FSDP Example for T5 Fine-Tuning and PyTorchJob #2286

andreyvelich · 2024-10-16T19:38:08Z

Related: #2040

I added simple FSDP example for T5 fine-tuning and PyTorchJob.

I refactored example from the PyTorch tutorial: https://pytorch.org/tutorials/intermediate/FSDP_adavnced_tutorial.html

/assign @kubeflow/wg-training-leads

Signed-off-by: Andrey Velichkevich <[email protected]>

review-notebook-app · 2024-10-16T19:38:15Z

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

andreyvelich · 2024-10-16T19:38:55Z

cc @kuizhiqing @Syulin7

coveralls · 2024-10-16T19:42:44Z

Pull Request Test Coverage Report for Build 11409185571

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

For more information on this, see Tracking coverage changes with pull request builds.
To avoid this issue with future PRs, see these Recommended CI Configurations.
For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

0 of 0 changed or added relevant lines in 0 files are covered.
No unchanged relevant lines lost coverage.
Overall coverage remained the same at 100.0%

Totals
Change from base Build 11330381194:	0.0%
Covered Lines:	73
Relevant Lines:	73

💛 - Coveralls

kuizhiqing

Good job.

kuizhiqing · 2024-10-18T09:21:50Z

examples/pytorch/fsdp/fine-tune-t5-with-fsdp.ipynb

+    "\n",
+    "Wikihow dataset: https://github.com/mahnazkoupaee/WikiHow-Dataset\n",
+    "\n",
+    "This Notebook will use **4** GPUs to fine-tune T5 model on 2 Nodes and 4 GPUs. This example is based on [the official PyTorch FSDP tutorial](https://pytorch.org/tutorials/intermediate/FSDP_adavnced_tutorial.html)."


Maybe ~~and 4 GPUs~~ ?

Let's just say on 2 Nodes.

kuizhiqing · 2024-10-18T09:23:25Z

examples/pytorch/fsdp/fine-tune-t5-with-fsdp.ipynb

+    "\n",
+    "Use `TrainingClient()` to create PyTorchJob which will fine-tine T5 on **2 workers** using **2 GPU** for each worker.\n",
+    "\n",
+    "If you don't have enough GPU resources, you can decrease number of workers or number of GPUs per worker."


Do we have minimal required resources, user may suffer OOM otherwise.

I was able to run this example on a single V100 Nvidia GPU.
T5-base has just 223M params: https://huggingface.co/google-t5/t5-base

kuizhiqing · 2024-10-18T09:34:28Z

examples/pytorch/fsdp/fine-tune-t5-with-fsdp.ipynb

+    "def train_function(parameters):\n",
+    "    import os\n",
+    "    import time\n",
+    "    import functools\n",
+    "\n",
+    "    import torch\n",
+    "    import torch.distributed as dist\n",
+    "    from torch.utils.data.distributed import DistributedSampler\n",
+    "    from torch.distributed.fsdp import FullyShardedDataParallel as FSDP\n",
+    "    from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy\n",
+    "\n",
+    "    from transformers import T5Tokenizer, T5ForConditionalGeneration\n",
+    "    from transformers.models.t5.modeling_t5 import T5Block\n",
+    "    from datasets import Dataset\n",


It's not a problem, just curious that why prefer encapsule the whole script as a function and include import in it.

This is how create_job() API works today with train function.
As you can see, we wrap the whole function body under containers args: https://github.com/kubeflow/training-operator/blob/master/sdk/python/kubeflow/training/utils/utils.py#L134-L140, similar to KFP lightweight Python components.

Do you have any suggestions/ideas on how we can serialize function without asking users to make imports defined under the function ?

Got it. Thanks for your information.
I have no better solution in mind now for this kind of usage. While I was thinking something like function level scheduling thing, despite KFP lightweight Python Components and Ray, please share more if you are familiar with other good practices.

