels to preprocess.
+
+    Returns:
+      A tuple (preprocessed predictions, labels).
+    """
+    # Handle padding and special tokens.
+    predictions = np.where(
+        predictions != -100, predictions, tokenizer.pad_token_id
+    )
+    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
+
+    # Decode to text.
+    pred_texts = tokenizer.batch_decode(predictions, skip_special_tokens=True)
+    label_texts = tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+    # Clean up text.
+    cleaned_pred_texts = [_cleanup_text(text) for text in pred_texts]
+    cleaned_label_texts = [_cleanup_text(text) for text in label_texts]
+
+    return cleaned_pred_texts, cleaned_label_texts
+
+  def _compute_metrics_with_tokenizer(
+      eval_pred: transformers.EvalPrediction,
+  ) -> MutableMapping[str, float]:
+    """Computes metrics using Hugging Face evaluate library.
+
+    Args:
+      eval_pred: The evaluation prediction.
+
+    Returns:
+      A dictionary of metrics.
+    """
+    predictions, perplexities = eval_pred.predictions
+    labels = eval_pred.label_ids
+
+    pred_texts, label_texts = _preprocess_data(predictions, labels)
+
+    metrics = {}
+
+    for eval_metric, computed_eval_metric in eval_metrics.items():
+      match eval_metric:
+        case "perplexity":
+          # We don't use the perplexity from HF Evaluate since it loads the
+          # model again. This causes an increase in the GPU utilization and
+          # hence an OOM. Due to this, we compute the perplexity ourselves
+          # using the eval_loss over the unmasked tokens in
+          # preprocess_logits_for_metrics fn.
+          metrics[eval_metric] = np.mean(perplexities)
+        case "bleu" | "google_bleu":
+          num_valid_labels = len(list(filter(None, label_texts)))
+          if num_valid_labels:
+            eval_score = computed_eval_metric.compute(
+                predictions=pred_texts,
+                references=[[text] for text in label_texts],
+            )
+            metrics[eval_metric] = eval_score[eval_metric]
+          else:
+            metrics[eval_metric] = 0.0
+        case "rouge1" | "rouge2" | "rougeL" | "rougeLsum":
+          rouge_scores = computed_eval_metric.compute(
+              predictions=pred_texts,
+              references=label_texts,
+              use_stemmer=True,
+          )
+          metrics[eval_metric] = rouge_scores[eval_metric]
+
+    pred_lengths = [len(pred.split()) for pred in pred_texts]
+    label_lengths = [len(label.split()) for label in label_texts]
+
+    metrics["gen_len"] = np.mean(pred_lengths)
+    metrics["ref_len"] = np.mean(label_lengths)
+    metrics["length_ratio"] = np.mean(
+        [len(p) / len(r) if r else 0 for p, r in zip(pred_texts, label_texts)]
+    )
+
+    # Round all metrics to 4 decimal places.
+    return {k: round(float(v), 4) for k, v in metrics.items()}
+
+  return _compute_metrics_with_tokenizer
+
+
+def preprocess_logits_for_metrics(
+    logits: torch.Tensor, labels: torch.Tensor
+) -> tuple[torch.Tensor, torch.Tensor]:
+  """Preprocesses the logits before caching them for eval metric calculation.
+
+  Args:
+    logits: Logits predicted by the model.
+    labels: Ground truth labels.
+
+  Returns:
+    A tuple (pred_ids, perplexities).
+  """
+  # Calculate prediction IDs.
+  pred_ids = logits.argmax(dim=-1)
+
+  # This step shifts the logits and labels to align them correctly, where we are
+  # predicting the next token in a sequence. The last logit doesn't have a
+  # corresponding label, and the first label doesn't have a preceding logit to
+  # predict it. This calculation of perplexity is inspired from
+  # https://github.com/huggingface/evaluate/blob/main/metrics/perplexity/perplexity.py.
+  shift_logits = logits[..., :-1, :].contiguous()
+  shift_labels = labels[..., 1:].contiguous()
+  attn_mask = shift_labels != -100
+  loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+
+  perplexities = torch.exp(
+      (loss_fct(shift_logits.transpose(1, 2), shift_labels) * attn_mask).sum(1)
+      / attn_mask.sum(1)
+  )
+
+  return (pred_ids, perplexities)
@@ -0,0 +1,905 @@
+"""Instruct/Chat with LoRA models."""
+
+from collections.abc import Callable, Mapping, Sequence
+import datetime
+import json
+import os
+from typing import Any
+import warnings
+
+from absl import app
+from absl import flags
+from absl import logging
+from accelerate import DistributedType
+from accelerate import PartialState
+import bitsandbytes as bnb
+import evaluate
+from peft import get_peft_model
+from peft import LoraConfig
+import torch
+import transformers
+import trl
+import wandb
+
+from util import dataset_validation_util
+from util import dataset_stats
+from util import device_stats
+from vertex_vision_model_garden_peft.train.vmg import callbacks
+from vertex_vision_model_garden_peft.train.vmg import eval_lib
+from vertex_vision_model_garden_peft.train.vmg import utils
+from util import constants
+from util import fileutils
+
+
+_PRETRAINED_MODEL_NAME_OR_PATH = flags.DEFINE_string(
+    'pretrained_model_name_or_path',
+    None,
+    'The pretrained model name or path. Supported models can be causal language'
+    ' modeling models from https://github.com/huggingface/peft/tree/main. Note,'
+    ' there might be different paddings for different models. This tool assumes'
+    ' the pretrained_model_name_or_path contains model name, and then choose'
+    ' proper padding methods. e.g. it must contain `llama` for `Llama2'
+    ' models`.',
+    required=True,
+)
+
+_HUGGINGFACE_ACCESS_TOKEN = flags.DEFINE_string(
+    'huggingface_access_token',
+    None,
+    'The access token for loading huggingface gated models.',
+)
+
+_TRAIN_DATASET = flags.DEFINE_string(
+    'train_dataset',
+    None,
+    'The training dataset name in huggingface or path.',
+)
+
+_OUTPUT_DIR = flags.DEFINE_string(
+    'output_dir',
+    None,
+    'The output directory.',
+)
+
+_LOGGING_OUTPUT_DIR = flags.DEFINE_string(
+    'logging_output_dir',
+    '',
+    'The logging output directory, which defaults to same as output_dir.',
+)
+
+_PRECISION_MODE = flags.DEFINE_enum(
+    'precision_mode',
+    constants.PRECISION_MODE_16,
+    [
+        constants.PRECISION_MODE_4,
+        constants.PRECISION_MODE_8,
+        constants.PRECISION_MODE_16,
+        constants.PRECISION_MODE_16B,
+        constants.PRECISION_MODE_32,
+    ],
+    'Precision to load model weights for finetuning.',
+)
+
+_LORA_RANK = flags.DEFINE_integer(
+    'lora_rank',
+    16,
+    'The rank of the update matrices, expressed in int. Lower rank results in'
+    ' smaller update matrices with fewer trainable parameters, referring to'
+    ' https://huggingface.co/docs/peft/conceptual_guides/lora.',
+)
+
+_LORA_ALPHA = flags.DEFINE_integer(
+    'lora_alpha',
+    32,
+    'LoRA scaling factor, referring to'
+    ' https://huggingface.co/docs/peft/conceptual_guides/lora.',
+)
+
+_LORA_DROPOUT = flags.DEFINE_float(
+    'lora_dropout',
+    0.05,
+    'dropout probability of the LoRA layers, referring to'
+    ' https://huggingface.co/docs/peft/task_guides/token-classification-lora.',
+)
+
+_WARMUP_STEPS = flags.DEFINE_integer(
+    'warmup_steps',
+    10,
+    'Number of steps for the warmup in the learning rate scheduler.',
+)
+
+_WARMUP_RATIO = flags.DEFINE_float(
+    'warmup_ratio',
+    0.03,
+    'The warmup ratio in the learning rate scheduler.',
+)
+
+_WEIGHT_DECAY = flags.DEFINE_float(
+    'weight_decay',
+    0.001,
+    'The weight decay in the learning rate scheduler.',
+)
+
+_NUM_TRAIN_EPOCHS = flags.DEFINE_float(
+    'num_train_epochs',
+    None,
+    'The number of training epochs. Only used for'
+    ' "sequence-classification-lora" with an integer value and for'
+    ' "instruct-lora" with a float value allowed.',
+)
+
+_MAX_STEPS = flags.DEFINE_integer(
+    'max_steps',
+    None,
+    'Total number of training steps. Overrides num_train_epochs if set. Only'
+    ' used for "instruct-lora."',
+)
+
+_MAX_SEQ_LENGTH = flags.DEFINE_integer(
+    'max_seq_length',
+    512,
+    'The maximum sequence length.',
+)
+
+_LEARNING_RATE = flags.DEFINE_float(
+    'learning_rate',
+    2e-4,
+    'The learning rate after the potential warmup period.',
+)
+
+_TRAIN_COLUMN = flags.DEFINE_string(
+    'train_column',
+    constants.DEFAULT_TRAIN_COLUMN,
+    'The instruct column in dataset.',
+)
+
+_REPORT_TO = flags.DEFINE_string(
+    'report_to',
+    constants.REPORT_TO_NONE,
+    'Where logging is reported to, which can be tensorboard or none.',
+)
+
+_PER_DEVICE_TRAIN_BATCH_SIZE = flags.DEFINE_integer(
+    'per_device_train_batch_size',
+    4,
+    'The per device train batch size.',
+)
+
+_GRADIENT_ACCUMULATION_STEPS = flags.DEFINE_integer(
+    'gradient_accumulation_steps',
+    4,
+    'The gradient accumulation steps.',
+)
+
+_GRADIENT_CHECKPOINTING = flags.DEFINE_boolean(
+    'gradient_checkpointing',
+    False,
+    'Whether to enable gradient checkpointing.',
+)
+
+_ENABLE_PEFT = flags.DEFINE_boolean(
+    'enable_peft',
+    True,
+    'Whether to enable peft.',
+)
+_TRAIN_TEMPLATE = flags.DEFINE_string(
+    'train_template',
+    None,
+    'Template for formatting language model training data. Must be a filename'
+    ' under `templates` folder, without `.json` extension, e.g. `alpaca`, or a'
+    ' Cloud Storage URI to a JSON file.',
+)
+
+_OPTIMIZER = flags.DEFINE_string(
+    'optimizer',
+    'adamw_torch',
+    'The optimizer.',
+)
+
+_LR_SCHEDULER_TYPE = flags.DEFINE_string(
+    'lr_scheduler_type',
+    'cosine',
+    'The learning rate scheduler type.',
+)
+
+_SAVE_STEPS = flags.DEFINE_integer(
+    'save_steps',
+    10,
+    'The save steps.',
+)
+
+_LOGGING_STEPS = flags.DEFINE_integer(
+    'logging_steps',
+    10,
+    'The logging steps.',
+)
+
+_EVAL_STEPS = flags.DEFINE_integer(
+    'eval_steps',
+    10,
+    'The number of training steps between evaluations.',
+)
+
+_TRAIN_SPLIT = flags.DEFINE_string(
+    'train_split',
+    'train',
+    'The train split name.',
+)
+
+_PER_DEVICE_EVAL_BATCH_SIZE = flags.DEFINE_integer(
+    'per_device_eval_batch_size',
+    1,
+    'The per device batch size for model evaluation.',
+)
+
+
+_EVAL_LIMIT = flags.DEFINE_float(
+    'eval_limit',
+    None,
+    'Limit the number of examples per task. If <1, limit is a percentage of the'
+    ' total number of examples.',
+)
+
+_EVAL_METRIC_NAME = flags.DEFINE_list(
+    'eval_metric_name',
+    ['loss'],
+    'A comma-separated list of metric names to aggregate during model'
+    ' evaluation. The supported metrics are: '
+    + ', '.join(constants.SUPPORTED_EVAL_METRICS),
+)
+
+_EVAL_DATASET = flags.DEFINE_string(
+    'eval_dataset',
+    None,
+    'The Hugging Face dataset name or path to use for evaluation.',
+)
+
+# We set the default eval split as `test`, based on observation from
+# https://huggingface.co/datasets/timdettmers/openassistant-guanaco/viewer/default/test.
+_EVAL_SPLIT = flags.DEFINE_string(
+    'eval_split',
+    'test',
+    'Eval split name in the eval dataset.',
+)
+
+_EVAL_TEMPLATE = flags.DEFINE_string(
+    'eval_template',
+    None,
+    'Template for formatting language model evaluation data.'
+    ' Must be a filename under `templates` folder, without `.json` extension,'
+    ' e.g. `alpaca`, or a Cloud Storage URI to a JSON file.',
+)
+
+_EVAL_COLUMN = flags.DEFINE_string(
+    'eval_column',
+    None,
+    'Eval column name in the eval dataset.',
+)
+
+_METRIC_FOR_BEST_MODEL = flags.DEFINE_string(
+    'metric_for_best_model',
+    None,
+    'If set, the best model is saved at the end of training based on the'
+    ' metric',
+)
+
+_TRAIN_PRECISION = flags.DEFINE_enum(
+    'train_precision',
+    constants.PRECISION_MODE_16B,
+    [
+        constants.PRECISION_MODE_16,
+        constants.PRECISION_MODE_16B,
+        constants.PRECISION_MODE_32,
+    ],
+    'Precision to train the model.',
+)
+
+_EXAMPLE_PACKING = flags.DEFINE_boolean(
+    'example_packing',
+    False,
+    'Enables example packing during training, which uses '
+    '`ConstantLengthDataset` under the hood.',
+)
+
+_INPUT_MASKING = flags.DEFINE_boolean(
+    'input_masking',
+    False,
+    'If set, it uses DataCollatorForCompletionOnlyLM to train the model on the'
+    ' generated prompts only, i.e., masking out the input',
+)
+
+_ATTN_IMPLEMENTATION = flags.DEFINE_string(
+    'attn_implementation',
+    None,
+    'Attention implementation, can be `eager`, `sdpa` or `flash_attention_2`',
+)
+
+_MAX_GRAD_NORM = flags.DEFINE_float(
+    'max_grad_norm',
+    0.3,
+    'Maximum gradient norm used for gradient clipping',
+)
+
+_WARNINGS_FILTER = flags.DEFINE_string(
+    'warnings_filter',
+    'ignore',
+    'Warning filter as defined in '
+    'https://docs.python.org/3/library/warnings.html#the-warnings-filter',
+)
+
+_LOGGER_LEVEL = flags.DEFINE_string(
+    'logger_level',
+    'passive',
+    'logging level passed to TrainingArguments. Note that this is for python'
+    ' logging module, NOT the one from absl',
+)
+
+_BENCHMARK_OUT_FILE = flags.DEFINE_string(
+    'benchmark_out_file', None, 'file path for writing benchmark result'
+)
+
+
+_NCCL_TIMEOUT = flags.DEFINE_integer(
+    'nccl_timeout', 6000, 'nccl timeout in seconds'
+)
+
+_TUNING_DATA_STATS_FILE = flags.DEFINE_string(
+    'tuning_data_stats_file', None, 'file path for writing tuning data stats.'
+)
+
+_TARGET_MODULES = flags.DEFINE_list(
+    'target_modules', None, 'The names of the modules to apply LoRA adapter to.'
+)
+
+_MAX_GPU_MEMORY_FRACTION = flags.DEFINE_float(
+    'max_gpu_memory_fraction',
+    '0.9',
+    'Maximum GPU memory a caching allocator is allowed to use per GPU.',
+)
+
+
+@flags.multi_flags_validator(
+    [
+        _INPUT_MASKING.name,
+        _EXAMPLE_PACKING.name,
+    ],
+    message='`example_packing=True` does not work with `input_masking=True`',
+)
+def check_example_packing(flags_dict: Mapping[str, Any]) -> bool:
+  """Check to make sure example packing is enabled properly.
+
+  Args:
+    flags_dict: Dictionary containing flags to check.
+
+  Returns:
+    If `example_packing` is set properly.
+  """
+  if flags_dict[_INPUT_MASKING.name] and flags_dict[_EXAMPLE_PACKING.name]:
+    return False
+  return True
+
+
+@flags.multi_flags_validator(
+    [
+        _INPUT_MASKING.name,
+        _TRAIN_TEMPLATE.name,
+    ],
+    message='`train_template` should be provided if using `input_masking=True`',
+)
+def check_input_masking(flags_dict: Mapping[str, Any]) -> bool:
+  """Check to make sure input_masking is enabled properly.
+
+  Args:
+    flags_dict: Dictionary containing flags to check.
+
+  Returns:
+    If `input_masking` is set properly
+  """
+  if (
+      flags_dict[_INPUT_MASKING.name]
+      and flags_dict[_TRAIN_TEMPLATE.name] is None
+  ):
+    return False
+  return True
+
+
+@flags.multi_flags_validator(
+    [
+        _EVAL_DATASET.name,
+        _EVAL_METRIC_NAME.name,
+    ],
+    message=(
+        '`eval_metric_name` should be a valid metric name and present when'
+        ' eval_dataset is provided.'
