Merge branch 'main' into main

Author: CollapsedSilhouette

.github/workflows/components-integration-tests.yaml

@@ -24,8 +24,6 @@ jobs:
             platform: ubuntu-latest
           - scheduler: "local_docker"
             platform: linux.24_04.4x
-          - scheduler: "ray"
-            platform: ubuntu-latest
       fail-fast: false
     runs-on: ${{ matrix.platform }}
     permissions:

.github/workflows/kfp-integration-tests.yaml

@@ -10,7 +10,7 @@ jobs:
   unittest:
     strategy:
       matrix:
-        python-version: [3.9, "3.10", 3.11, 3.12]
+        python-version: ["3.10", 3.11, 3.12]
         platform: [linux.24_04.4x]
         include:
           - python-version: 3.12

.github/workflows/python-unittests.yaml

@@ -39,11 +39,6 @@ lintrunner init
 lintrunner -a
 ```
 
-## Integration Tests
-
-See the [KFP integration test](scripts/kfpint.py) file for more details on setup
-and running them.
-
 ## License
 By contributing to TorchX, you agree that your contributions will be licensed
 under the LICENSE file in the root directory of this source tree.

CONTRIBUTING.md

@@ -19,7 +19,6 @@ TorchX currently supports:
 * AWS Batch
 * Docker
 * Local
-* Ray (prototype)
 * GCP Batch (prototype)
 
 Need a scheduler not listed? [Let us know!](https://github.com/pytorch/torchx/issues?q=is%3Aopen+is%3Aissue+label%3Ascheduler-request)
@@ -55,15 +54,9 @@ pip install torchx
 # install torchx sdk and CLI -- all dependencies
 pip install "torchx[dev]"
 
-# install torchx kubeflow pipelines (kfp) support
-pip install "torchx[kfp]"
-
 # install torchx Kubernetes / Volcano support
 pip install "torchx[kubernetes]"
 
-# install torchx Ray support
-pip install "torchx[ray]"
-
 # install torchx GCP Batch support
 pip install "torchx[gcp_batch]"
 ```

README.md

@@ -12,9 +12,6 @@ google-cloud-logging==3.10.0
 google-cloud-runtimeconfig==0.34.0
 hydra-core
 ipython
-kfp==1.8.22
-# pin protobuf to the version that is required by kfp
-protobuf==3.20.3
 mlflow-skinny
 moto~=5.0.8
 pyre-extensions
@@ -32,7 +29,6 @@ torchtext==0.18.0
 torchvision==0.23.0
 typing-extensions
 ts==0.5.1
-ray[default]
 wheel
 
 # lint (linter versions are managed by lintrunner)
@@ -45,4 +41,3 @@ grpcio==1.62.1
 grpcio-status==1.48.1
 googleapis-common-protos==1.63.0
 google-api-core==2.18.0
-protobuf==3.20.3 # kfp==1.8.22 needs protobuf < 4

dev-requirements.txt

@@ -14,8 +14,7 @@ The top level modules in TorchX are:
 4. :mod:`torchx.cli`: CLI tool
 5. :mod:`torchx.runner`: given an app spec, submits the app as a job on a scheduler
 6. :mod:`torchx.schedulers`: backend job schedulers that the runner supports
-7. :mod:`torchx.pipelines`: adapters that convert the given app spec to a "stage" in an ML pipeline platform
-8. :mod:`torchx.runtime`: util and abstraction libraries you can use in authoring apps (not app spec)
+7. :mod:`torchx.runtime`: util and abstraction libraries you can use in authoring apps (not app spec)
 
 Below is a UML diagram
 
@@ -32,8 +31,7 @@ the actual application. In scheduler lingo, this is a ``JobDefinition`` and a
 similar concept in Kubernetes is the ``spec.yaml``. To disambiguate between the
 application binary (logic) and the spec, we typically refer to a TorchX
 ``AppDef`` as an "app spec" or ``specs.AppDef``. It
-is the common interface understood by ``torchx.runner``
-and ``torchx.pipelines`` allowing you to run your app as a standalone job
+is the common interface understood by ``torchx.runner`` allowing you to run your app as a standalone job
 or as a stage in an ML pipeline.
 