@kuizhiqing One of the options could be to use cloudpickle: https://github.com/cloudpipe/cloudpickle
That allows us to serialize function to byte string and then de-serialize it on the container.
That supports imports outside of serialized function.

Signed-off-by: Andrey Velichkevich <[email protected]>

Syulin7 · 2024-10-20T15:37:24Z

examples/pytorch/fsdp/fine-tune-t5-with-fsdp.ipynb

+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## FSDP with multi-node multi-worker training"


the official PyTorch FSDP tutorial showcase the training on a single node.

In this example, multi-node is used for training, and the model is not saved, correct?
If the model needs to be saved, is it necessary to use PVC and S3?

Similar to this example, we can save model on the process with RANK=0:
https://pytorch.org/tutorials/intermediate/FSDP_adavnced_tutorial.html#:~:text=if%20rank%20%3D%3D%200%3A%0A%20%20%20%20%20%20%20%20%20%20%20%20%20%20%20%20print(f%22%2D%2D%3E%20saving,)%0A%0A%20%20%20%20%20%20%20%20%20%20%20%20%20%20%20%20torch.save(cpu_state%2C%20save_name)

Syulin7 · 2024-10-22T01:06:03Z

Good work!
/lgtm

andreyvelich · 2024-10-22T12:17:47Z

Thanks for the review!
/approve

google-oss-prow · 2024-10-22T12:17:56Z

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: andreyvelich

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Needs approval from an approver in each of these files:

~~OWNERS~~ [andreyvelich]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

tenzen-y · 2024-10-22T20:32:14Z

Thanks for the review! /approve

Why the approve label is not automatically added to this PR?
IIUC, the PRs created by approver have approve label.

andreyvelich · 2024-10-23T02:02:09Z

@tenzen-y Actually, we disabled it as part of this PR: GoogleCloudPlatform/oss-test-infra#2271.
That will help us to not accidentally merge PRs that are not ready and got /lgtm from the GitHub members.
We noticed that, not all contributors add the /hold label if PR is not ready.

tenzen-y · 2024-10-23T02:08:10Z

@tenzen-y Actually, we disabled it as part of this PR: GoogleCloudPlatform/oss-test-infra#2271.

That will help us to not accidentally merge PRs that are not ready and got /lgtm from the GitHub members.

We noticed that, not all contributors add the /hold label if PR is not ready.

That makes sense. Thank you for clarifications. That seems the limitations (accidentally merging) for collaborations between GitHub Actions and Prow :(

* FSDP Example with PyTorchJob and T5 Fine-Tuning Signed-off-by: Andrey Velichkevich <[email protected]> * Modify text Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]>