+    ),
+)
+def _validate_eval_metrics(flags_dict: Mapping[str, Any]) -> bool:
+  """Validates the eval metric name.
+
+  Args:
+    flags_dict: Dictionary containing flags to check.
+
+  Returns:
+    If the eval metrics are valid.
+  """
+  if flags_dict[_EVAL_DATASET.name] is None:
+    return True
+  eval_metrics = flags_dict[_EVAL_METRIC_NAME.name]
+  for eval_metric in eval_metrics:
+    if eval_metric not in constants.SUPPORTED_EVAL_METRICS:
+      raise flags.ValidationError(f'Invalid eval metric: {eval_metric}')
+  if 'perplexity' in eval_metrics and 'loss' not in eval_metrics:
+    _EVAL_METRIC_NAME.value.append('loss')
+    logging.warning(
+        'Adding `loss` to eval_metric_name because `perplexity` is present.'
+    )
+  return True
+
+
+@flags.multi_flags_validator(
+    [
+        _METRIC_FOR_BEST_MODEL.name,
+        _EVAL_METRIC_NAME.name,
+    ],
+    message='`metric_for_best_model` should be in `eval_metric_name`.',
+)
+def _validate_metric_for_best_model(flags_dict: Mapping[str, Any]) -> bool:
+  """Validates the metric for best model.
+
+  Args:
+    flags_dict: Dictionary containing flags to check.
+
+  Returns:
+    If the metric for best model is valid.
+  """
+  if flags_dict[_METRIC_FOR_BEST_MODEL.name] is None:
+    return True
+
+  metric_for_best_model = flags_dict[_METRIC_FOR_BEST_MODEL.name]
+  eval_metric_name = flags_dict[_EVAL_METRIC_NAME.name]
+
+  if metric_for_best_model not in eval_metric_name:
+    raise flags.ValidationError(
+        'Invalid metric for picking the best model:'
+        f' {metric_for_best_model}. The metric should be one'
+        f' of the {eval_metric_name}.'
+    )
+  return True
+
+
+# References:
+# Huggingface SFT trainer example:
+# https://github.com/huggingface/trl/blob/main/examples/scripts/sft_trainer.py.
+# Huggingface sagemaker example:
+# https://github.com/huggingface/notebooks/blob/main/sagemaker/28_train_llms_with_qlora/scripts/run_clm.py.
+
+
+def _calculate_hf_eval_metrics(
+    tokenizer: transformers.PreTrainedTokenizerBase,
+    eval_config: eval_lib.EvalConfig | None,
+) -> tuple[
+    Callable[[transformers.EvalPrediction], Mapping[str, float]], torch.Tensor
+]:
+  """Calculates the HF evaluation metrics.
+
+  Args:
+    tokenizer: The tokenizer to use for evaluation.
+    eval_config: The evaluation config to use.
+
+  Returns:
+    The compute metrics and preprocess logits for metrics.
+  """
+  if eval_config is None:
+    return None, None
+  hf_eval_metrics = {}
+  for metric in eval_config.metric_name:
+    if metric in constants.SUPPORTED_HF_EVAL_METRICS:
+      if metric in constants.ROUGE_VARIANTS:
+        hf_eval_metrics[metric] = evaluate.load('rouge')
+      else:
+        hf_eval_metrics[metric] = evaluate.load(metric)
+
+  if not hf_eval_metrics:
+    return None, None
+  return (
+      eval_lib.create_compute_metrics(tokenizer, hf_eval_metrics),
+      eval_lib.preprocess_logits_for_metrics,
+  )
+
+
+# Copied from https://github.com/artidoro/qlora/blob/main/qlora.py.
+def find_all_linear_names(
+    model: transformers.AutoModelForCausalLM, precision_mode: str
+) -> Sequence[str]:
+  """Finds all linear module names."""
+  if precision_mode == constants.PRECISION_MODE_4:
+    cls = bnb.nn.Linear4bit
+  elif precision_mode == constants.PRECISION_MODE_8:
+    cls = bnb.nn.Linear8bitLt
+  else:
+    cls = torch.nn.Linear
+  lora_module_names = set()
+  for name, module in model.named_modules():
+    if isinstance(module, cls):
+      names = name.split('.')
+      lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+  if 'lm_head' in lora_module_names:  # needed for 16-bit
+    lora_module_names.remove('lm_head')
+  return list(lora_module_names)
+
+
+def finetune_instruct(
+    pretrained_model_name_or_path: str,
+    train_dataset: str,
+    output_dir: str,
+    logging_output_dir: str,
+    lora_rank: int = 64,
+    lora_alpha: int = 16,
+    lora_dropout: float = 0.1,
+    warmup_ratio: int = 0.03,
+    num_train_epochs: float | None = None,
+    max_steps: int | None = None,
+    warmup_steps: int = 10,
+    max_seq_length: int = 512,
+    learning_rate: float = 2e-4,
+    precision_mode: str = None,
+    train_column: str = constants.DEFAULT_TRAIN_COLUMN,
+    per_device_train_batch_size: int = 4,
+    gradient_accumulation_steps: int = 4,
+    optim: str = 'paged_adamw_32bit',
+    weight_decay: float = 0.001,
+    gradient_checkpointing: bool = False,
+    enable_peft: bool = True,
+    train_template: str = None,
+    lr_scheduler_type: str = 'constant',
+    save_steps: int = 10,
+    logging_steps: int = 10,
+    train_split: str = 'train',
+    eval_config: eval_lib.EvalConfig | None = None,
+    report_to: str = constants.REPORT_TO_NONE,
+    access_token: str | None = None,
+    train_precision: str = constants.PRECISION_MODE_16B,
+    example_packing: bool = False,
+    attn_implementation: str | None = None,
+    max_grad_norm: float = 0.3,
+    input_masking: bool = False,
+    logger_level: str = 'passive',
+    benchmark_out_file: str | None = None,
+    tuning_data_stats_file: str | None = None,
+    target_modules: str | None = None,
+) -> None:
+  """Finetunes instruct."""
+  logging.info(
+      'on entering instruct_lora, %s,\n%s',
+      device_stats.gpu_stats_str(),
+      device_stats.cpu_stats_str(),
+  )
+  gradient_checkpointing_kwargs = {}
+  # DDP provides limited support with the reentrant variant of gradient
+  # checkpoint [1]. Below is an indirect way of checking whether DDP will be
+  # used. It is "indirect" because there are complex logic under the hood of
+  # `SFTTrainer` and since those are not public API, they might change as we
+  # update the library.
+  if PartialState().distributed_type == DistributedType.MULTI_GPU:
+    gradient_checkpointing_kwargs['use_reentrant'] = False
+
+  tokenizer = dataset_validation_util.load_tokenizer(
+      pretrained_model_name_or_path,
+      'right',
+      access_token=access_token,
+  )
+
+  train_dataset, train_dataset_with_template = (
+      dataset_validation_util.load_dataset_with_template(
+          train_dataset,
+          split=train_split,
+          input_column=train_column,
+          template=train_template,
+          tokenizer=tokenizer,
+      )
+  )
+  train_dataset_with_template = dataset_validation_util.get_filtered_dataset(
+      dataset=train_dataset_with_template,
+      input_column=train_column,
+      max_seq_length=max_seq_length,
+      tokenizer=tokenizer,
+  )
+
+  if tuning_data_stats_file:
+    with PartialState().main_process_first():
+      effective_batch_size = (
+          per_device_train_batch_size
+          * gradient_accumulation_steps
+          * PartialState().num_processes
+      )
+      logging.info(
+          'getting tuning data stats with effective batch size %s',
+          effective_batch_size,
+      )
+      train_dataset_stats = dataset_stats.get_dataset_stats(
+          raw=train_dataset,
+          templated=train_dataset_with_template,
+          template=train_template,
+          tokenizer=tokenizer,
+          column=train_column,
+          effective_batch_size=effective_batch_size,
+      )
+      logging.info('stats: %s', train_dataset_stats)
+      tuning_data_stats_file = dataset_validation_util.force_gcs_fuse_path(
+          tuning_data_stats_file
+      )
+      with open(tuning_data_stats_file, 'w') as out_f:
+        json.dump(train_dataset_stats, out_f)
+
+  model = utils.load_model(
+      pretrained_model_name_or_path=pretrained_model_name_or_path,
+      tokenizer=tokenizer,
+      precision_mode=precision_mode,
+      gradient_checkpointing=gradient_checkpointing,
+      access_token=access_token,
+      attn_implementation=attn_implementation,
+      train_precision=train_precision,
+  )
+
+  if enable_peft:
+    if target_modules is None:
+      target_modules = find_all_linear_names(
+          model, precision_mode=precision_mode
+      )
+    logging.info('applying lora adapters to modules: %s', target_modules)
+    peft_config = LoraConfig(
+        lora_alpha=lora_alpha,
+        lora_dropout=lora_dropout,
+        r=lora_rank,
+        bias='none',
+        task_type='CAUSAL_LM',
+        target_modules=target_modules,
+    )
+    # If we pass in `peft_config` to SFTTrainer, it does a lot of magic under
+    # the hood, e.g., calling `prepare_model_for_kbit_training` before calling
+    # `get_peft_model`, which may revert other changes we did before. That's why
+    # we are calling `get_peft_model` explicitly here.
+    model = get_peft_model(model, peft_config)
+    adapter_for_eval_dir = os.path.join(output_dir, 'adapter_for_eval')
+    logging.info('saving adapter for evaluation to %s...', adapter_for_eval_dir)
+    peft_config.save_pretrained(adapter_for_eval_dir)
+    # This is to work-around mix-precision training. This issue is not fixed as
+    # of transformers==4.41.2.
+    # See b/332760883#comment30 for more details.
+    if precision_mode in (
+        constants.PRECISION_MODE_16,
+        constants.PRECISION_MODE_16B,
+    ):
+      for param in filter(lambda p: p.requires_grad, model.parameters()):
+        param.data = param.data.to(torch.float32)
+
+  if not logging_output_dir:
+    logging_output_dir = output_dir
+
+  # To use singleton PartialState() without re-initializing it. See
+  # b/357970482#comment3
+  accelerator_config = {'use_configured_state': True}
+
+  training_arguments = transformers.TrainingArguments(
+      report_to=report_to,
+      output_dir=output_dir,
+      per_device_train_batch_size=per_device_train_batch_size,
+      gradient_accumulation_steps=gradient_accumulation_steps,
+      optim=optim,
+      save_steps=save_steps,
+      save_total_limit=3,
+      logging_dir=os.path.join(logging_output_dir, 'logs'),
+      logging_steps=logging_steps,
+      learning_rate=learning_rate,
+      fp16=(train_precision == constants.PRECISION_MODE_16),
+      bf16=(train_precision == constants.PRECISION_MODE_16B),
+      max_grad_norm=max_grad_norm,
+      num_train_epochs=num_train_epochs if num_train_epochs else -1,
+      max_steps=max_steps if max_steps else -1,
+      warmup_ratio=warmup_ratio,
+      warmup_steps=warmup_steps,
+      group_by_length=False,
+      lr_scheduler_type=lr_scheduler_type,
+      gradient_checkpointing=gradient_checkpointing,
+      gradient_checkpointing_kwargs=gradient_checkpointing_kwargs,
+      weight_decay=weight_decay,
+      log_level=logger_level,
+      accelerator_config=accelerator_config,
+      include_num_input_tokens_seen=True,
+  )
+  trainer_kwargs = {}
+  if input_masking and train_template:
+    template_json = dataset_validation_util.get_template(
+        template_path=train_template
+    )
+    instruction_sep = dataset_validation_util.get_instruction_separator(
+        template_json
+    )
+    response_sep = dataset_validation_util.get_response_separator(template_json)
+    if not response_sep:
+      raise ValueError(
+          '`response_separator` must be provided to use'
+          ' `DataCollatorForCompletionOnlyLM`'
+      )
+
+    trainer_kwargs['data_collator'] = trl.DataCollatorForCompletionOnlyLM(
+        instruction_template=instruction_sep,
+        response_template=response_sep,
+        tokenizer=tokenizer,
+    )
+    logging.info('using DataCollatorForCompletionOnlyLM')
+
+  trainer_stats_callback = callbacks.TrainerStatsCallback(
+      max_seq_length, benchmark_out_file
+  )
+  compute_metrics, preprocess_logits = _calculate_hf_eval_metrics(
+      tokenizer, eval_config
+  )
+
+  trainer = eval_lib.create_trainer(
+      cls=trl.SFTTrainer,
+      eval_config=eval_config,
+      model=model,
+      train_dataset=train_dataset_with_template,
+      dataset_text_field=train_column,
+      max_seq_length=max_seq_length,
+      tokenizer=tokenizer,
+      args=training_arguments,
+      packing=example_packing,
+      callbacks=[trainer_stats_callback],
+      compute_metrics=compute_metrics,
+      preprocess_logits_for_metrics=preprocess_logits,
+      **trainer_kwargs,
+  )
+
+  # `eval_lib.create_trainer` might modify the training args. Printing here
+  # should capture what will be used by the trainer.
+  if PartialState().is_main_process:
+    logging.info('training args: %s', trainer.args)
+
+  if enable_peft:
+    trainer.model.print_trainable_parameters()
+
+  if trainer.is_fsdp_enabled:
+    logging.info('Trainer running with FSDP.')
+  elif trainer.is_deepspeed_enabled:
+    logging.info('Trainer running with DeepSpeed.')
+  else:
+    logging.info('Trainer running without parallelism.')
+
+  trainer.train()
+
+  # Always save the final checkpoint.
+  final_checkpoint = utils.get_final_checkpoint_path(output_dir)
+  logging.info('The final checkpoint is: %s.', final_checkpoint)
+
+  if trainer.is_fsdp_enabled:
+    trainer.accelerator.state.fsdp_plugin.set_state_dict_type('FULL_STATE_DICT')
+    # This method saves the sharded weights like `accelerator.save_state`, see
+    # https://huggingface.co/docs/accelerate/en/usage_guides/fsdp#saving-and-loading
+    trainer.save_model(output_dir)
+    model = trainer.model
+    state_dict = trainer.accelerator.get_state_dict(model)
+    # To aggregate the weights from all the devices, we need to use
+    # `state_dict=state_dict`.
+    model.save_pretrained(
+        final_checkpoint,
+        state_dict=state_dict,
+        is_main_process=PartialState().is_main_process,
+        save_embedding_layers=False,  # Only pad token is added. See go/lora-adapter-pad-token #pylint: disable=line-too-long
+    )
+  else:
+    trainer.model.save_pretrained(
+        final_checkpoint,
+        is_main_process=PartialState().is_main_process,
+        save_embedding_layers=False,  # Only pad token is added. See go/lora-adapter-pad-token #pylint: disable=line-too-long
+    )
+
+  if eval_config is not None and trainer.eval_dataset is not None:
+    metrics = trainer.evaluate(metric_key_prefix='eval')
+    # Both `log_metrics` and `save_metrics` are multiple process safe.
+    # https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/trainer_pt_utils.py#L911 #pylint: disable=line-too-long
+    # https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/trainer_pt_utils.py#L1001 #pylint: disable=line-too-long
+    trainer.log_metrics('eval', metrics)
+    trainer.save_metrics('eval', metrics)
+
+  if not enable_peft:
+    tokenizer.save_pretrained(
+        final_checkpoint, is_main_process=PartialState().is_main_process
+    )
+
+
+def main(unused_argv: Sequence[str]) -> None:
+  # This needs to be called before any other PartialState() calls.
+  utils.init_partial_state(
+      timeout=datetime.timedelta(seconds=_NCCL_TIMEOUT.value)
+  )
+
+  torch.cuda.set_per_process_memory_fraction(
+      _MAX_GPU_MEMORY_FRACTION.value, device=PartialState().local_process_index
+  )
+
+  utils.print_library_versions()
+  warnings.simplefilter(_WARNINGS_FILTER.value)
+
+  pretrained_model_name_or_path = fileutils.force_gcs_path(
+      _PRETRAINED_MODEL_NAME_OR_PATH.value
+  )
+  if dataset_validation_util.is_gcs_path(pretrained_model_name_or_path):
+    pretrained_model_name_or_path = (
+        dataset_validation_util.download_gcs_uri_to_local(
+            pretrained_model_name_or_path
+        )
+    )
+
+  # GCS Fuse does not sync flushed files if not closed. See b/361771727.
+  logging_output_dir = fileutils.force_gcs_path(_LOGGING_OUTPUT_DIR.value)
+
+  # Creates evaluation config.