 Below is a simple example of an ``specs.AppDef`` that echos "hello world"
@@ -119,10 +117,6 @@ can be achieved through python function composition rather than object compositi
 However **we do not recommend component composition** for maintainability
 purposes.
 
-**PROTIP 2:** To define dependencies between components, use a pipelining DSL.
-See :ref:`basics:Pipeline Adapters` section below to understand how TorchX components
-are used in the context of pipelines.
-
 Before authoring your own component, browse through the library of
 :ref:`Components` that are included with TorchX
 to see if one fits your needs.
@@ -141,34 +135,11 @@ There are two ways to access runners in TorchX:
 See :ref:`Schedulers` for a list of schedulers that the runner can
 launch apps to.
 
-Pipeline Adapters
-~~~~~~~~~~~~~~~~~~~~~~
-While runners launch components as standalone jobs, ``torchx.pipelines``
-makes it possible to plug components into an ML pipeline/workflow. For a
-specific target pipeline platform (e.g. kubeflow pipelines), TorchX
-defines an adapter that converts a TorchX app spec to whatever the
-"stage" representation is in the target platform. For instance,
-``torchx.pipelines.kfp`` adapter for kubeflow pipelines converts an
-app spec to a ``kfp.ContainerOp`` (or more accurately, a kfp "component spec" yaml).
-
-
-In most cases an app spec would map to a "stage" (or node) in a pipeline.
-However advanced components, especially those that have a mini control flow
-of its own (e.g. HPO), may map to a "sub-pipeline" or an "inline-pipeline".
-The exact semantics of how these advanced components map to the pipeline
-is dependent on the target pipeline platform. For example, if the
-pipeline DSL allows dynamically adding stages to a pipeline from an upstream
-stage, then TorchX may take advantage of such feature to "inline" the
-sub-pipeline to the main pipeline. TorchX generally tries its best to adapt
-app specs to the **most canonical** representation in the target pipeline platform.
-
-See :ref:`Pipelines` for a list of supported pipeline platforms.
-
 Runtime
 ~~~~~~~~
 .. important:: ``torchx.runtime`` is by no means is a requirement to use TorchX.
                If your infrastructure is fixed and you don't need your application
-               to be portable across different types of schedulers and pipelines,
+               to be portable across different types of schedulers,
                you can skip this section.
 
 Your application (not the app spec, but the actual app binary) has **ZERO** dependencies

docs/source/basics.rst

@@ -341,7 +341,7 @@ def handle_item(fieldarg, content):
     code_url = f"https://github.com/pytorch/torchx/archive/refs/heads/{notebook_version}.tar.gz"
 
 first_notebook_cell = f"""
-!pip install torchx[kfp]
+!pip install torchx
 !wget --no-clobber {code_url}
 !tar xf {notebook_version}.tar.gz --strip-components=1
 
@@ -351,7 +351,6 @@ def handle_item(fieldarg, content):
 sphinx_gallery_conf = {
     "examples_dirs": [
         "../../torchx/examples/apps",
-        "../../torchx/examples/pipelines",
     ],
     "gallery_dirs": [
         "examples_apps",

docs/source/conf.py

@@ -4,8 +4,6 @@ TorchX
 ==================
 
 TorchX is a universal job launcher for PyTorch applications.
-TorchX is designed to have fast iteration time for training/research and support
-for E2E production ML pipelines when you're ready.
 
 **GETTING STARTED?** Follow the :ref:`quickstart guide<quickstart:Quickstart>`.
 
@@ -76,7 +74,6 @@ Works With
    schedulers/kubernetes
    schedulers/kubernetes_mcad
    schedulers/slurm
-   schedulers/ray
    schedulers/aws_batch
    schedulers/aws_sagemaker
    schedulers/lsf
@@ -91,14 +88,6 @@ Works With
 
       schedulers/fb/*
 
-.. _Pipelines:
-.. toctree::
-   :maxdepth: 1
-   :caption: Pipelines
-
-   pipelines/kfp
-   pipelines/airflow.md
-
 .. fbcode::
 
    .. toctree::
@@ -116,7 +105,6 @@ Examples
    :caption: Examples
 
    examples_apps/index
-   examples_pipelines/index
 
 
 Components Library
@@ -165,7 +153,6 @@ Reference
    runner
    schedulers
    workspace
-   pipelines
 
 .. toctree::
    :maxdepth: 1