* Added test for create-pytorchjob.ipynb Signed-off-by: sailesh duddupudi <[email protected]> * fix yaml syntax Signed-off-by: sailesh duddupudi <[email protected]> * Fix uses path Signed-off-by: sailesh duddupudi <[email protected]> * Add actions/checkout Signed-off-by: sailesh duddupudi <[email protected]> * Add bash to action.yaml Signed-off-by: sailesh duddupudi <[email protected]> * Install pip dependencies step Signed-off-by: sailesh duddupudi <[email protected]> * Add quotes for args Signed-off-by: sailesh duddupudi <[email protected]> * Add jupyter Signed-off-by: sailesh duddupudi <[email protected]> * Add nbformat_minor: 5 to fix invalid format error Signed-off-by: sailesh duddupudi <[email protected]> * Fix job name Signed-off-by: sailesh duddupudi <[email protected]> * test papermill-args-yaml Signed-off-by: sailesh duddupudi <[email protected]> * testing multi line args Signed-off-by: sailesh duddupudi <[email protected]> * testing multi line args1 Signed-off-by: sailesh duddupudi <[email protected]> * testing multi line args2 Signed-off-by: sailesh duddupudi <[email protected]> * testing multi line args3 Signed-off-by: sailesh duddupudi <[email protected]> * Parameterize sdk install Signed-off-by: sailesh duddupudi <[email protected]> * Remove unnecessary output Signed-off-by: sailesh duddupudi <[email protected]> * nbformat normailze Signed-off-by: sailesh duddupudi <[email protected]> * [SDK] Training Client Conditions related unit tests (#2253) * test: add unit test for get_job_conditions function of training client Signed-off-by: Bobbins228 <[email protected]> * test: add unit test for is_job_created function of training client Signed-off-by: Bobbins228 <[email protected]> * test: add unit test for is_job_running function of training client Signed-off-by: Bobbins228 <[email protected]> * test: add unit test for is_job_restarting function of training client Signed-off-by: Bobbins228 <[email protected]> * test: add unit test for is_job_failed function of training client Signed-off-by: Bobbins228 <[email protected]> * test: add unit test for is_job_succeded function of training client Signed-off-by: Bobbins228 <[email protected]> * test: improve job condition unit tests efficiency Signed-off-by: Bobbins228 <[email protected]> --------- Signed-off-by: Bobbins228 <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * [SDK] test: add unit test for list_jobs method of the training_client (#2267) Signed-off-by: wei-chenglai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Generate clientset, openapi spec for the V2 APIs (#2273) Generate clientset, informers, listers and open api spec for v2alpha1 APIs. Signed-off-by: Varsha Prasad Narsing <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * [SDK] Use torchrun to create PyTorchJob from function (#2276) * [SDK] Use torchrun to create PyTorchJob from function Signed-off-by: Andrey Velichkevich <[email protected]> * Update PyTorchJob SDK example Signed-off-by: Andrey Velichkevich <[email protected]> * Add consts for entrypoint Signed-off-by: Andrey Velichkevich <[email protected]> * Add check for num procs per worker Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * [SDK] test: add unit test for get_job_logs method of the training_client (#2275) Signed-off-by: wei-chenglai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * [v2alpha] Move GV related codebase (#2281) Move GV related codebase in v2alpha Signed-off-by: Varsha Prasad Narsing <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Implement runtime framework (#2248) * KEP-2170: Implement runtime framework interfaces Signed-off-by: Yuki Iwai <[email protected]> * Remove grep dependency Signed-off-by: Yuki Iwai <[email protected]> * KEP-2170: Implement ValidateObjects interface to the runtime framework Signed-off-by: Yuki Iwai <[email protected]> * KEP-2170: Expose the TrainingRuntime and ClusterTrainingRuntime Kind Signed-off-by: Yuki Iwai <[email protected]> * KEP-2170: Remove unneeded scheme field from the internal TrainingRuntime Signed-off-by: Yuki Iwai <[email protected]> * Rephrase the error message Signed-off-by: Yuki Iwai <[email protected]> * Distinguish TrainingRuntime and ClusterTrainingRuntime when creating indexes for the TrainJobs Signed-off-by: Yuki Iwai <[email protected]> * Propagate the TrainJob labels and annotations to the JobSet Signed-off-by: Yuki Iwai <[email protected]> * Remove PodAnnotations from the runtime info Signed-off-by: Yuki Iwai <[email protected]> * Implement TrainingRuntime