+  if _EVAL_DATASET.value:
+    eval_config = eval_lib.EvalConfig(
+        per_device_batch_size=_PER_DEVICE_EVAL_BATCH_SIZE.value,
+        limit=_EVAL_LIMIT.value,
+        metric_name=_EVAL_METRIC_NAME.value,
+        steps=_EVAL_STEPS.value,
+        dataset_path=dataset_validation_util.force_gcs_fuse_path(
+            _EVAL_DATASET.value
+        ),
+        split=_EVAL_SPLIT.value,
+        template=_EVAL_TEMPLATE.value,
+        column=_EVAL_COLUMN.value,
+        tokenize_dataset=False,
+        metric_for_best_model=_METRIC_FOR_BEST_MODEL.value,
+    )
+  else:
+    eval_config = None
+
+  if _REPORT_TO.value == constants.REPORT_TO_WANDB:
+    wandb.login()
+
+  finetune_instruct(
+      pretrained_model_name_or_path=pretrained_model_name_or_path,
+      train_dataset=_TRAIN_DATASET.value,
+      output_dir=_OUTPUT_DIR.value,
+      logging_output_dir=logging_output_dir,
+      precision_mode=_PRECISION_MODE.value,
+      lora_rank=_LORA_RANK.value,
+      lora_alpha=_LORA_ALPHA.value,
+      lora_dropout=_LORA_DROPOUT.value,
+      warmup_ratio=_WARMUP_RATIO.value,
+      num_train_epochs=_NUM_TRAIN_EPOCHS.value,
+      warmup_steps=_WARMUP_STEPS.value,
+      max_steps=_MAX_STEPS.value,
+      max_seq_length=_MAX_SEQ_LENGTH.value,
+      learning_rate=_LEARNING_RATE.value,
+      train_column=_TRAIN_COLUMN.value,
+      per_device_train_batch_size=_PER_DEVICE_TRAIN_BATCH_SIZE.value,
+      optim=_OPTIMIZER.value,
+      weight_decay=_WEIGHT_DECAY.value,
+      gradient_accumulation_steps=_GRADIENT_ACCUMULATION_STEPS.value,
+      gradient_checkpointing=_GRADIENT_CHECKPOINTING.value,
+      enable_peft=_ENABLE_PEFT.value,
+      train_template=_TRAIN_TEMPLATE.value,
+      lr_scheduler_type=_LR_SCHEDULER_TYPE.value,
+      save_steps=_SAVE_STEPS.value,
+      logging_steps=_LOGGING_STEPS.value,
+      train_split=_TRAIN_SPLIT.value,
+      eval_config=eval_config,
+      report_to=_REPORT_TO.value,
+      access_token=_HUGGINGFACE_ACCESS_TOKEN.value,
+      train_precision=_TRAIN_PRECISION.value,
+      example_packing=_EXAMPLE_PACKING.value,
+      attn_implementation=_ATTN_IMPLEMENTATION.value,
+      max_grad_norm=_MAX_GRAD_NORM.value,
+      input_masking=_INPUT_MASKING.value,
+      logger_level=_LOGGER_LEVEL.value,
+      benchmark_out_file=_BENCHMARK_OUT_FILE.value,
+      tuning_data_stats_file=_TUNING_DATA_STATS_FILE.value,
+      target_modules=_TARGET_MODULES.value,
+  )
+  # Frees the model from GPU.
+  utils.force_gc()
+
+
+if __name__ == '__main__':
+  app.run(main)
@@ -0,0 +1,108 @@
+"""Script to merge PEFT adapter with base model."""
+
+from collections.abc import Mapping, Sequence
+from typing import Any
+
+from absl import app
+from absl import flags
+
+from util import dataset_validation_util
+from vertex_vision_model_garden_peft.train.vmg import utils
+from util import constants
+from util import fileutils
+
+
+_PRETRAINED_MODEL_NAME_OR_PATH = flags.DEFINE_string(
+    'pretrained_model_name_or_path',
+    None,
+    'The pretrained model id. Supported models can be causal language modeling'
+    ' models from https://github.com/huggingface/peft/tree/main. Note, there'
+    ' might be different paddings for different models. This tool assumes the'
+    ' pretrained_model_name_or_path contains model name, and then choose proper'
+    ' padding methods. e.g. it must contain `llama` for `Llama2 models`.',
+    required=True,
+)
+
+_MERGE_BASE_AND_LORA_OUTPUT_DIR = flags.DEFINE_string(
+    'merge_base_and_lora_output_dir',
+    None,
+    'The directory to store the merged model with the base and lora adapter.',
+)
+
+_MERGE_MODEL_PRECISION_MODE = flags.DEFINE_enum(
+    'merge_model_precision_mode',
+    constants.PRECISION_MODE_16B,
+    [
+        constants.PRECISION_MODE_4,
+        constants.PRECISION_MODE_8,
+        constants.PRECISION_MODE_FP8,
+        constants.PRECISION_MODE_16,
+        constants.PRECISION_MODE_16B,
+        constants.PRECISION_MODE_32,
+    ],
+    'Merging model precision mode.',
+)
+
+_FINETUNED_LORA_MODEL_DIR = flags.DEFINE_string(
+    'finetuned_lora_model_dir',
+    None,
+    'The directory storing finetuned LoRA model weights.',
+)
+
+_HUGGINGFACE_ACCESS_TOKEN = flags.DEFINE_string(
+    'huggingface_access_token',
+    None,
+    'The access token for loading huggingface gated models.',
+)
+
+
+@flags.multi_flags_validator(
+    [
+        _PRETRAINED_MODEL_NAME_OR_PATH.name,
+        _FINETUNED_LORA_MODEL_DIR.name,
+        _MERGE_BASE_AND_LORA_OUTPUT_DIR.name,
+    ],
+)
+def check_merge_lora_model_flags(flags_dict: Mapping[str, Any]) -> bool:
+  """Check if required flags are set on merge model LoRA task.
+
+  Args:
+    flags_dict: Dictionary containing task and flags to check.
+
+  Returns:
+    If required flags are not None.
+  """
+  return all(map(lambda x: x is not None, flags_dict.values()))
+
+
+def main(unused_argv: Sequence[str]) -> None:
+  pretrained_model_name_or_path = fileutils.force_gcs_path(
+      _PRETRAINED_MODEL_NAME_OR_PATH.value
+  )
+  if dataset_validation_util.is_gcs_path(pretrained_model_name_or_path):
+    pretrained_model_name_or_path = (
+        dataset_validation_util.download_gcs_uri_to_local(
+            pretrained_model_name_or_path
+        )
+    )
+
+  finetuned_lora_model_dir = fileutils.force_gcs_path(
+      _FINETUNED_LORA_MODEL_DIR.value
+  )
+  if dataset_validation_util.is_gcs_path(finetuned_lora_model_dir):
+    finetuned_lora_model_dir = (
+        dataset_validation_util.download_gcs_uri_to_local(
+            finetuned_lora_model_dir
+        )
+    )
+  utils.merge_causal_language_model_with_lora(
+      pretrained_model_name_or_path=pretrained_model_name_or_path,
+      precision_mode=_MERGE_MODEL_PRECISION_MODE.value,
+      finetuned_lora_model_dir=finetuned_lora_model_dir,
+      merged_model_output_dir=_MERGE_BASE_AND_LORA_OUTPUT_DIR.value,
+      access_token=_HUGGINGFACE_ACCESS_TOKEN.value,
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)
@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# Run copybara first:
+# cloud/ml/applications/vision/model_garden/copybara/run_copybara_local.sh
+# Run docker build:
+# cloud/ml/applications/vision/model_garden/model_oss/peft/train/vmg/scripts/build_train_docker.sh
+
+set -x
+set -e
+
+COPYBARA_DIR="/tmp/train_docker/"
+
+pushd "${COPYBARA_DIR}"
+
+PROJECT="cloud-nas-260507"
+IMAGE_TAG="gcr.io/${PROJECT}/pytorch-peft-train:${USER}-test"
+
+docker build -f model_oss/peft/train/vmg/dockerfile/train.Dockerfile . -t "${IMAGE_TAG}"
+docker push "${IMAGE_TAG}"
+
+popd
@@ -0,0 +1,98 @@
+# Vertex Model Garden Training Dataset Template
+
+## Overview
+
+Vertex Model Garden training provides templates for streamlined preprocessing of
+datasets. Although datasets often have intricate structures, the supported LLM
+models accept only flat strings. A template facilitates parsing a dataset and
+preprocessing it to be compatible with the model.
+
+When fine-tuning a pretrained model, it is advisable to maintain the same format
+as the original training data. A template helps replicate the format, ensuring
+consistency and potentially enhancing the fine-tuning process.
+
+Both multi-turn messages and single instruction-response pairs are supported.
+Multi-turn messages are accommodated using a more general `chat_template` field,
+whereas simple instruction-response pair datasets are supported through the
+`prompt_input` field.
+
+A template is a JSON file consisting of string key-value pairs. Refer to the
+following for the definitions of the supported fields.
+
+## Template field documentation
+
+**description**: An explanation of the template.
+
+**source**: Information about the origin of the template.
+
+**chat_template**: A
+[jinja template](https://jinja.palletsprojects.com/en/3.1.x/templates/) that can
+be used to parse a chat dataset. This is the same format as
+[HF chat templates](https://huggingface.co/docs/transformers/main/en/chat_templating).
+To create a chat_template, use the `messages` variable to be filled with the
+sample. The flag `--instruct_column_in_dataset` identifies which column will be
+passed to the `messages` variable in the chat_template. This field is mutually
+exclusive with `prompt_input` and `prompt_no_input`.
+
+**prompt_input**: A string template that is used when value for the input column
+exists in the sample. It should be able to be formatted with the
+[str.format](https://docs.python.org/3/library/stdtypes.html#str.format) method.
+The input column is specified with the flag `--instruct_column_in_dataset`. Used
+for instruction dataset. This field is mutually exclusive with `chat_template`.
+
+**prompt_no_input**: A string template that is used when value for the input
+column does not exist in the sample. It should be able to be formatted with the
+[str.format](https://docs.python.org/3/library/stdtypes.html#str.format) method.
+The input column is specified with the flag `--instruct_column_in_dataset`. Used
+for instruction dataset. This field is mutually exclusive with `chat_template`.
+
+**instruction_separator**: A unique string used to indicate the start of the
+instructions. If not specified, every token after response_separator will be
+treated as a response, and every token before the first response_separator will
+be treated as instruction.
+
+**response_separator**: A unique string used to indicate the start of the
+response. This field is required if `--completion_only` flag is set to `True`.
+
+## Example templates
+
+-   See the list of all supported templates [here](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/main/community-content/vertex_model_garden/model_oss/peft/train/vmg/templates).
+-   For an example with `chat_template` see the JSON template below.
+
+```
+{
+  "description": "Chat template used by Llama 3.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
+```
+
+-   For an example with `prompt_input` see the JSON template below. In this case
+    the flag `--instruct_column_in_dataset=text` should be set, and there must
+    be a column named `text` in the dataset.
+
+```
+{
+  "description": "Template for openassistant-guanaco dataset.",
+  "source": "https://huggingface.co/datasets/timdettmers/openassistant-guanaco",
+  "prompt_input": "{text}",
+  "instruction_separator": "### Human:",
+  "response_separator": "### Assistant:"
+}
+```
+
+-   For an example with `prompt_no_input` see the JSON template below. In this
+    case the flag `--instruct_column_in_dataset=input` should be set, and there
+    must be columns named `input` and `instruction` in the dataset.
+
+```
+{
+  "description": "Template used by Alpaca-LoRA.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca.json",
+  "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
+```
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by Alpaca-LoRA.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca.json",
+  "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "A shorter template to experiment with.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca_short.json",
+  "prompt_input": "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Gemma. 'assistant' role is replaced by 'model'",
+  "source": "https://huggingface.co/google/gemma-1.1-2b-it/blob/bf4924f313df5166dee1467161e886e55f2eb4d4/tokenizer_config.json#L1507",
+  "chat_template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
+  "instruction_separator": "<start_of_turn>user\n",
+  "response_separator": "<start_of_turn>model\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by Llama 3, accepting text-bison format.",
+  "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-text-models-supervised#dataset-format",
+  "prompt_input": "\n\n<|start_header_id|>user<|end_header_id|>\n\n{input_text}<|eot_id|>\n\n<|start_header_id|>assistant<|end_header_id|>\n\n{output_text}<|eot_id|>",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Llama 3.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '\n\n<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '\n\n<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Mistral.",
+  "source": "https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/utils/chat_templates.py",
+  "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
+  "instruction_separator": "[INST]",
+  "response_separator": "[/INST]"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by openai chat.",
+  "source": "https://platform.openai.com/docs/api-reference/fine-tuning/chat-input",
+  "chat_template": "{% set loop_messages = messages %}{% set content = '' %}{% for message in loop_messages %}{% set content = content ~ '\n\n<|start_header_id|>' ~ message.role ~ '<|end_header_id|>\n\n' %}{% if message.content is string %}{% set content = content ~ message.content|trim ~ '<|eot_id|>' %}{% else %}{% set content = content ~ message.content|join(' ', attribute='text')|trim ~ '<|eot_id|>' %}{% endif %}{% if loop.index0 == 0 %}{% set content = bos_token ~ content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by openai completion.",
+  "source": "https://platform.openai.com/docs/api-reference/fine-tuning/completions-input",
+  "prompt_input": "\n\n<|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|>\n\n<|start_header_id|>assistant<|end_header_id|>\n\n{completion}<|eot_id|>",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template for openassistant-guanaco dataset.",
+  "source": "https://huggingface.co/datasets/timdettmers/openassistant-guanaco",
+  "prompt_input": "{text}",
+  "instruction_separator": "### Human:",
+  "response_separator": "### Assistant:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Qwen 2.5.",
+  "source": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/tokenizer_config.json#L198",
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "instruction_separator": "<|im_start|>user\n",
+  "response_separator": "<|im_start|>assistant\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used for chat based models.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '\n\n<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '\n\n<|start_header_id|>model<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>model<|end_header_id|>\n\n"
+}
@@ -0,0 +1,453 @@
+# pylint: disable=W,C,R
+
+# DO NOT MODIFY: this file is auto-generated
+# See go/vmg-oss-peft-tests#command-builder-genpy
+
+
+class InstructLoraCommandBuilder:
+
+  def __init__(self):
+    self._config_file = None
+    self._task = None
+    self._gcs_rsync_interval_secs = None
+    self._pretrained_model_name_or_path = None
+    self._train_dataset = None
+    self._train_split = None
+    self._train_template = None
+    self._train_column = None
+    self._output_dir = None
+    self._merge_base_and_lora_output_dir = None
+    self._logging_output_dir = None
+    self._per_device_train_batch_size = None
+    self._gradient_accumulation_steps = None
+    self._lora_rank = None
+    self._lora_alpha = None
+    self._lora_dropout = None
+    self._max_steps = None
+    self._num_train_epochs = None
+    self._max_seq_length = None
+    self._learning_rate = None
+    self._lr_scheduler_type = None
+    self._precision_mode = None
+    self._train_precision = None
+    self._gradient_checkpointing = None
+    self._example_packing = None
+    self._attn_implementation = None
+    self._optimizer = None
+    self._warmup_ratio = None
+    self._report_to = None
+    self._save_steps = None
+    self._logging_steps = None
+    self._huggingface_access_token = None
+    self._eval_dataset = None
+    self._eval_column = None
+    self._eval_template = None
+    self._eval_split = None
+    self._eval_steps = None
+    self._eval_metric_name = None
+    self._metric_for_best_model = None
+    self._input_masking = None
+    self._max_grad_norm = None
+    self._logger_level = None
+    self._benchmark_out_file = None
+    self._tuning_data_stats_file = None
+    self._enable_peft = None
+    self._merge_model_precision_mode = None
+    self._target_modules = None
+    self._unnamed_args = None
+
+  @property
+  def config_file(self):
+    return self._config_file
+
+  @config_file.setter
+  def config_file(self, val: str):
+    self._config_file = val
+
+  @property
+  def task(self):
+    return self._task
+
+  @task.setter
+  def task(self, val: str):
+    self._task = val
+
+  @property
+  def gcs_rsync_interval_secs(self):
+    return self._gcs_rsync_interval_secs
+
+  @gcs_rsync_interval_secs.setter
+  def gcs_rsync_interval_secs(self, val: str):
+    self._gcs_rsync_interval_secs = val