ReplicatedJob validation Signed-off-by: Yuki Iwai <[email protected]> * Add TODO comments Signed-off-by: Yuki Iwai <[email protected]> * Replace queueSuspendedTrainJob with queueSuspendedTrainJobs Signed-off-by: Yuki Iwai <[email protected]> --------- Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Add DeepSpeed Example with Pytorch Operator (#2235) Signed-off-by: Syulin7 <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Rename TrainingRuntimeRef to RuntimeRef API (#2283) * KEP-2170: Rename TrainingRuntimeRef to RuntimeRef API Signed-off-by: Andrey Velichkevich <[email protected]> * Rename RuntimeRef in runtime framework Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Adding CEL validations on v2 TrainJob CRD (#2260) Signed-off-by: Akshay Chitneni <[email protected]> Co-authored-by: Akshay Chitneni <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Upgrade Deepspeed demo dependencies (#2294) Signed-off-by: Syulin7 <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Add manifests for Kubeflow Training V2 (#2289) * KEP-2170: Add manifests for Kubeflow Training V2 Signed-off-by: Andrey Velichkevich <[email protected]> * Fix invalid name for webhook config in cert Signed-off-by: Andrey Velichkevich <[email protected]> * Fix integration tests Signed-off-by: Andrey Velichkevich <[email protected]> * Move kubebuilder markers to runtime framework Signed-off-by: Andrey Velichkevich <[email protected]> * Use Kubernetes recommended labels Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * FSDP Example for T5 Fine-Tuning and PyTorchJob (#2286) * FSDP Example with PyTorchJob and T5 Fine-Tuning Signed-off-by: Andrey Velichkevich <[email protected]> * Modify text Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Implement TrainJob Reconciler to manage objects (#2295) * KEP-2170: Implement TrainJob Reconciler to manage objects Signed-off-by: Yuki Iwai <[email protected]> * Mode dep-crds to manifests/external-crds Signed-off-by: Yuki Iwai <[email protected]> * Rename run with runtime Signed-off-by: Yuki Iwai <[email protected]> --------- Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Remove Prometheus Monitoring doc (#2301) Signed-off-by: Sophie <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Decouple JobSet from TrainJob (#2296) Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Strictly verify the CRD marker validation and defaulting in the integration testings (#2304) Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Initialize runtimes before the manager starts (#2306) Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Generate Python SDK for Kubeflow Training V2 (#2310) * Generate SDK models for the Training V2 APIs Signed-off-by: Andrey Velichkevich <[email protected]> * Create pyproject.toml config Signed-off-by: Andrey Velichkevich <[email protected]> * Remove comments Signed-off-by: Andrey Velichkevich <[email protected]> * Fix pre-commit Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Create model and dataset initializers (#2303) * KEP-2170: Create model and dataset initializers Signed-off-by: Andrey Velichkevich <[email protected]> * Add abstract classes Signed-off-by: Andrey Velichkevich <[email protected]> * Add storage URI to config Signed-off-by: Andrey Velichkevich <[email protected]> * Update .gitignore Co-authored-by: Kevin Hannon <[email protected]> Signed-off-by: Andrey Velichkevich <[email protected]> * Fix the misspelling for initializer Signed-off-by: Andrey Velichkevich <[email protected]> * Add .pt and .pth to ignore_patterns Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Co-authored-by: Kevin Hannon <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Implement JobSet, PlainML, and Torch Plugins (#2308) * KEP-2170: Implement JobSet and PlainML Plugins Signed-off-by: Andrey Velichkevich <[email protected]> * Fix nil pointer exception for Trainer Signed-off-by: Andrey Velichkevich <[email protected]> * Fix unit tests in runtime package Signed-off-by: Andrey Velichkevich <[email protected]> * Fix unit tests Signed-off-by: Andrey Velichkevich <[email protected]> * Fix integration tests Signed-off-by: Andrey Velichkevich <[email protected]> * Fix lint Signed-off-by: Andrey Velichkevich <[email protected]> * Implement Torch Plugin Signed-off-by: Andrey Velichkevich <[email protected]> * Use list for the Info envs Signed-off-by: Andrey Velichkevich <[email protected]> * Fix golang ci Signed-off-by: Andrey Velichkevich <[email protected]> * Fix Torch plugin Signed-off-by: Andrey