+
+  @property
+  def pretrained_model_name_or_path(self):
+    return self._pretrained_model_name_or_path
+
+  @pretrained_model_name_or_path.setter
+  def pretrained_model_name_or_path(self, val: str):
+    self._pretrained_model_name_or_path = val
+
+  @property
+  def train_dataset(self):
+    return self._train_dataset
+
+  @train_dataset.setter
+  def train_dataset(self, val: str):
+    self._train_dataset = val
+
+  @property
+  def train_split(self):
+    return self._train_split
+
+  @train_split.setter
+  def train_split(self, val: str):
+    self._train_split = val
+
+  @property
+  def train_template(self):
+    return self._train_template
+
+  @train_template.setter
+  def train_template(self, val: str):
+    self._train_template = val
+
+  @property
+  def train_column(self):
+    return self._train_column
+
+  @train_column.setter
+  def train_column(self, val: str):
+    self._train_column = val
+
+  @property
+  def ckpt_dir(self):
+    return self._output_dir
+
+  @ckpt_dir.setter
+  def ckpt_dir(self, val: str):
+    self._output_dir = val
+
+  @property
+  def merged_model_dir(self):
+    return self._merge_base_and_lora_output_dir
+
+  @merged_model_dir.setter
+  def merged_model_dir(self, val: str):
+    self._merge_base_and_lora_output_dir = val
+
+  @property
+  def logging_dir(self):
+    return self._logging_output_dir
+
+  @logging_dir.setter
+  def logging_dir(self, val: str):
+    self._logging_output_dir = val
+
+  @property
+  def per_device_batch_size(self):
+    return self._per_device_train_batch_size
+
+  @per_device_batch_size.setter
+  def per_device_batch_size(self, val: int):
+    self._per_device_train_batch_size = val
+
+  @property
+  def gradient_accumulation_steps(self):
+    return self._gradient_accumulation_steps
+
+  @gradient_accumulation_steps.setter
+  def gradient_accumulation_steps(self, val: int):
+    self._gradient_accumulation_steps = val
+
+  @property
+  def lora_rank(self):
+    return self._lora_rank
+
+  @lora_rank.setter
+  def lora_rank(self, val: int):
+    self._lora_rank = val
+
+  @property
+  def lora_alpha(self):
+    return self._lora_alpha
+
+  @lora_alpha.setter
+  def lora_alpha(self, val: int):
+    self._lora_alpha = val
+
+  @property
+  def lora_dropout(self):
+    return self._lora_dropout
+
+  @lora_dropout.setter
+  def lora_dropout(self, val: float):
+    self._lora_dropout = val
+
+  @property
+  def max_steps(self):
+    return self._max_steps
+
+  @max_steps.setter
+  def max_steps(self, val: int):
+    self._max_steps = val
+
+  @property
+  def num_train_epochs(self):
+    return self._num_train_epochs
+
+  @num_train_epochs.setter
+  def num_train_epochs(self, val: float):
+    self._num_train_epochs = val
+
+  @property
+  def max_seq_length(self):
+    return self._max_seq_length
+
+  @max_seq_length.setter
+  def max_seq_length(self, val: int):
+    self._max_seq_length = val
+
+  @property
+  def learning_rate(self):
+    return self._learning_rate
+
+  @learning_rate.setter
+  def learning_rate(self, val: float):
+    self._learning_rate = val
+
+  @property
+  def lr_scheduler_type(self):
+    return self._lr_scheduler_type
+
+  @lr_scheduler_type.setter
+  def lr_scheduler_type(self, val: str):
+    self._lr_scheduler_type = val
+
+  @property
+  def load_precision(self):
+    return self._precision_mode
+
+  @load_precision.setter
+  def load_precision(self, val: str):
+    self._precision_mode = val
+
+  @property
+  def train_precision(self):
+    return self._train_precision
+
+  @train_precision.setter
+  def train_precision(self, val: str):
+    self._train_precision = val
+
+  @property
+  def gradient_checkpointing(self):
+    return self._gradient_checkpointing
+
+  @gradient_checkpointing.setter
+  def gradient_checkpointing(self, val: bool):
+    self._gradient_checkpointing = val
+
+  @property
+  def example_packing(self):
+    return self._example_packing
+
+  @example_packing.setter
+  def example_packing(self, val: bool):
+    self._example_packing = val
+
+  @property
+  def attn_implementation(self):
+    return self._attn_implementation
+
+  @attn_implementation.setter
+  def attn_implementation(self, val: str):
+    self._attn_implementation = val
+
+  @property
+  def optimizer(self):
+    return self._optimizer
+
+  @optimizer.setter
+  def optimizer(self, val: str):
+    self._optimizer = val
+
+  @property
+  def warmup_ratio(self):
+    return self._warmup_ratio
+
+  @warmup_ratio.setter
+  def warmup_ratio(self, val: float):
+    self._warmup_ratio = val
+
+  @property
+  def report_to(self):
+    return self._report_to
+
+  @report_to.setter
+  def report_to(self, val: str):
+    self._report_to = val
+
+  @property
+  def save_steps(self):
+    return self._save_steps
+
+  @save_steps.setter
+  def save_steps(self, val: int):
+    self._save_steps = val
+
+  @property
+  def logging_steps(self):
+    return self._logging_steps
+
+  @logging_steps.setter
+  def logging_steps(self, val: int):
+    self._logging_steps = val
+
+  @property
+  def huggingface_access_token(self):
+    return self._huggingface_access_token
+
+  @huggingface_access_token.setter
+  def huggingface_access_token(self, val: str):
+    self._huggingface_access_token = val
+
+  @property
+  def eval_dataset(self):
+    return self._eval_dataset
+
+  @eval_dataset.setter
+  def eval_dataset(self, val: str):
+    self._eval_dataset = val
+
+  @property
+  def eval_column(self):
+    return self._eval_column
+
+  @eval_column.setter
+  def eval_column(self, val: str):
+    self._eval_column = val
+
+  @property
+  def eval_template(self):
+    return self._eval_template
+
+  @eval_template.setter
+  def eval_template(self, val: str):
+    self._eval_template = val
+
+  @property
+  def eval_split(self):
+    return self._eval_split
+
+  @eval_split.setter
+  def eval_split(self, val: str):
+    self._eval_split = val
+
+  @property
+  def eval_steps(self):
+    return self._eval_steps
+
+  @eval_steps.setter
+  def eval_steps(self, val: int):
+    self._eval_steps = val
+
+  @property
+  def eval_metric_name(self):
+    return self._eval_metric_name
+
+  @eval_metric_name.setter
+  def eval_metric_name(self, val: str):
+    self._eval_metric_name = val
+
+  @property
+  def metric_for_best_model(self):
+    return self._metric_for_best_model
+
+  @metric_for_best_model.setter
+  def metric_for_best_model(self, val: str):
+    self._metric_for_best_model = val
+
+  @property
+  def input_masking(self):
+    return self._input_masking
+
+  @input_masking.setter
+  def input_masking(self, val: bool):
+    self._input_masking = val
+
+  @property
+  def max_grad_norm(self):
+    return self._max_grad_norm
+
+  @max_grad_norm.setter
+  def max_grad_norm(self, val: float):
+    self._max_grad_norm = val
+
+  @property
+  def logger_level(self):
+    return self._logger_level
+
+  @logger_level.setter
+  def logger_level(self, val: str):
+    self._logger_level = val
+
+  @property
+  def benchmark_out_file(self):
+    return self._benchmark_out_file
+
+  @benchmark_out_file.setter
+  def benchmark_out_file(self, val: str):
+    self._benchmark_out_file = val
+
+  @property
+  def tuning_data_stats_file(self):
+    return self._tuning_data_stats_file
+
+  @tuning_data_stats_file.setter
+  def tuning_data_stats_file(self, val: str):
+    self._tuning_data_stats_file = val
+
+  @property
+  def enable_peft(self):
+    return self._enable_peft
+
+  @enable_peft.setter
+  def enable_peft(self, val: bool):
+    self._enable_peft = val
+
+  @property
+  def merge_model_precision_mode(self):
+    return self._merge_model_precision_mode
+
+  @merge_model_precision_mode.setter
+  def merge_model_precision_mode(self, val: str):
+    self._merge_model_precision_mode = val
+
+  @property
+  def target_modules(self):
+    return self._target_modules
+
+  @target_modules.setter
+  def target_modules(self, val: str):
+    self._target_modules = val
+
+  @property
+  def unnamed_args(self):
+    return self._unnamed_args
+
+  @unnamed_args.setter
+  def unnamed_args(self, val: list):
+    self._unnamed_args = val
+
+  def build_cmd(self) -> list[str]:
+    cmd = []
+    args = ''
+    for k, v in self.__dict__.items():
+      if k == '_unnamed_args' and v is not None:
+        args += ' '.join(v)
+        continue
+      if v is not None:
+        cmd.append(f'--{k[1:]}={v}')
+    cmd.append(f'{args}')
+    return cmd
@@ -0,0 +1,142 @@
+# pylint: disable=W,C,R
+
+# DO NOT MODIFY: this file is auto-generated
+# See go/vmg-oss-peft-tests#command-builder-genpy
+
+
+class QuantizeModelCommandBuilder:
+
+  def __init__(self):
+    self._task = None
+    self._pretrained_model_name_or_path = None
+    self._quantization_method = None
+    self._quantization_precision_mode = None
+    self._quantization_dataset_name = None
+    self._text_column_in_quantization_dataset = None
+    self._quantization_output_dir = None
+    self._device_map = None
+    self._max_memory = None
+    self._group_size = None
+    self._desc_act = None
+    self._damp_percent = None
+    self._cache_examples_on_gpu = None
+    self._awq_version = None
+
+  @property
+  def task(self):
+    return self._task
+
+  @task.setter
+  def task(self, val: str):
+    self._task = val
+
+  @property
+  def pretrained_model_name_or_path(self):
+    return self._pretrained_model_name_or_path
+
+  @pretrained_model_name_or_path.setter
+  def pretrained_model_name_or_path(self, val: str):
+    self._pretrained_model_name_or_path = val
+
+  @property
+  def quantization_method(self):
+    return self._quantization_method
+
+  @quantization_method.setter
+  def quantization_method(self, val: str):
+    self._quantization_method = val
+
+  @property
+  def quantization_precision_mode(self):
+    return self._quantization_precision_mode
+
+  @quantization_precision_mode.setter
+  def quantization_precision_mode(self, val: str):
+    self._quantization_precision_mode = val
+
+  @property
+  def quantization_dataset_name(self):
+    return self._quantization_dataset_name
+
+  @quantization_dataset_name.setter
+  def quantization_dataset_name(self, val: str):
+    self._quantization_dataset_name = val
+
+  @property
+  def text_column_in_quantization_dataset(self):
+    return self._text_column_in_quantization_dataset
+
+  @text_column_in_quantization_dataset.setter
+  def text_column_in_quantization_dataset(self, val: str):
+    self._text_column_in_quantization_dataset = val
+
+  @property
+  def quantization_output_dir(self):
+    return self._quantization_output_dir
+
+  @quantization_output_dir.setter
+  def quantization_output_dir(self, val: str):
+    self._quantization_output_dir = val
+
+  @property
+  def device_map(self):
+    return self._device_map
+
+  @device_map.setter
+  def device_map(self, val: str):
+    self._device_map = val
+
+  @property
+  def max_memory(self):
+    return self._max_memory
+
+  @max_memory.setter
+  def max_memory(self, val: str):
+    self._max_memory = val
+
+  @property
+  def group_size(self):
+    return self._group_size
+
+  @group_size.setter
+  def group_size(self, val: int):
+    self._group_size = val
+
+  @property
+  def desc_act(self):
+    return self._desc_act
+
+  @desc_act.setter
+  def desc_act(self, val: bool):
+    self._desc_act = val
+
+  @property
+  def damp_percent(self):
+    return self._damp_percent
+
+  @damp_percent.setter
+  def damp_percent(self, val: float):
+    self._damp_percent = val
+
+  @property
+  def cache_examples_on_gpu(self):
+    return self._cache_examples_on_gpu
+
+  @cache_examples_on_gpu.setter
+  def cache_examples_on_gpu(self, val: bool):
+    self._cache_examples_on_gpu = val
+
+  @property
+  def awq_version(self):
+    return self._awq_version
+
+  @awq_version.setter
+  def awq_version(self, val: str):
+    self._awq_version = val
+
+  def build_cmd(self) -> str:
+    cmd = []
+    for k, v in self.__dict__.items():
+      if v is not None:
+        cmd.append(f'--{k[1:]}={v}')
+    return cmd
@@ -0,0 +1,61 @@
+"""Run the tests from docker command line."""
+
+import subprocess
+import sys
+from typing import Sequence
+
+from absl import app
+from absl import flags
+
+
+_ALLOWED_TEST_FILE_PATHS = (
+    "test_instruct_lora_adapters",
+    "test_instruct_lora_features",
+    "test_instruct_lora_throughput",
+    "test_instruct_lora_trained_model_quality",
+    "test_validate_dataset_with_template",
+)
+
+_TEST_FILE_PATH = flags.DEFINE_multi_enum(
+    "test_file_path",
+    None,
+    _ALLOWED_TEST_FILE_PATHS + ("all",),
+    "The test file path.",
+    required=True,
+)
+
+_IS_AUTOMATED_TEST = flags.DEFINE_bool(
+    "is_automated_test",
+    True,
+    "Whether the test is an automated test.",
+)
+
+
+def main(argv: Sequence[str]) -> None:
+  if len(argv) > 1:
+    raise app.UsageError("Too many command-line arguments.")
+  test_file_path = _TEST_FILE_PATH.value
+  if "all" in test_file_path:
+    test_file_path = _ALLOWED_TEST_FILE_PATHS
+
+  for test_file in test_file_path:
+    cmd = [
+        "python3",
+        f"vertex_vision_model_garden_peft/tests/{test_file}.py",
+    ]
+    if (
+        test_file == "test_instruct_lora_throughput"
+        and _IS_AUTOMATED_TEST.value
+    ):
+      subprocess.run(
+          cmd + ["--", "-k", "peft_train_image_automated_test"],
+          stdout=sys.stdout,
+          stderr=sys.stdout,
+          check=True,
+      )
+    else:
+      subprocess.run(cmd, stdout=sys.stdout, stderr=sys.stdout, check=True)
+
+
+if __name__ == "__main__":
+  app.run(main)
@@ -0,0 +1,106 @@
+# pylint: disable=missing-function-docstring
+# pylint: disable=missing-class-docstring
+"""Tests adapters of PEFT train docker."""
+
+import inspect
+import os
+import time
+
+from absl.testing import absltest
+from absl.testing import parameterized
+import instruct_lora_command_builder as task_cmd_builder
+from safetensors import safe_open
+import test_util
+
+
+class AdapterTest(test_util.TestBase):
+
+  # Needs to be accessible outside docker to check artifacts.
+  _TEST_OUTPUT_DIR = os.path.expanduser('~/output')
+  _MODULES_NEED_TO_BE_EXCLUDED_IN_ADAPTER = ['lm_head', 'embed_tokens']
+
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    cls.test_suite_output_dir = os.path.join(
+        cls._TEST_OUTPUT_DIR,
+        os.path.splitext(os.path.basename(__file__))[0],
+        cls.__class__.__name__,
+    )
+
+  def setUp(self):
+    super().setUp()
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'peft_train_sample.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'input_text'
+    self.task_cmd_builder.train_template = 'llama3-text-bison'
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.max_steps = 1
+    self.task_cmd_builder.max_seq_length = 256
+    self.task_cmd_builder.load_precision = '4bit'
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.save_steps = 10
+    self.task_cmd_builder.max_steps = 3
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+
+  def setup_output_dir(self, testcase_name: str):
+    testcase_output_dir = os.path.join(
+        self.test_suite_output_dir, testcase_name, test_util.get_timestamp()
+    )
+    self.task_cmd_builder.ckpt_dir = os.path.join(
+        testcase_output_dir, 'adapter'
+    )
+    self.task_cmd_builder.logging_dir = os.path.join(
+        testcase_output_dir, 'logs'
+    )
+
+  def check_adapter_for_bad_modules(self, adapter_path):
+    unwanted_modules = set()
+    with safe_open(adapter_path, framework='pt', device='cpu') as f:
+      for key in f.keys():
+        for module in self._MODULES_NEED_TO_BE_EXCLUDED_IN_ADAPTER:
+          if module in key:
+            unwanted_modules.add(key)
+    assert (
+        not unwanted_modules
+    ), f'Adapter includes unwanted modules: {unwanted_modules}'
+
+  @parameterized.named_parameters(
+      ('llama3.1-8b', 'llama3.1-8b-hf'),
+      ('llama3.1-70b', 'llama3.1-70b-hf'),
+      ('llama2-7b', 'llama2-7b-hf'),
+  )
+  def test_llama_adapters(self, model_name):
+    test_function_name = inspect.stack()[0][3]
+    self.setup_output_dir(f'{test_function_name}-{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 1000)
+
+    adapter = os.path.join(
+        self.task_cmd_builder.ckpt_dir,
+        'checkpoint-final/adapter_model.safetensors',
+    )
+    self.check_adapter_for_bad_modules(adapter)
+
+
+if __name__ == '__main__':
+  absltest.main()
@@ -0,0 +1,367 @@
+# pylint: disable=missing-function-docstring
+# pylint: disable=missing-class-docstring
+"""Tests various features of PEFT train docker."""