Velichkevich <[email protected]> * Use K8s sets Update error return Use ptr.Deref() for nil values Signed-off-by: Andrey Velichkevich <[email protected]> * Use client.Object for Build() call Signed-off-by: Andrey Velichkevich <[email protected]> * Remove DeepCopy Signed-off-by: Andrey Velichkevich <[email protected]> * Remove MLPolicy and PodGroupPolicy from the Info object Signed-off-by: Andrey Velichkevich <[email protected]> * Inline error Signed-off-by: Andrey Velichkevich <[email protected]> * Remove SDK jar file Signed-off-by: Andrey Velichkevich <[email protected]> * Add integration test for Torch plugin Signed-off-by: Andrey Velichkevich <[email protected]> * Add TODO to calculate PodGroup values in unit tests Signed-off-by: Andrey Velichkevich <[email protected]> * Revert the change to add original Runtime Policies to Info Signed-off-by: Andrey Velichkevich <[email protected]> * Create const for the DefaultJobReplicas Signed-off-by: Andrey Velichkevich <[email protected]> * Check if PodLabels is empty Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Implement Initializer builders in the JobSet plugin (#2316) * KEP-2170: Implement Initializer builder in the JobSet plugin Signed-off-by: Andrey Velichkevich <[email protected]> * Update the SDK models Signed-off-by: Andrey Velichkevich <[email protected]> * Remove Info from Initializer builder Signed-off-by: Andrey Velichkevich <[email protected]> * Update manifests Signed-off-by: Andrey Velichkevich <[email protected]> * Update pkg/constants/constants.go Co-authored-by: Yuki Iwai <[email protected]> Signed-off-by: Andrey Velichkevich <[email protected]> * Use var for envs Signed-off-by: Andrey Velichkevich <[email protected]> * Remove check manifests from GitHub actions Signed-off-by: Andrey Velichkevich <[email protected]> * Move consts to JobSet plugin Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Co-authored-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Add the TrainJob state transition design (#2298) * KEP-2170: Add the TrainJob state transition design Signed-off-by: Yuki Iwai <[email protected]> * Replace actual jobs with TrainJob Signed-off-by: Yuki Iwai <[email protected]> * Remove the JobSet conditions propagation and Add expanding runtime framework interfaces for each plugin Signed-off-by: Yuki Iwai <[email protected]> * Expand the Creation Failed reasons Signed-off-by: Yuki Iwai <[email protected]> * Rename Completed to Complete Signed-off-by: Yuki Iwai <[email protected]> --------- Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Update tf job examples to tf v2 (#2270) * mnist with summaries updaetd to TF v2 Signed-off-by: yelias <[email protected]> * tf_sample updaetd to TF v2 Signed-off-by: yelias <[email protected]> * Add mnist_utils and update dist-mnist Signed-off-by: yelias <[email protected]> * Add mnist_utils and update dist-mnist Signed-off-by: yelias <[email protected]> * Remove old example - estimator-API, this example has been replaced by distribution_strategy Signed-off-by: yelias <[email protected]> * Small fix Signed-off-by: yelias <[email protected]> * Remove unsupported powerPC dockerfiles Signed-off-by: yelias <[email protected]> * Fix typo in copyright Signed-off-by: yelias <[email protected]> --------- Signed-off-by: yelias <[email protected]> Co-authored-by: yelias <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Add TrainJob conditions (#2322) * KEP-2170: Implement TrainJob conditions Signed-off-by: Yuki Iwai <[email protected]> * Fix API comments Signed-off-by: Yuki Iwai <[email protected]> * Make condition message constants Signed-off-by: Yuki Iwai <[email protected]> * Stop connecting condition type and reason in JobSet plugin Signed-off-by: Yuki Iwai <[email protected]> --------- Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Pin Gloo repository in JAX Dockerfile to a specific commit (#2329) This commit pins the Gloo repository to a specific commit (43b7acbf) in the JAX Dockerfile to prevent build failures caused by a recent bug introduced in the Gloo codebase. By locking the version of Gloo to a known working commit, we ensure that the JAX build remains stable and functional until the issue is resolved upstream. The build failure occurs when compiling the gloo/transport/tcp/buffer.cc file due to an undefined __NR_gettid constant, which was introduced after the pinned commit. By using this commit, we bypass the issue and allow the build to complete successfully. Signed-off-by: Sandipan Panda <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * [fix] Resolve v2alpha API exceptions (#2317) Resolve v2alpha API exceptions by adding necessary listType validations. Signed-off-by: Varsha Prasad Narsing <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Upgrade Kubernetes to v1.30.7 (#2332) * Upgrade Kubernetes to v1.30.7 Signed-off-by: Antonin Stefanutti <[email protected]> * Use typed event handlers and predicates in job controllers Signed-off-by: Antonin Stefanutti <[email protected]> * Re-organize pkg/common/util/reconciler.go Signed-off-by: Antonin Stefanutti <[email protected]> * Update installation instructions in README Signed-off-by: Antonin Stefanutti <[email protected]> --------- Signed-off-by: Antonin Stefanutti <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Ignore cache exporting errors in the image building workflows (#2336) Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * KEP-2170: Add Torch Distributed Runtime (#2328) * KEP-2170: Add Torch Distributed Runtime Signed-off-by: Andrey Velichkevich <[email protected]> * Add pip list Signed-off-by: Andrey Velichkevich <[email protected]> --------- Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Refine the server-side apply installation args (#2337) Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Add openapi-generator CLI option to skip SDK v2 test generation (#2338) Signed-off-by: Antonin Stefanutti <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Upgrade kustomization files to Kustomize v5 (#2326) Signed-off-by: oksanabaza <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Pin accelerate package version in trainer (#2340) * Pin accelerate package version in trainer Signed-off-by: Gavrish Prabhu <[email protected]> * include new line to pass pre-commit hook Signed-off-by: Gavrish Prabhu <[email protected]> --------- Signed-off-by: Gavrish Prabhu <[email protected]> Signed-off-by: sailesh duddupudi <[email protected]> * Replace papermill command with bash script Signed-off-by: sailesh duddupudi <[email protected]> * Typo fix Signed-off-by: sailesh duddupudi <[email protected]> * Move Checkout step outside action.yaml file Signed-off-by: sailesh duddupudi <[email protected]> * Add newline EOF in script Signed-off-by: sailesh duddupudi <[email protected]> * Pass python dependencies as args and pin versions Signed-off-by: sailesh duddupudi <[email protected]> * Update Usage Signed-off-by: sailesh duddupudi <[email protected]> * Install dependencies in yaml Signed-off-by: sailesh duddupudi <[email protected]> * fix ipynb Signed-off-by: sailesh duddupudi <[email protected]> * set bash flags Signed-off-by: sailesh duddupudi <[email protected]> * Update script args and add more kubernetes versions for tests Signed-off-by: sailesh duddupudi <[email protected]> * add gang-scheduler-name to template Signed-off-by: sailesh duddupudi <[email protected]> * move go setup to template Signed-off-by: sailesh duddupudi <[email protected]> * remove -p parameter from script Signed-off-by: sailesh duddupudi <[email protected]> --------- Signed-off-by: sailesh duddupudi <[email protected]> Signed-off-by: Bobbins228 <[email protected]> Signed-off-by: wei-chenglai <[email protected]> Signed-off-by: Varsha Prasad Narsing <[email protected]> Signed-off-by: Andrey Velichkevich <[email protected]> Signed-off-by: Yuki Iwai <[email protected]> Signed-off-by: Syulin7 <[email protected]> Signed-off-by: Akshay Chitneni <[email protected]> Signed-off-by: Sophie <[email protected]> Signed-off-by: yelias <[email protected]> Signed-off-by: Sandipan Panda <[email protected]> Signed-off-by: Antonin Stefanutti <[email protected]> Signed-off-by: oksanabaza <[email protected]> Signed-off-by: Gavrish Prabhu <[email protected]> Co-authored-by: Mark Campbell <[email protected]> Co-authored-by: Wei-Cheng Lai <[email protected]> Co-authored-by: Varsha <[email protected]> Co-authored-by: Andrey Velichkevich <[email protected]> Co-authored-by: Yuki Iwai <[email protected]> Co-authored-by: yu lin <[email protected]> Co-authored-by: Akshay Chitneni <[email protected]> Co-authored-by: Akshay Chitneni <[email protected]> Co-authored-by: Sophie Hsu <[email protected]> Co-authored-by: Kevin Hannon <[email protected]> Co-authored-by: YosiElias <[email protected]> Co-authored-by: yelias <[email protected]> Co-authored-by: Sandipan Panda <[email protected]> Co-authored-by: Antonin Stefanutti <[email protected]> Co-authored-by: Oksana Bazylieva <[email protected]> Co-authored-by: Gavrish Prabhu <[email protected]>

FSDP Example with PyTorchJob and T5 Fine-Tuning

92d4a9d

Signed-off-by: Andrey Velichkevich <[email protected]>

google-oss-prow bot requested review from jinchihe and kuizhiqing October 16, 2024 19:38

google-oss-prow bot added the size/XL label Oct 16, 2024

kuizhiqing reviewed Oct 18, 2024

View reviewed changes

Modify text

a35b3e2

Signed-off-by: Andrey Velichkevich <[email protected]>

Syulin7 reviewed Oct 20, 2024

View reviewed changes

google-oss-prow bot assigned Syulin7 Oct 22, 2024

google-oss-prow bot added the lgtm label Oct 22, 2024

google-oss-prow bot added the approved label Oct 22, 2024

google-oss-prow bot merged commit 66ebecd into kubeflow:master Oct 22, 2024
41 checks passed

andreyvelich deleted the fsdp-example-torch branch October 22, 2024 12:22

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FSDP Example for T5 Fine-Tuning and PyTorchJob #2286

FSDP Example for T5 Fine-Tuning and PyTorchJob #2286

andreyvelich commented Oct 16, 2024

review-notebook-app bot commented Oct 16, 2024

andreyvelich commented Oct 16, 2024

coveralls commented Oct 16, 2024 •

edited

Loading

kuizhiqing left a comment

kuizhiqing Oct 18, 2024

andreyvelich Oct 18, 2024

kuizhiqing Oct 18, 2024

andreyvelich Oct 18, 2024 •

edited

Loading

kuizhiqing Oct 18, 2024

andreyvelich Oct 18, 2024

kuizhiqing Oct 19, 2024

andreyvelich Oct 21, 2024

Syulin7 Oct 20, 2024

andreyvelich Oct 21, 2024

Syulin7 commented Oct 22, 2024

andreyvelich commented Oct 22, 2024

google-oss-prow bot commented Oct 22, 2024

tenzen-y commented Oct 22, 2024

andreyvelich commented Oct 23, 2024 •

edited

Loading

tenzen-y commented Oct 23, 2024

FSDP Example for T5 Fine-Tuning and PyTorchJob #2286

FSDP Example for T5 Fine-Tuning and PyTorchJob #2286

Conversation

andreyvelich commented Oct 16, 2024

review-notebook-app bot commented Oct 16, 2024

andreyvelich commented Oct 16, 2024

coveralls commented Oct 16, 2024 • edited Loading

Pull Request Test Coverage Report for Build 11409185571

Warning: This coverage report may be inaccurate.

Details

💛 - Coveralls

kuizhiqing left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

andreyvelich Oct 18, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Syulin7 commented Oct 22, 2024

andreyvelich commented Oct 22, 2024

google-oss-prow bot commented Oct 22, 2024

tenzen-y commented Oct 22, 2024

andreyvelich commented Oct 23, 2024 • edited Loading

tenzen-y commented Oct 23, 2024

coveralls commented Oct 16, 2024 •

edited

Loading

andreyvelich Oct 18, 2024 •

edited

Loading

andreyvelich commented Oct 23, 2024 •

edited

Loading