+
+import os
+import time
+import unittest
+from absl.testing import absltest
+from absl.testing import parameterized
+import instruct_lora_command_builder as task_cmd_builder
+import test_util
+
+
+class EvalConfigTest(test_util.TestBase):
+
+  def setUp(self):
+    super().setUp()
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.learning_rate = 5e-5
+    self.task_cmd_builder.warmup_ratio = 0.01
+    self.task_cmd_builder.max_steps = 10
+    self.task_cmd_builder.save_steps = 1000
+    self.task_cmd_builder.logging_steps = 1
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.example_packing = True
+    self.task_cmd_builder.train_dataset = 'mlabonne/guanaco-llama2'
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'text'
+    self.task_cmd_builder.train_template = 'openassistant-guanaco'
+    self.task_cmd_builder.ckpt_dir = '/tmp/adapter'
+    self.task_cmd_builder.logging_dir = '/tmp/logs'
+    self.task_cmd_builder.eval_steps = 10
+    self.task_cmd_builder.eval_dataset = 'mlabonne/guanaco-llama2'
+    self.task_cmd_builder.eval_split = 'test'
+    self.task_cmd_builder.eval_column = 'text'
+    self.task_cmd_builder.eval_template = 'openassistant-guanaco'
+
+  @parameterized.named_parameters(
+      ('all_eval_metric', 'loss,perplexity,bleu,google_bleu,rouge1', 0),
+      ('invalid_metric', 'invalid_metric', 1),
+      ('only_loss', 'loss', 0),
+      ('perplexity_without_loss', 'perplexity,bleu', 0),
+      ('unsupported_eval_metric', 'f1', 1),
+  )
+  def test_hf_eval_metrics(self, eval_metric_name, expected_return_code):
+    self.task_cmd_builder.eval_metric_name = eval_metric_name
+    self.assertEqual(self.run_cmd(), expected_return_code)
+
+  @parameterized.named_parameters(
+      ('valid_best_model_metric', 'loss,perplexity', 'perplexity', 0),
+      ('only_loss', None, 'loss', 0),
+      ('invalid_best_model_metric', 'loss', 'invalid_metric', 1),
+  )
+  def test_metric_for_best_model(
+      self, eval_metric_name, metric_for_best_model, expected_return_code
+  ):
+    self.task_cmd_builder.eval_metric_name = eval_metric_name
+    self.task_cmd_builder.metric_for_best_model = metric_for_best_model
+    self.assertEqual(self.run_cmd(), expected_return_code)
+
+
+class GcsUploadDownloadTest(test_util.TestBase):
+
+  def setUp(self):
+    super().setUp()
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0')
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'peft_train_sample.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'input_text'
+    self.task_cmd_builder.train_template = 'llama3-text-bison'
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.max_steps = 1
+    self.task_cmd_builder.max_seq_length = 256
+    self.task_cmd_builder.load_precision = '4bit'
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.ckpt_dir = '/tmp'
+
+  @parameterized.named_parameters(
+      (
+          'llama3_8b_gcs',
+          'gs://vertex-model-garden-public-us/llama3/llama3-8b-hf',
+      ),
+      ('llama2_7b_hf', 'NousResearch/Llama-2-7b-hf'),
+  )
+  def test_model_download_single_process(self, pretrained_model_name_or_path):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(
+            pretrained_model_name_or_path
+        )
+    )
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 5 * 60.0)
+
+  @parameterized.named_parameters(
+      (
+          'llama3_8b_gcs',
+          'gs://vertex-model-garden-public-us/llama3/llama3-8b-hf',
+      ),
+      ('llama2_7b_hf', 'NousResearch/Llama-2-7b-hf'),
+  )
+  def test_model_download_multi_process(self, pretrained_model_name_or_path):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(
+            pretrained_model_name_or_path
+        )
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 5 * 60.0)
+
+  def test_8b_model_download(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        'gs://vertex-model-garden-public-us/llama3/llama3-8b-hf'
+    )
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 10 * 60.0)
+
+  @parameterized.named_parameters(
+      ('merged-without-upload', '/tmp/merged'),
+      ('merged-and-upload-to-gcs', 'gs://vmg-test-ttl-1y/tests/merged'),
+  )
+  def test_model_merge(self, output_dir):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+
+    ckpt_dir = os.path.join(
+        output_dir,
+        f'output-{test_util.get_timestamp()}',
+    )
+    self.task_cmd_builder.ckpt_dir = ckpt_dir
+    self.task_cmd_builder.merged_model_dir = os.path.join(ckpt_dir, 'merged')
+    self.task_cmd_builder.logging_dir = '/tmp/logging'
+
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 5 * 60.0)
+
+  @unittest.skipIf(
+      not test_util.is_gpu_h100(),
+      'Skipping because this test is only for H100',
+  )
+  def test_model_fp8_conversion(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+
+    ckpt_dir = f'/tmp/output/output-{test_util.get_timestamp()}'
+    self.task_cmd_builder.ckpt_dir = ckpt_dir
+    self.task_cmd_builder.merged_model_dir = os.path.join(ckpt_dir, 'merged')
+    self.task_cmd_builder.logging_dir = os.path.join(ckpt_dir, 'logging')
+    self.task_cmd_builder.merge_model_precision_mode = 'float8'
+
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 5 * 60.0)
+
+  @parameterized.named_parameters(
+      ('merged-without-upload', '/tmp/merged'),
+      ('merged-and-upload-to-gcs', 'gs://vmg-test-ttl-1y/tests/merged'),
+  )
+  def test_model_merge_and_upload_deepspeed(self, merged_model_dir):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/deepspeed_zero2_8gpu.yaml'
+    )
+    self.task_cmd_builder.merged_model_dir = os.path.join(
+        merged_model_dir, f'merged-{test_util.get_timestamp()}'
+    )
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 5 * 60.0)
+
+  @parameterized.named_parameters(
+      ('save-only-last', 10),
+      ('save-multiple-times', 1),
+  )
+  def test_llama3_8b_save_and_merge_8_gpus_fsdp(self, save_steps):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.save_steps = save_steps
+    self.task_cmd_builder.max_steps = 3
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.merged_model_dir = '/tmp/merged'
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+    start_time = time.time()
+    self.assertEqual(self.run_cmd(), 0)
+    end_time = time.time()
+    self.assertLess(end_time - start_time, 9 * 60.0)
+
+
+class TemplateAndDataStatsTest(test_util.TestBase):
+
+  def setUp(self):
+    super().setUp()
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0')
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.max_steps = 1
+    self.task_cmd_builder.max_seq_length = 256
+    self.task_cmd_builder.load_precision = '4bit'
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.ckpt_dir = '/tmp'
+
+  @parameterized.named_parameters(
+      ('multi-chat-string-content', 'openai-multi-chat-example-data.jsonl'),
+      (
+          'multi-chat-array-content',
+          'openai-multi-chat-example-data-array-content.jsonl',
+      ),
+  )
+  def test_openai_chat_template(self, example_dataset):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        example_dataset
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'messages'
+    self.task_cmd_builder.train_template = 'openai-chat'
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  def test_openai_completion_template(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'openai-completion-example-data.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'prompt'
+    self.task_cmd_builder.train_template = 'openai-completion'
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  def test_data_stats_chat_template(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'openai-multi-chat-example-data.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'messages'
+    self.task_cmd_builder.train_template = 'llama3'
+    self.task_cmd_builder.tuning_data_stats_file = '/tmp/data-stats.json'
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  def test_data_stats_completion_template(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'openai-completion-example-data.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'prompt'
+    self.task_cmd_builder.train_template = 'openai-completion'
+    self.task_cmd_builder.tuning_data_stats_file = '/tmp/data-stats.json'
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+    self.assertEqual(self.run_cmd(), 0)
+
+
+class TargetModulesTest(test_util.TestBase):
+
+  def setUp(self):
+    super().setUp()
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0')
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'peft_train_sample.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'input_text'
+    self.task_cmd_builder.train_template = 'llama3-text-bison'
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.max_steps = 1
+    self.task_cmd_builder.max_seq_length = 256
+    self.task_cmd_builder.load_precision = '4bit'
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.ckpt_dir = '/tmp'
+
+  def test_target_modules(self):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path('llama3.1-8b-hf')
+    )
+    self.task_cmd_builder.target_modules = 'q_proj, v_proj, k_proj'
+
+    self.assertEqual(self.run_cmd(), 0)
+
+
+if __name__ == '__main__':
+  absltest.main()
@@ -0,0 +1,276 @@
+# pylint: disable=missing-function-docstring
+# pylint: disable=missing-class-docstring
+"""Tests to check training throughput and GPU memory consumption."""
+
+import os
+import pathlib
+
+from absl.testing import absltest
+from absl.testing import parameterized
+import instruct_lora_command_builder as task_cmd_builder
+import test_util
+
+
+class TrainerThroughputTest(test_util.TestBase):
+
+  _TEST_OUTPUT_DIR = os.path.expanduser('~/throughput_tests')
+
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    cls.test_suite_output_dir = os.path.join(
+        cls._TEST_OUTPUT_DIR, os.path.splitext(os.path.basename(__file__))[0]
+    )
+    if not os.path.isdir(cls.test_suite_output_dir):
+      pathlib.Path(cls.test_suite_output_dir).mkdir(parents=True)
+
+  def setUp(self):
+    super().setUp()
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.learning_rate = 5e-5
+    self.task_cmd_builder.warmup_ratio = 0.01
+    self.task_cmd_builder.max_steps = 10
+    self.task_cmd_builder.save_steps = 1000
+    self.task_cmd_builder.logging_steps = 1
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.example_packing = True
+    self.task_cmd_builder.train_dataset = 'mlabonne/guanaco-llama2-1k'
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'text'
+    self.task_cmd_builder.train_template = 'openassistant-guanaco'
+    self.task_cmd_builder.ckpt_dir = '/tmp/adapter'
+    self.task_cmd_builder.logging_dir = '/tmp/logs'
+
+  def run_cmd_and_handle_failure(self):
+    ret = self.run_cmd()
+    if ret != 0:
+      with open(self.task_cmd_builder.benchmark_out_file, 'a') as f:
+        max_seq_length = self.task_cmd_builder.max_seq_length
+        f.write(f'{max_seq_length/1024.0:.1f} | failed | n/a\n')
+    return ret
+
+  @parameterized.product(
+      model_name=[
+          'llama3.1-8b-hf',
+          'llama3.1-70b-hf',
+          'Mistral-7B-v0.1',
+          'Mixtral-8x7B-v0.1',
+          'gemma-2-9b-it',
+          'Qwen2.5-32B-Instruct',
+      ],
+      precision=['4bit', '8bit', 'bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+  )
+  def test_model_single_gpu(self, model_name, precision, max_seq_length):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    self.task_cmd_builder.benchmark_out_file = os.path.join(
+        self.test_suite_output_dir, f'bm_{model_name}_{precision}.txt'
+    )
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0')
+
+    self.assertEqual(self.run_cmd_and_handle_failure(), 0)
+
+  @parameterized.product(
+      model_name=[
+          'llama3.1-8b-hf',
+          'llama3.1-70b-hf',
+          'Mistral-7B-v0.1',
+          'Mixtral-8x7B-v0.1',
+          'gemma-2-9b-it',
+          'Qwen2.5-32B-Instruct',
+      ],
+      precision=['4bit', '8bit', 'bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+      num_gpus=[8],
+      config=['deepspeed_zero2'],
+  )
+  def test_model_multi_gpu_deepspeed(
+      self, model_name, precision, max_seq_length, num_gpus, config
+  ):
+    self.assertTrue(num_gpus == 4 or num_gpus == 8)
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    self.task_cmd_builder.benchmark_out_file = os.path.join(
+        self.test_suite_output_dir,
+        f'bm_{config}_{num_gpus}gpu_{model_name}_{precision}.txt',
+    )
+
+    self.task_cmd_builder.config_file = (
+        f'vertex_vision_model_garden_peft/{config}_{num_gpus}gpu.yaml'
+    )
+    self.command_builder.add_env_var(
+        'CUDA_VISIBLE_DEVICES', ','.join([str(x) for x in range(0, num_gpus)])
+    )
+
+    self.assertEqual(self.run_cmd_and_handle_failure(), 0)
+
+  @parameterized.product(
+      model_name=[
+          'llama3.1-8b-hf',
+          'llama3.1-70b-hf',
+          'Qwen2.5-32B-Instruct',
+      ],
+      precision=['4bit', '8bit', 'bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+      num_gpus=[8],
+  )
+  def test_model_multi_gpu_fsdp_lora(
+      self, model_name, precision, max_seq_length, num_gpus
+  ):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    self.task_cmd_builder.benchmark_out_file = os.path.join(
+        self.test_suite_output_dir,
+        f'bm_fsdp_{num_gpus}gpu_{model_name}_{precision}.txt',
+    )
+    if 'llama' in model_name.lower():
+      self.task_cmd_builder.config_file = (
+          'vertex_vision_model_garden_peft/llama2_fsdp_8gpu.yaml'
+      )
+    elif 'qwen' in model_name.lower():
+      self.task_cmd_builder.config_file = (
+          'vertex_vision_model_garden_peft/qwen2_fsdp_8gpu.yaml'
+      )
+    else:
+      self.fail(f'Unsupported model: {model_name}')
+
+    self.command_builder.add_env_var(
+        'CUDA_VISIBLE_DEVICES', ','.join([str(x) for x in range(0, num_gpus)])
+    )
+
+    self.assertEqual(self.run_cmd_and_handle_failure(), 0)
+
+  @parameterized.product(
+      model_name=['llama3.1-8b-hf', 'llama3.1-70b-hf'],
+      precision=['4bit', 'bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+      config=['deepspeed_zero2', 'fsdp'],
+  )
+  def test_peft_train_image_automated_test_llama(
+      self, model_name, precision, max_seq_length, config
+  ):
+    num_gpus = 8
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    benchmark_out_file = os.path.join(
+        self.test_suite_output_dir,
+        f'bm_{config}_{num_gpus}gpu_{model_name}_{precision}.txt',
+    )
+    self.task_cmd_builder.benchmark_out_file = benchmark_out_file
+    if config == 'fsdp':
+      self.task_cmd_builder.config_file = (
+          f'vertex_vision_model_garden_peft/llama_{config}_{num_gpus}gpu.yaml'
+      )
+    else:
+      self.task_cmd_builder.config_file = (
+          f'vertex_vision_model_garden_peft/{config}_{num_gpus}gpu.yaml'
+      )
+
+    self.command_builder.add_env_var(
+        'CUDA_VISIBLE_DEVICES', ','.join([str(x) for x in range(0, num_gpus)])
+    )
+    self.run_cmd_and_handle_failure()
+    if test_util.is_gpu_h100():
+      self.assertEqual(
+          test_util.check_benchmark_results(
+              benchmark_out_file, 'llama', 10.0, max_seq_length
+          ),
+          True,
+      )
+
+  @parameterized.product(
+      model_name=['gemma-2-2b-it', 'gemma-2-9b-it', 'gemma-2-27b-it'],
+      precision=['4bit', 'bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+      config=['deepspeed_zero2', 'deepspeed_zero3', 'fsdp'],
+  )
+  def test_peft_train_image_automated_test_gemma(
+      self, model_name, precision, max_seq_length, config
+  ):
+    num_gpus = 8
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.attn_implementation = 'eager'
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    benchmark_out_file = os.path.join(
+        self.test_suite_output_dir,
+        f'bm_{config}_{num_gpus}gpu_{model_name}_{precision}.txt',
+    )
+    self.task_cmd_builder.benchmark_out_file = benchmark_out_file
+    if config == 'fsdp':
+      self.task_cmd_builder.config_file = (
+          f'vertex_vision_model_garden_peft/gemma2_{config}_{num_gpus}gpu.yaml'
+      )
+    else:
+      self.task_cmd_builder.config_file = (
+          f'vertex_vision_model_garden_peft/{config}_{num_gpus}gpu.yaml'
+      )
+
+    self.command_builder.add_env_var(
+        'CUDA_VISIBLE_DEVICES', ','.join([str(x) for x in range(0, num_gpus)])
+    )
+    self.run_cmd_and_handle_failure()
+    if test_util.is_gpu_h100():
+      self.assertEqual(
+          test_util.check_benchmark_results(
+              benchmark_out_file, 'gemma', 10.0, max_seq_length
+          ),
+          True,
+      )
+
+  @parameterized.product(
+      model_name=['llama3.1-8b-hf', 'llama3.1-70b-hf'],
+      precision=['bfloat16'],
+      max_seq_length=list(range(4 * 1024, 24 * 1024 + 1, 4 * 1024)),
+      num_gpus=[8],
+  )
+  def test_model_multi_gpu_fsdp_full_finetuning(
+      self, model_name, precision, max_seq_length, num_gpus
+  ):
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.load_precision = precision
+    self.task_cmd_builder.benchmark_out_file = os.path.join(
+        self.test_suite_output_dir,
+        f'bm_fsdp_full_finetuning_{num_gpus}gpu_{model_name}_{precision}.txt',
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.enable_peft = False
+
+    self.command_builder.add_env_var(
+        'CUDA_VISIBLE_DEVICES', ','.join([str(x) for x in range(0, num_gpus)])
+    )
+
+    self.assertEqual(self.run_cmd_and_handle_failure(), 0)
+
+
+if __name__ == '__main__':
+  absltest.main()
@@ -0,0 +1,209 @@
+# pylint: disable=missing-function-docstring
+# pylint: disable=missing-class-docstring
+"""Tests to make sure trained model achieves decent quality.
+
+Right now, the metric is loss decreasing and we'll eyeball the TB graphs.
+"""
+
+import os
+
+from absl.testing import absltest
+from absl.testing import parameterized
+import instruct_lora_command_builder as task_cmd_builder
+import test_util
+
+
+class TrainedModelQualityTest(test_util.TestBase):
+
+  _TEST_OUTPUT_DIR = os.path.expanduser('~/output')
+
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    cls.test_suite_output_dir = os.path.join(
+        cls._TEST_OUTPUT_DIR,
+        os.path.splitext(os.path.basename(__file__))[0],
+        cls.__class__.__name__,
+    )
+
+  def setUp(self):
+    super().setUp()
+
+    self.task_cmd_builder = task_cmd_builder.InstructLoraCommandBuilder()
+    self.task_cmd_builder.task = 'instruct-lora'
+    self.task_cmd_builder.eval_metric_name = 'loss'
+    self.task_cmd_builder.per_device_batch_size = 1
+    self.task_cmd_builder.gradient_accumulation_steps = 8
+    self.task_cmd_builder.lora_rank = 16
+    self.task_cmd_builder.lora_alpha = 32
+    self.task_cmd_builder.lora_dropout = 0.05
+    self.task_cmd_builder.learning_rate = 5e-5
+    self.task_cmd_builder.num_train_epochs = 2.0
+    self.task_cmd_builder.warmup_ratio = 0.01
+    self.task_cmd_builder.max_steps = -1
+    self.task_cmd_builder.save_steps = 10
+    self.task_cmd_builder.eval_steps = 10
+    self.task_cmd_builder.max_seq_length = 4096
+    self.task_cmd_builder.load_precision = '4bit'
+    self.task_cmd_builder.gradient_checkpointing = True
+    self.task_cmd_builder.input_masking = True
+    self.task_cmd_builder.attn_implementation = 'flash_attention_2'
+    self.task_cmd_builder.report_to = 'tensorboard'
+
+  def setup_output_dir(self, testcase_name: str):
+    testcase_output_dir = os.path.join(
+        self.test_suite_output_dir, testcase_name
+    )
+    self.task_cmd_builder.ckpt_dir = os.path.join(
+        testcase_output_dir, 'adapter'
+    )
+    self.task_cmd_builder.logging_dir = os.path.join(
+        testcase_output_dir, 'logs'
+    )
+    self.task_cmd_builder.merged_model_dir = os.path.join(
+        testcase_output_dir, 'merged'
+    )
+
+  @parameterized.named_parameters(
+      ('llama3-8b', 'llama3-8b-hf'),
+      ('llama3.1-8b', 'llama3.1-8b-hf'),
+  )
+  def test_8b_model_deepspeed(self, model_name):
+    self.setup_output_dir(f'test_deepspeed_{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'peft_train_sample.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'input_text'
+    self.task_cmd_builder.train_template = 'llama3-text-bison'
+    self.task_cmd_builder.eval_dataset = test_util.get_test_data_path(
+        'peft_eval_sample.jsonl'
+    )
+    self.task_cmd_builder.eval_split = 'train'
+    self.task_cmd_builder.eval_column = self.task_cmd_builder.train_column
+    self.task_cmd_builder.eval_template = self.task_cmd_builder.train_template
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0')
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  @parameterized.named_parameters(
+      ('llama3-70b', 'llama3-70b-hf'),
+      ('llama3.1-70b', 'llama3.1-70b-hf'),
+  )
+  def test_70b_model_deepspeed(self, model_name):
+    self.setup_output_dir(f'test_deepspeed_{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/deepspeed_zero2_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = 'timdettmers/openassistant-guanaco'
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'text'
+    self.task_cmd_builder.train_template = 'openassistant-guanaco'
+    self.task_cmd_builder.eval_dataset = self.task_cmd_builder.train_dataset
+    self.task_cmd_builder.eval_split = 'test'
+    self.task_cmd_builder.eval_column = self.task_cmd_builder.train_column
+    self.task_cmd_builder.eval_template = self.task_cmd_builder.train_template
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  @parameterized.named_parameters(
+      ('llama3-70b', 'llama3-70b-hf'),
+      ('llama3.1-70b', 'llama3.1-70b-hf'),
+  )
+  def test_70b_model_fsdp(self, model_name):
+    self.setup_output_dir(f'test_fsdp_{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/llama_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = 'timdettmers/openassistant-guanaco'
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'text'
+    self.task_cmd_builder.train_template = 'openassistant-guanaco'
+    self.task_cmd_builder.eval_dataset = self.task_cmd_builder.train_dataset
+    self.task_cmd_builder.eval_split = 'test'
+    self.task_cmd_builder.eval_column = self.task_cmd_builder.train_column
+    self.task_cmd_builder.eval_template = self.task_cmd_builder.train_template
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  @parameterized.named_parameters(
+      ('Qwen2.5-32B-Instruct', 'Qwen2.5-32B-Instruct'),
+  )
+  def test_qwen_model_deepspeed(self, model_name):
+    self.setup_output_dir(f'test_deepspeed_{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = model_name
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/deepspeed_zero2_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'llama-tuning-test/opposite-examples-train.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'messages'
+    self.task_cmd_builder.train_template = 'qwen2_5'
+    self.task_cmd_builder.eval_dataset = test_util.get_test_data_path(
+        'llama-tuning-test/opposite-examples-eval.jsonl'
+    )
+    self.task_cmd_builder.eval_split = 'train'
+    self.task_cmd_builder.eval_column = self.task_cmd_builder.train_column
+    self.task_cmd_builder.eval_template = self.task_cmd_builder.train_template
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+    # Note(lavrai): The following parameters are needed for the opposite-word
+    # dataset to converge properly.
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.num_train_epochs = 10.0
+    self.task_cmd_builder.logging_steps = 1
+
+    self.assertEqual(self.run_cmd(), 0)
+
+  @parameterized.named_parameters(
+      ('Qwen2.5-32B-Instruct', 'Qwen2.5-32B-Instruct'),
+  )
+  def test_qwen_model_fsdp(self, model_name):
+    self.setup_output_dir(f'test_fsdp_{model_name}')
+    self.task_cmd_builder.pretrained_model_name_or_path = (
+        test_util.get_pretrained_model_name_or_path(model_name)
+    )
+    self.task_cmd_builder.config_file = (
+        'vertex_vision_model_garden_peft/qwen2_fsdp_8gpu.yaml'
+    )
+    self.task_cmd_builder.train_dataset = test_util.get_test_data_path(
+        'llama-tuning-test/opposite-examples-train.jsonl'
+    )
+    self.task_cmd_builder.train_split = 'train'
+    self.task_cmd_builder.train_column = 'messages'
+    self.task_cmd_builder.train_template = 'qwen2_5'
+    self.task_cmd_builder.eval_dataset = test_util.get_test_data_path(
+        'llama-tuning-test/opposite-examples-eval.jsonl'
+    )
+    self.task_cmd_builder.eval_split = 'train'
+    self.task_cmd_builder.eval_column = self.task_cmd_builder.train_column
+    self.task_cmd_builder.eval_template = self.task_cmd_builder.train_template
+
+    self.command_builder.add_env_var('CUDA_VISIBLE_DEVICES', '0,1,2,3,4,5,6,7')
+    # Note(lavrai): The following parameters are needed for the opposite-word
+    # dataset to converge properly.
+    self.task_cmd_builder.gradient_accumulation_steps = 1
+    self.task_cmd_builder.num_train_epochs = 10.0
+    self.task_cmd_builder.logging_steps = 1
+
+    self.assertEqual(self.run_cmd(), 0)
+
+
+if __name__ == '__main__':
+  absltest.main()
@@ -0,0 +1,353 @@
+"""Tests validate the dataset with template task in PEFT docker."""
+
+from absl.testing import absltest
+from absl.testing import parameterized
+import test_util
+import validate_dataset_with_template_command_builder as task_cmd_builder
+
+
+class ValidateDatasetWithTemplateTest(test_util.TestBase):
+  """Test the validate dataset with template task in PEFT docker."""
+
+  def setUp(self):
+    super().setUp()
+
+    self.task_cmd_builder = (
+        task_cmd_builder.ValidateDatasetWithTemplateCommandBuilder()
+    )
+    self.task_cmd_builder.task = "validate-dataset-with-template"
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="valid_rows",
+          validate_top_k_rows=100,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="negative_rows",
+          validate_top_k_rows=-10,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="out_of_range_rows",
+          validate_top_k_rows=100000,
+          expected_result=0,
+      ),
+  )
+  def test_validate_dataset_with_template_top_k_rows(
+      self,
+      validate_top_k_rows,
+      expected_result,
+  ):
+    self.task_cmd_builder.dataset_name = "timdettmers/openassistant-guanaco"
+    self.task_cmd_builder.train_split = "train"
+    self.task_cmd_builder.train_column = "text"
+    self.task_cmd_builder.template = (
+        "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+    )
+    self.task_cmd_builder.validate_percentage_of_dataset = None
+    self.task_cmd_builder.validate_k_rows_of_dataset = validate_top_k_rows
+    self.task_cmd_builder.use_multiprocessing = True
+    result = self.run_cmd()
+    self.assertEqual(result, expected_result)
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="valid_positive_x_percent",
+          validate_percentage_of_dataset=10,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="valid_negative_x_percent",
+          validate_percentage_of_dataset=-10,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="invalid_positive_x_percent",
+          validate_percentage_of_dataset=110,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="invalid_negative_x_percent",
+          validate_percentage_of_dataset=-110,
+          expected_result=1,
+      ),
+  )
+  def test_validate_dataset_with_template_x_percent(
+      self,
+      validate_percentage_of_dataset,
+      expected_result,
+  ):
+    self.task_cmd_builder.dataset_name = "timdettmers/openassistant-guanaco"
+    self.task_cmd_builder.train_split = "train"
+    self.task_cmd_builder.train_column = "text"
+    self.task_cmd_builder.template = (
+        "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+    )
+    self.task_cmd_builder.validate_percentage_of_dataset = (
+        validate_percentage_of_dataset
+    )
+    self.task_cmd_builder.validate_k_rows_of_dataset = None
+    self.task_cmd_builder.use_multiprocessing = True
+    result = self.run_cmd()
+    self.assertEqual(result, expected_result)
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="small_max_seq_length",
+          max_seq_length=10,
+      ),
+      dict(
+          testcase_name="large_max_seq_length",
+          max_seq_length=1024,
+      ),
+  )
+  def test_validate_dataset_with_template_max_seq_length(
+      self,
+      max_seq_length,
+  ):
+    self.task_cmd_builder.dataset_name = "timdettmers/openassistant-guanaco"
+    self.task_cmd_builder.train_split = "train"
+    self.task_cmd_builder.train_column = "text"
+    self.task_cmd_builder.template = (
+        "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+    )
+    self.task_cmd_builder.max_seq_length = max_seq_length
+    self.task_cmd_builder.validate_k_rows_of_dataset = None
+    self.task_cmd_builder.use_multiprocessing = True
+    result = self.run_cmd()
+    self.assertEqual(result, 0)
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="invalid_default_input_column",
+          dataset_name="timdettmers/openassistant-guanaco",
+          split="train",
+          input_column="",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="invalid_percentage",
+          dataset_name="timdettmers/openassistant-guanaco",
+          split="train",
+          input_column="text",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=110,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="negative_percentage",
+          dataset_name="timdettmers/openassistant-guanaco",
+          split="train",
+          input_column="text",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=-110,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="empty_dataset",
+          dataset_name="",
+          split="train",
+          input_column="text",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="empty_split",
+          dataset_name="timdettmers/openassistant-guanaco",
+          split="",
+          input_column="text",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="empty_template",
+          dataset_name="timdettmers/openassistant-guanaco",
+          split="train",
+          input_column="text",
+          template="",
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="wrong_gcs_template",
+          dataset_name="gs://cloud-nas-260507-tmp-20240724/model-evaluation/peft_train_sample.jsonl",
+          split="train",
+          input_column="text",
+          template="gs://cloud-nas-260507-tmp-20240724/sample-template.json",
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+      dict(
+          testcase_name="wrong_gcs_dataset_name",
+          dataset_name="gs://cloud-nas-260507-tmp-20240724/model-evaluation/peft-train_sample.jsonl",
+          split="train",
+          input_column="text",
+          template="gs://cloud-nas-260507-tmp-20240724/sample_template.json",
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=1,
+      ),
+  )
+  def test_validate_dataset_with_template_invalid_input(
+      self,
+      dataset_name,
+      split,
+      input_column,
+      template,
+      validate_percentage_of_dataset,
+      validate_top_k_rows,
+      use_multiprocessing,
+      expected_result,
+  ):
+    self.task_cmd_builder.dataset_name = dataset_name
+    self.task_cmd_builder.train_split = split
+    self.task_cmd_builder.train_column = input_column
+    self.task_cmd_builder.template = template
+    self.task_cmd_builder.validate_percentage_of_dataset = (
+        validate_percentage_of_dataset
+    )
+    self.task_cmd_builder.validate_k_rows_of_dataset = validate_top_k_rows
+    self.task_cmd_builder.use_multiprocessing = use_multiprocessing
+    result = self.run_cmd()
+    self.assertEqual(result, expected_result)
+
+  @parameterized.named_parameters(
+      dict(
+          testcase_name="full_hf_dataset_with_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="full_gcs_dataset_with_multiprocessing",
+          dataset_name="gs://cloud-nas-260507-tmp-20240724/model-evaluation/peft_train_sample.jsonl",
+          template="gs://cloud-nas-260507-tmp-20240724/sample_template.json",
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="half_dataset_with_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=50,
+          validate_top_k_rows=None,
+          use_multiprocessing=True,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="top_100_rows_with_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=100,
+          use_multiprocessing=True,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="full_gcs_dataset_without_multiprocessing",
+          dataset_name="gs://cloud-nas-260507-tmp-20240724/model-evaluation/peft_train_sample.jsonl",
+          template="gs://cloud-nas-260507-tmp-20240724/sample_template.json",
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=False,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="full_hf_dataset_without_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=None,
+          use_multiprocessing=False,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="half_dataset_without_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=50,
+          validate_top_k_rows=None,
+          use_multiprocessing=False,
+          expected_result=0,
+      ),
+      dict(
+          testcase_name="top_100_rows_without_multiprocessing",
+          dataset_name="timdettmers/openassistant-guanaco",
+          template=(
+              "gs://cloud-nas-260507-tmp-20240724/openassistant-guanaco.json"
+          ),
+          validate_percentage_of_dataset=None,
+          validate_top_k_rows=100,
+          use_multiprocessing=True,
+          expected_result=0,
+      ),
+  )
+  def test_validate_dataset_with_template_multiprocessing_option(
+      self,
+      dataset_name,
+      template,
+      validate_percentage_of_dataset,
+      validate_top_k_rows,
+      use_multiprocessing,
+      expected_result,
+  ):
+    self.task_cmd_builder.dataset_name = dataset_name
+    self.task_cmd_builder.train_split = "train"
+    self.task_cmd_builder.train_column = "text"
+    self.task_cmd_builder.template = template
+    self.task_cmd_builder.validate_percentage_of_dataset = (
+        validate_percentage_of_dataset
+    )
+    self.task_cmd_builder.validate_k_rows_of_dataset = validate_top_k_rows
+    self.task_cmd_builder.use_multiprocessing = use_multiprocessing
+    result = self.run_cmd()
+    self.assertEqual(result, expected_result)
+
+
+if __name__ == "__main__":
+  absltest.main()
@@ -0,0 +1,97 @@
+# pylint: disable=W,C,R
+
+# DO NOT MODIFY: this file is auto-generated
+# See go/vmg-oss-peft-tests#command-builder-genpy
+
+
+class ValidateDatasetWithTemplateCommandBuilder:
+
+  def __init__(self):
+    self._task = None
+    self._template = None
+    self._dataset_name = None
+    self._train_split = None
+    self._train_column = None
+    self._max_seq_length = None
+    self._use_multiprocessing = None
+    self._validate_k_rows_of_dataset = None
+    self._validate_percentage_of_dataset = None
+
+  @property
+  def task(self):
+    return self._task
+
+  @task.setter
+  def task(self, val: str):
+    self._task = val
+
+  @property
+  def template(self):
+    return self._template
+
+  @template.setter
+  def template(self, val: str):
+    self._template = val
+
+  @property
+  def dataset_name(self):
+    return self._dataset_name
+
+  @dataset_name.setter
+  def dataset_name(self, val: str):
+    self._dataset_name = val
+
+  @property
+  def train_split(self):
+    return self._train_split
+
+  @train_split.setter
+  def train_split(self, val: str):
+    self._train_split = val
+
+  @property
+  def train_column(self):
+    return self._train_column
+
+  @train_column.setter
+  def train_column(self, val: str):
+    self._train_column = val
+
+  @property
+  def max_seq_length(self):
+    return self._max_seq_length
+
+  @max_seq_length.setter
+  def max_seq_length(self, val: int):
+    self._max_seq_length = val
+
+  @property
+  def use_multiprocessing(self):
+    return self._use_multiprocessing
+
+  @use_multiprocessing.setter
+  def use_multiprocessing(self, val: bool):
+    self._use_multiprocessing = val
+
+  @property
+  def validate_k_rows_of_dataset(self):
+    return self._validate_k_rows_of_dataset
+
+  @validate_k_rows_of_dataset.setter
+  def validate_k_rows_of_dataset(self, val: int):
+    self._validate_k_rows_of_dataset = val
+
+  @property
+  def validate_percentage_of_dataset(self):
+    return self._validate_percentage_of_dataset
+
+  @validate_percentage_of_dataset.setter
+  def validate_percentage_of_dataset(self, val: int):
+    self._validate_percentage_of_dataset = val
+
+  def build_cmd(self) -> str:
+    cmd = []
+    for k, v in self.__dict__.items():
+      if v is not None:
+        cmd.append(f'--{k[1:]}={v}')
+    return cmd
@@ -0,0 +1,292 @@
+"""Entrypoint for peft train docker.
+
+Dispatches to different scripts based on `task` type.
+
+For task type in `_TASK_TO_SCRIPT`, if `--config_file` is specified, the script
+will dispatch the call to `accelerate`, which is friendly for multi-GPU
+environment. Otherwise, `python3` is used.
+"""
+
+import argparse
+from collections.abc import MutableSequence, Sequence
+import multiprocessing
+import subprocess
+import sys
+from absl import app
+from absl import flags
+from absl import logging
+from util import dataset_validation_util
+from util import cluster_spec
+from vertex_vision_model_garden_peft.train.vmg import utils
+from util import constants
+from util import gcs_syncer
+from util import hypertune_utils
+
+
+_TEXT_TO_IMAGE_TASKS_SCRIPTS = {
+    constants.TEXT_TO_IMAGE: 'text_to_image/train_text_to_image.py',
+    constants.TEXT_TO_IMAGE_LORA: 'text_to_image/train_text_to_image_lora.py',
+    constants.TEXT_TO_IMAGE_DREAMBOOTH: 'dreambooth/train_dreambooth.py',
+    constants.TEXT_TO_IMAGE_DREAMBOOTH_LORA: (
+        'dreambooth/train_dreambooth_lora.py'
+    ),
+    constants.TEXT_TO_IMAGE_DREAMBOOTH_LORA_SDXL: (
+        'dreambooth/train_dreambooth_lora_sdxl.py'
+    ),
+}
+
+_TASK_TO_SCRIPT = {
+    constants.INSTRUCT_LORA: (
+        'vertex_vision_model_garden_peft/train/vmg/instruct_lora.py'
+    ),
+    constants.MERGE_CAUSAL_LANGUAGE_MODEL_LORA: (
+        'vertex_vision_model_garden_peft/train/vmg/merge_causal_language_model_lora.py'
+    ),
+    constants.VALIDATE_DATASET_WITH_TEMPLATE: (
+        'vertex_vision_model_garden_peft/train/vmg/validate_dataset_with_template.py'
+    ),
+    constants.RUN_TESTS: 'vertex_vision_model_garden_peft/tests/run_tests.py',
+}
+
+
+def launch_script_cmd(
+    script: str,
+    config_file: str | None,
+    accelerate_args: argparse.Namespace = argparse.Namespace(),
+) -> MutableSequence[str]:
+  """Returns the command to launch the script."""
+  if config_file:
+    cmd = [
+        'accelerate',
+        'launch',
+        '--config_file={}'.format(config_file),
+    ]
+  else:
+    cmd = ['python3']
+
+  _append_args_to_command_in_place(accelerate_args, cmd)
+  cmd.append(script)
+
+  return cmd
+
+
+def _get_accelerate_args() -> argparse.Namespace:
+  """Returns the accelerate args."""
+  primary_node_addr, primary_node_port, node_rank, num_nodes = (
+      cluster_spec.get_cluster_spec()
+  )
+  accelerate_args = argparse.Namespace()
+  if num_nodes > 1:
+    accelerate_args.machine_rank = node_rank
+    accelerate_args.num_machines = num_nodes
+    accelerate_args.main_process_ip = primary_node_addr
+    accelerate_args.main_process_port = primary_node_port
+    accelerate_args.max_restarts = 0
+    accelerate_args.monitor_interval = 120
+
+  return accelerate_args
+
+
+def _append_args_to_command_in_place(
+    args: argparse.Namespace, command: MutableSequence[str]
+):
+  for key, value in vars(args).items():
+    # If not specified, skip.
+    if value is not None:
+      command.append(f'--{key}={value}')
+
+
+def _get_train_and_maybe_merge_cmd_and_dirs_to_sync(
+    task_type: str, config_file: str, unknown: Sequence[str]
+) -> Sequence[Sequence[str]]:
+  """Returns the training and merge command(if applicable) and dirs to sync.
+
+  Args:
+    task_type: The task type.
+    config_file: The accelerate config file path.
+    unknown: The unknown args which are not recognised by the parser.
+
+  Returns:
+    The bash commands to execute and the directories to sync.
+  """
+  dirs_to_sync = []
+  # Only populated when multi-node is used.
+  accelerate_args = _get_accelerate_args()
+  node_rank = getattr(accelerate_args, 'machine_rank', 0)
+  training_cmd = launch_script_cmd(
+      _TASK_TO_SCRIPT[task_type],
+      config_file,
+      accelerate_args=accelerate_args,
+  )
+
+  # Training only flag.
+  train_parser = argparse.ArgumentParser()
+  train_parser.add_argument('--output_dir', required=True)
+  training_args, unknown = train_parser.parse_known_args(unknown)
+  # Checks for `hypertune_utils._ENVIRONMENT_VARIABLE_FOR_TRIAL_ID` env var and
+  # appends the trial id if it exists.
+  training_args.output_dir = hypertune_utils.maybe_append_trial_id(
+      dataset_validation_util.force_gcs_fuse_path(training_args.output_dir)
+  )
+
+  local_output_dir, gcs_output_dir = gcs_syncer.manage_sync_path(
+      training_args.output_dir, node_rank
+  )
+  training_args.output_dir = local_output_dir
+  if gcs_syncer.is_gcs_or_gcsfuse_path(gcs_output_dir):
+    dirs_to_sync.append((local_output_dir, gcs_output_dir))
+
+  # Merge only flags.
+  merge_parser = argparse.ArgumentParser()
+  merge_parser.add_argument('--merge_model_precision_mode')
+  merge_parser.add_argument('--merge_base_and_lora_output_dir')
+  merge_args, unknown = merge_parser.parse_known_args(unknown)
+
+  if merge_args.merge_base_and_lora_output_dir:
+    merge_local_dir, merge_gcs_dir = gcs_syncer.manage_sync_path(
+        merge_args.merge_base_and_lora_output_dir, None
+    )
+    merge_args.merge_base_and_lora_output_dir = merge_local_dir
+    if gcs_syncer.is_gcs_or_gcsfuse_path(merge_gcs_dir):
+      dirs_to_sync.append((merge_local_dir, merge_gcs_dir))
+
+  # Common flags shared by merging and training.
+  common_parser = argparse.ArgumentParser()
+  common_parser.add_argument('--pretrained_model_name_or_path', required=True)
+  common_parser.add_argument('--huggingface_access_token')
+  common_args, remaining = common_parser.parse_known_args(unknown)
+
+  # Add flags for training.
+  _append_args_to_command_in_place(training_args, training_cmd)
+  _append_args_to_command_in_place(common_args, training_cmd)
+  training_cmd.extend(remaining)  # Remaining args are passed to training cmd.
+  commands = [training_cmd]
+
+  # Only the main node runs merging.
+  if merge_args.merge_base_and_lora_output_dir and node_rank == 0:
+    lora_dir = utils.get_final_checkpoint_path(training_args.output_dir)
+    lora_local_dir, lora_gcs_dir = gcs_syncer.manage_sync_path(
+        lora_dir, node_rank
+    )
+    if gcs_syncer.is_gcs_or_gcsfuse_path(lora_gcs_dir):
+      dirs_to_sync.append((lora_local_dir, lora_gcs_dir))
+
+    merge_cmd = [
+        'WORLD_SIZE=1',  # To ignore other nodes in multi-node setting.
+        'python3',
+        _TASK_TO_SCRIPT[constants.MERGE_CAUSAL_LANGUAGE_MODEL_LORA],
+        f'--finetuned_lora_model_dir={lora_local_dir}',
+    ]
+    _append_args_to_command_in_place(merge_args, merge_cmd)
+    _append_args_to_command_in_place(common_args, merge_cmd)
+
+    # Run in a conda environment.
+    conda_run_cmd = [
+        '/bin/bash',
+        '-c',
+        f'conda run -n merge {" ".join(merge_cmd)}',
+    ]
+    commands.append(conda_run_cmd)
+
+  return commands, dirs_to_sync
+
+
+def _get_merge_cmd_and_dirs_to_sync(
+    task_type: str, config_file: str, unknown: Sequence[str]
+) -> Sequence[Sequence[str]]:
+  """Returns the merge command and dirs to sync.
+
+  Args:
+    task_type: The task type.
+    config_file: The accelerate config file path.
+    unknown: The unknown args which are not recognised by the parser.
+
+  Returns:
+    The bash commands to execute and the directories to sync.
+  """
+  # Merge only flags.
+  merge_parser = argparse.ArgumentParser()
+  merge_parser.add_argument('--merge_base_and_lora_output_dir')
+  merge_args, unknown = merge_parser.parse_known_args(unknown)
+
+  dirs_to_sync = []
+  if merge_args.merge_base_and_lora_output_dir:
+    merge_local_dir, merge_gcs_dir = gcs_syncer.manage_sync_path(
+        merge_args.merge_base_and_lora_output_dir, None
+    )
+    merge_args.merge_base_and_lora_output_dir = merge_local_dir
+    if gcs_syncer.is_gcs_or_gcsfuse_path(merge_gcs_dir):
+      dirs_to_sync.append((merge_local_dir, merge_gcs_dir))
+
+  cmd = launch_script_cmd(_TASK_TO_SCRIPT[task_type], config_file)
+  _append_args_to_command_in_place(merge_args, cmd)
+  cmd.extend(unknown)
+  return [cmd], dirs_to_sync
+
+
+def main(unused_argv: Sequence[str]) -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--config_file')
+  parser.add_argument('--task')
+  parser.add_argument('--gcs_rsync_interval_secs', type=int, default=60)
+  args, unknown = parser.parse_known_args()
+
+  task = args.task
+  dirs_to_sync = None
+
+  if task in _TEXT_TO_IMAGE_TASKS_SCRIPTS:
+    # Setup accelerate config before running trainer.
+    config_gen_cmd = [
+        'python',
+        '-c',
+        (
+            'from accelerate.utils import write_basic_config;'
+            ' write_basic_config(mixed_precision="fp16")'
+        ),
+    ]
+    task_cmd = [
+        'accelerate',
+        'launch',
+        _TEXT_TO_IMAGE_TASKS_SCRIPTS[task],
+    ] + list(map(dataset_validation_util.force_gcs_fuse_path, unknown))
+    commands = [config_gen_cmd, task_cmd]
+  elif task in [constants.INSTRUCT_LORA]:
+    commands, dirs_to_sync = _get_train_and_maybe_merge_cmd_and_dirs_to_sync(
+        task_type=task, config_file=args.config_file, unknown=unknown
+    )
+  elif task in [constants.MERGE_CAUSAL_LANGUAGE_MODEL_LORA]:
+    commands, dirs_to_sync = _get_merge_cmd_and_dirs_to_sync(
+        task_type=task, config_file=args.config_file, unknown=unknown
+    )
+  else:
+    assert task in _TASK_TO_SCRIPT
+    cmd = launch_script_cmd(_TASK_TO_SCRIPT[task], args.config_file)
+    cmd.extend(unknown)
+    commands = [cmd]
+
+  rsync_process = None
+  mp_queue = multiprocessing.Queue(maxsize=1)
+  if dirs_to_sync:
+    rsync_process = gcs_syncer.setup_gcs_rsync(
+        dirs_to_sync, mp_queue, args.gcs_rsync_interval_secs
+    )
+
+  for cmd in commands:
+    logging.info('launching task=%s with cmd: \n%s', task, ' \\\n'.join(cmd))
+    # Both absl logging and python's logging module writes to stderr by default.
+    # Redirect output to stdout on purpose, such that log entries do not get
+    # marked as `Error` in Cloud's Log Explorer.
+    try:
+      subprocess.run(cmd, stdout=sys.stdout, stderr=sys.stdout, check=True)
+    except subprocess.CalledProcessError as e:
+      if rsync_process is not None and rsync_process.is_alive():
+        logging.info('Terminating GCS rsync process.')
+        rsync_process.terminate()
+      raise e
+  if rsync_process is not None:
+    gcs_syncer.cleanup_gcs_rsync(rsync_process, mp_queue)
+
+
+if __name__ == '__main__':
+  logging.get_absl_handler().python_handler.stream = sys.stdout
+  app.run(main, flags_parser=lambda _args: flags.FLAGS(_args, known_only=True))
@@ -0,0 +1,360 @@
+"""Common libraries for PEFT."""
+
+from collections.abc import Mapping, Sequence
+import datetime
+import gc
+import os
+from typing import Any
+
+from absl import logging
+import accelerate
+from accelerate import DistributedType
+from accelerate import PartialState
+import peft
+from peft import PeftModel
+from peft import prepare_model_for_kbit_training
+import torch
+import transformers
+from transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer
+from transformers import BitsAndBytesConfig
+from transformers import FbgemmFp8Config
+import trl
+
+from util import dataset_validation_util
+from util import constants
+
+_LLAMA_3_1_405B_MODEL_ID = "Meta-Llama-3.1-405B"
+_LOCAL_MERGED_MODEL_DIR = "/tmp/merged_model"
+_GEMMA2_MODEL = "gemma-2"
+
+
+def load_model(
+    pretrained_model_name_or_path: str,
+    tokenizer: AutoTokenizer,
+    precision_mode: str = None,
+    gradient_checkpointing: bool = False,
+    gradient_checkpointing_kwargs: Mapping[str, Any] | None = None,
+    access_token: str | None = None,
+    attn_implementation: str | None = None,
+    train_precision: str | None = None,
+    device_map: str | None = None,
+    is_training: bool = True,
+) -> AutoModelForCausalLM:
+  """Loads models from the local dir if specified or from huggingface."""
+  # The `distributed_type` we got through `PartialState` is incorrect for FSDP.
+  # And that's why `Accelerator` is used here.
+  # See b/357138252 for more details.
+  accelerator = accelerate.Accelerator()
+  logging.info("using distributed_type %s", accelerator.distributed_type)
+
+  if device_map is None:
+    if accelerator.distributed_type == DistributedType.MULTI_GPU:
+      # https://github.com/artidoro/qlora/issues/186#issuecomment-1943045599
+      # and b/342038175.
+      device_map = {"": accelerator.process_index}
+    elif accelerator.distributed_type == DistributedType.DEEPSPEED:
+      # Deepspeed Zero3 does not allow setting device_map.
+      # https://github.com/huggingface/transformers/blob/v4.38.2/src/transformers/modeling_utils.py#L2941-L2943
+      device_map = None
+    elif accelerator.distributed_type == DistributedType.FSDP:
+      if precision_mode in [
+          constants.PRECISION_MODE_4,
+          constants.PRECISION_MODE_8,
+      ]:
+        device_map = trl.get_kbit_device_map()
+      else:
+        device_map = None
+    elif (
+        accelerator.distributed_type == DistributedType.NO
+        and torch.cuda.device_count() > 1
+    ):
+      # Setting device map to None to avoid using model parallelism (MP) when
+      # there are multiple GPUs, which can have very inefficient GPU utilization
+      # (b/342252819). This setting should trigger torch's nn.DataParallel
+      # instead, which has better GPU utilization.
+      device_map = None
+    else:
+      device_map = "auto"
+  logging.info("using device_map %s", device_map)
+
+  if train_precision == constants.PRECISION_MODE_32:
+    train_dtype = torch.float32
+  elif train_precision == constants.PRECISION_MODE_16:
+    train_dtype = torch.float16
+  elif train_precision == constants.PRECISION_MODE_16B:
+    train_dtype = torch.bfloat16
+  else:
+    train_dtype = "auto"
+
+  quantization_config = None
+  # Note: use_cache is False when enable gradient checkpointing.
+  if precision_mode == constants.PRECISION_MODE_32:
+    torch_dtype = torch.float32
+  elif precision_mode == constants.PRECISION_MODE_16:
+    torch_dtype = torch.float16
+  elif precision_mode == constants.PRECISION_MODE_16B:
+    torch_dtype = torch.bfloat16
+  elif precision_mode == constants.PRECISION_MODE_8:
+    quantization_config = BitsAndBytesConfig(
+        load_in_8bit=True, int8_threshold=0
+    )
+    torch_dtype = train_dtype
+  elif precision_mode == constants.PRECISION_MODE_4:
+    quantization_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_use_double_quant=True,
+        bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=train_dtype,
+    )
+    # `bnb_4bit_quant_storage` must be set when using FSDP.
+    # https://huggingface.co/docs/bitsandbytes/main/en/fsdp_qlora
+    if accelerator.distributed_type == DistributedType.FSDP:
+      quantization_config.bnb_4bit_quant_storage = train_dtype
+    torch_dtype = train_dtype
+  else:
+    raise ValueError(f"Invalid precision mode: {precision_mode}")
+  logging.info("using torch_type=%s", torch_dtype)
+
+  model_kwargs = {
+      "use_cache": not gradient_checkpointing,
+      "device_map": device_map,
+      "torch_dtype": torch_dtype,
+      "quantization_config": quantization_config,
+      "trust_remote_code": False,
+      "token": access_token,
+      "attn_implementation": attn_implementation,
+  }
+  if _GEMMA2_MODEL in pretrained_model_name_or_path:
+    # The cache_implementation for Gemma 2 is set to hybrid by default. This
+    # param is only supported by Gemma 2. The default 'hybrid' value causes an
+    # issue when use_cache is set to False. So we have to use 'None' in such
+    # cases.
+    # https://github.com/huggingface/transformers/commit/238b13478df209ab534f2195a397dc64a3930883
+    model_kwargs["cache_implementation"] = (
+        None if gradient_checkpointing else "hybrid"
+    )
+
+  model = AutoModelForCausalLM.from_pretrained(
+      pretrained_model_name_or_path, **model_kwargs
+  )
+
+  if precision_mode in (constants.PRECISION_MODE_4, constants.PRECISION_MODE_8):
+    model = prepare_model_for_kbit_training(
+        model,
+        use_gradient_checkpointing=gradient_checkpointing,
+        gradient_checkpointing_kwargs=gradient_checkpointing_kwargs,
+    )
+
+  if gradient_checkpointing:
+    model.gradient_checkpointing_enable(
+        gradient_checkpointing_kwargs=gradient_checkpointing_kwargs
+    )
+
+  # Flash attention only supports fp16 or bf16 [1].
+  # prepare_model_for_kbit_training will force cast some layers to float32 [2]
+  #
+  # [1]: https://github.com/Dao-AILab/flash-attention/issues/882
+  # [2]: https://github.com/huggingface/peft/blob/v0.10.0/src/peft/utils/other.py#L79-L81 # pylint: disable=line-too-long
+  if attn_implementation == "flash_attention_2" and precision_mode in (
+      constants.PRECISION_MODE_4,
+      constants.PRECISION_MODE_8,
+  ):
+    for _, param in model.named_parameters():
+      if param.dtype == torch.float32:
+        param.data = param.data.to(torch_dtype)
+
+  if is_training:
+    # KV cache is useless during training
+    # https://stackoverflow.com/a/77408076
+    model.config.use_cache = False
+
+  if dataset_validation_util.should_add_pad_token(
+      pretrained_model_name_or_path
+  ):
+    model.resize_token_embeddings(len(tokenizer), mean_resizing=False)
+    if is_training:
+      # The following is needed since we added a new token that needs to be
+      # learned.
+      # https://github.com/QwenLM/Qwen/issues/405#issuecomment-1751680291
+      model.enable_input_require_grads()
+
+  return model
+
+
+def _merge_causal_language_model_with_lora_internal(
+    pretrained_model_name_or_path: str,
+    merge_precision_mode: str,
+    finetuned_lora_model_dir: str,
+    merged_model_output_dir: str,
+    access_token: str | None = None,
+) -> None:
+  """Internal function to merges the base model with the lora adapter."""
+  logging.info("loading tokenizer...")
+  tokenizer = dataset_validation_util.load_tokenizer(
+      pretrained_model_name_or_path
+  )
+
+  # Note: merging peft adapter requires loading model in 16 bits, so merging
+  # is done on CPU on purpose in case one GPU cannot hold the base model.
+  logging.info("loading model %s...", pretrained_model_name_or_path)
+  device_map = "cpu"
+  model = load_model(
+      pretrained_model_name_or_path=pretrained_model_name_or_path,
+      tokenizer=tokenizer,
+      precision_mode=merge_precision_mode,
+      access_token=access_token,
+      device_map=device_map,
+      is_training=False,
+  )
+
+  logging.info("loading LoRA model...")
+  model = PeftModel.from_pretrained(
+      model, finetuned_lora_model_dir, device_map=device_map
+  )
+
+  logging.info("merging base model with finetuned LoRA model...")
+  model = model.merge_and_unload()
+
+  logging.info("saving model to %s...", merged_model_output_dir)
+  model.save_pretrained(
+      merged_model_output_dir,
+      safe_serialization=False,
+      is_main_process=PartialState().is_main_process,
+  )
+
+  logging.info("saving tokenizer to %s...", merged_model_output_dir)
+  tokenizer.save_pretrained(
+      merged_model_output_dir,
+      is_main_process=PartialState().is_main_process,
+  )
+
+
+def merge_causal_language_model_with_lora(
+    pretrained_model_name_or_path: str,
+    precision_mode: str,
+    finetuned_lora_model_dir: str,
+    merged_model_output_dir: str,
+    access_token: str | None = None,
+) -> None:
+  """Merges the base model with the lora adapter."""
+
+  # Set merge related variables.
+  if precision_mode == constants.PRECISION_MODE_FP8:
+    # Merge as FP16. FP8 requires conversion after merge.
+    merge_precision_mode = constants.PRECISION_MODE_16
+    local_merged_model_dir = _LOCAL_MERGED_MODEL_DIR
+  else:
+    merge_precision_mode = precision_mode
+    local_merged_model_dir = merged_model_output_dir
+
+  if PartialState().is_main_process:
+    logging.info("Starting merging job...")
+    _merge_causal_language_model_with_lora_internal(
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        merge_precision_mode=merge_precision_mode,
+        finetuned_lora_model_dir=finetuned_lora_model_dir,
+        merged_model_output_dir=local_merged_model_dir,
+        access_token=access_token,
+    )
+    logging.info("merging job is done")
+
+  # Wait for all processes to sync here.
+  PartialState().wait_for_everyone()
+
+  if precision_mode == constants.PRECISION_MODE_FP8:
+    convert_model_to_fp8(
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        merged_model_output_dir=local_merged_model_dir,
+        quantized_model_output_dir=merged_model_output_dir,
+        access_token=access_token,
+    )
+
+
+def convert_model_to_fp8(
+    pretrained_model_name_or_path: str,
+    merged_model_output_dir: str,
+    quantized_model_output_dir: str,
+    access_token: str | None = None,
+) -> None:
+  """Converts the model to fp8.
+
+  Args:
+    pretrained_model_name_or_path: Original base model name or path.
+    merged_model_output_dir: Path to directory containing the merged model.
+    quantized_model_output_dir: Path to directory to save the quantized model.
+    access_token: Access token for accessing the model.
+  """
+  if PartialState().is_main_process:
+    quantization_config = FbgemmFp8Config(
+        modules_to_not_convert=_maybe_get_modules_to_not_convert_by_model_id(
+            pretrained_model_name_or_path
+        )
+    )
+    quantized_model = AutoModelForCausalLM.from_pretrained(
+        merged_model_output_dir,
+        device_map="cpu",
+        quantization_config=quantization_config,
+        trust_remote_code=False,
+        token=access_token,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(merged_model_output_dir)
+
+    quantized_model.save_pretrained(quantized_model_output_dir)
+    tokenizer.save_pretrained(quantized_model_output_dir)
+  PartialState().wait_for_everyone()
+
+
+def force_gc():
+  """Collects garbage immediately to release unused CPU/GPU resources."""
+  gc.collect()
+  torch.cuda.empty_cache()
+
+
+def init_partial_state(
+    timeout: datetime.timedelta = datetime.timedelta(seconds=600),
+) -> None:
+  """Initializes the partial state with timeout."""
+  # This needs to be called before any other PartialState() calls, and
+  # TrainingArguments needs `use_configured_state`. See b/357970482#comment3
+  # for more details.
+  PartialState(timeout=timeout)
+
+
+def print_library_versions():
+  if PartialState().is_main_process:
+    logging.info("======================")
+    logging.info("library versions")
+    logging.info("======================")
+    logging.info("accelerate: %s", accelerate.__version__)
+    logging.info("peft: %s", peft.__version__)
+    logging.info("transformers: %s", transformers.__version__)
+    logging.info("trl: %s", trl.__version__)
+  PartialState().wait_for_everyone()
+
+
+def get_final_checkpoint_path(output_dir: str) -> str:
+  """Returns the final checkpoint path."""
+  return os.path.join(output_dir, constants.FINAL_CHECKPOINT_DIRNAME)
+
+
+def _maybe_get_modules_to_not_convert_by_model_id(
+    pretrained_model_name_or_path: str,
+) -> Sequence[str] | None:
+  """Returns the modules to not convert for the model."""
+  if _LLAMA_3_1_405B_MODEL_ID in pretrained_model_name_or_path:
+    return _get_llama_3_1_405b_modules_to_not_convert()
+  else:
+    return None
+
+
+def _get_llama_3_1_405b_modules_to_not_convert() -> Sequence[str]:
+  """Returns the modules to not convert for Llama 3.1 405B model."""
+  modules_to_not_convert = ["lm_head"]
+  for idx in range(126):
+    for proj_name in ["k_proj", "o_proj", "q_proj", "v_proj"]:
+      modules_to_not_convert.append(f"model.layers.{idx}.self_attn.{proj_name}")
+  for proj_name in ["down_proj", "gate_proj", "up_proj"]:
+    modules_to_not_convert.append(f"model.layers.0.mlp.{proj_name}")
+    modules_to_not_convert.append(f"model.layers.125.mlp.{proj_name}")
+  return tuple(modules_to_not_convert)
@@ -0,0 +1,84 @@
+"""Validate the dataset with the template."""
+
+from typing import Sequence
+
+from absl import app
+from absl import flags
+
+from util import dataset_validation_util
+from vertex_vision_model_garden_peft.train.vmg import utils
+from util import constants
+
+
+_DATASET_NAME = flags.DEFINE_string(
+    'dataset_name',
+    None,
+    'The dataset name in huggingface.',
+    required=True,
+)
+
+_TRAIN_SPLIT = flags.DEFINE_string(
+    'train_split',
+    'train',
+    'The train split name.',
+)
+
+_TRAIN_COLUMN = flags.DEFINE_string(
+    'train_column',
+    constants.DEFAULT_TRAIN_COLUMN,
+    'The instruct column in dataset.',
+)
+
+_TEMPLATE = flags.DEFINE_string(
+    'template',
+    None,
+    'Template for formatting language model training data. Must be a filename'
+    ' under `templates` folder, without `.json` extension, e.g. `alpaca`, or a'
+    ' Cloud Storage URI to a JSON file.',
+    required=True,
+)
+
+_MAX_SEQ_LENGTH = flags.DEFINE_integer(
+    'max_seq_length',
+    None,
+    'The maximum sequence length.',
+)
+
+_VALIDATE_PERCENTAGE_OF_DATASET = flags.DEFINE_integer(
+    'validate_percentage_of_dataset',
+    None,
+    'The percentage of the dataset to validate with the template. If set to'
+    ' -1, it loads the full dataset.',
+)
+
+_VALIDATE_K_ROWS_OF_DATASET = flags.DEFINE_integer(
+    'validate_k_rows_of_dataset',
+    None,
+    'The top k rows of the dataset to validate with the template. If set to -1,'
+    ' it loads the full dataset.',
+)
+
+_USE_MULTIPROCESSING = flags.DEFINE_boolean(
+    'use_multiprocessing',
+    False,
+    'Whether to use multiprocessing for loading the dataset.',
+)
+
+
+def main(unused_argv: Sequence[str]) -> None:
+  utils.print_library_versions()
+
+  dataset_validation_util.validate_dataset_with_template(
+      dataset_name=_DATASET_NAME.value,
+      split=_TRAIN_SPLIT.value,
+      input_column=_TRAIN_COLUMN.value,
+      template=_TEMPLATE.value,
+      max_seq_length=_MAX_SEQ_LENGTH.value,
+      use_multiprocessing=_USE_MULTIPROCESSING.value,
+      validate_percentage_of_dataset=_VALIDATE_PERCENTAGE_OF_DATASET.value,
+      validate_k_rows_of_dataset=_VALIDATE_K_ROWS_OF_DATASET.value,
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)
@@ -1,6 +1,6 @@
 """Common utility lib for prediction on images."""

-from typing import Any, Dict, List
+from typing import Any, Dict, List, Tuple

 import numpy as np
 from PIL import Image
@@ -10,6 +10,28 @@ import yaml
 from util import image_format_converter


+def convert_list_to_label_map(
+    input_list: List[str],
+) -> Tuple[Dict[str, Dict[int, str]], List[int]]:
+  """Converts a list of labels to a dictionary and numerical encoding.
+
+  Args:
+    input_list: A list of strings representing class labels.
+
+  Returns:
+    A tuple containing:
+      label_map: A dictionary mapping unique labels to integer indices.
+      encoded_list: A list of integers corresponding to the labels in the input
+      list.
+  """
+  unique_labels = set(input_list)
+  label_map_reverse = {label: idx for idx, label in enumerate(unique_labels)}
+  label_map = {idx: label for idx, label in enumerate(unique_labels)}
+  encoded_list = [label_map_reverse[label] for label in input_list]
+
+  return {"label_map": label_map}, encoded_list
+
+
 def get_prediction_instances(image: Image.Image) -> List[Dict[str, Any]]:
  """Gets prediction instances.

@@ -40,14 +62,14 @@ def get_label_map(label_map_yaml_filepath: str) -> Dict[str, Any]:


 def get_object_detection_endpoint_predictions(
-    detection_endpoint: ...,
+    detector_endpoint: ...,
    input_image: np.ndarray,
    detection_thresh: float = 0.2,
 ) -> np.ndarray:
  """Gets endpoint predictions.

  Args:
-      detection_endpoint: image object detection endpoint.
+      detector_endpoint: image object detection endpoint.
      input_image: Input image.
      detection_thresh: Detection threshold.

@@ -55,9 +77,10 @@ def get_object_detection_endpoint_predictions(
      Object detection predictions from endpoints.
  """
  height, width, _ = input_image.shape
-  predictions = detection_endpoint.predict(
+  predictions = detector_endpoint.predict(
      get_prediction_instances(Image.fromarray(input_image))
  ).predictions
+
  detection_scores = np.array(predictions[0]["detection_scores"])
  detection_classes = np.array(predictions[0]["detection_classes"])
  detection_boxes = np.array(
@@ -66,6 +89,29 @@ def get_object_detection_endpoint_predictions(
          for b in predictions[0]["detection_boxes"]
      ]
  )
+  return merge_boxes_and_classes(
+      detection_scores, detection_boxes, detection_classes, detection_thresh
+  )
+
+
+def merge_boxes_and_classes(
+    detection_scores: np.ndarray,
+    detection_boxes: np.ndarray,
+    detection_classes: np.ndarray,
+    detection_thresh: float = 0.2,
+) -> np.ndarray:
+  """Merges prediction boxes and classes.
+
+  Args:
+      detection_scores: array of detection scores.
+      detection_boxes: array of detection boxes.
+      detection_classes: array of detection classes.
+      detection_thresh: float indicating the detection threshold.
+
+  Returns:
+      preds_merge_cls: a numpy array containing the detection boxes, scores and
+      classes.
+  """
  thresh_indices = [
      x for x, val in enumerate(detection_scores) if val > detection_thresh
  ]
@@ -76,4 +122,5 @@ def get_object_detection_endpoint_predictions(
  preds_merge_cls = np.column_stack(
      (preds_merge_conf, detection_classes[thresh_indices])
  )
+
  return preds_merge_cls
--- a/Show More
+++ b/Show More