(torchx/specs) Allow roles to specify their own workspaces #1139
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
meta-cla
bot
added
the
CLA Signed
|
@kiukchung has exported this pull request. If you are a Meta employee, you can view the originating Diff in D83793199. |
|
@kiukchung has exported this pull request. If you are a Meta employee, you can view the originating Diff in D83793199. |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #1139 +/- ##
==========================================
+ Coverage 91.63% 91.64% +0.01%
==========================================
Files 83 83
Lines 6431 6441 +10
==========================================
+ Hits 5893 5903 +10
Misses 538 538
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
kiukchung
added a commit
to kiukchung/torchx-1
that referenced
this pull request
Oct 3, 2025
…rch#1139) Summary: So far you can only specify the workspace in the runner's API: ``` runner.dryrun(appdef, cfg, workspace=...) ``` in which case the workspace applies to ONLY `role[0]`. This behavior was intentional since multi-role usecases of TorchX typically had a single "main" role that the application owner actually owned and the other roles were prepackaged apps (not part of your project). This is no longer the case with applications such as reenforcement learning where the project encompasses multiple applications (e.g. trainer, generator, etc) therefore we need a more flexible way to specify a workspace per Role. For BC this I'm maintaining the following behavior: 1. If `workspace` is specified as a runner argument then it takes precedence over `role[0].workspace` 2. Non-zero roles (e.g. `role[1], role[2], ...`) are unaffected by the workspace argument. That is their workspace attributes (e.g. `role[1].workspace`) are respected as is. 3. "disabling" workspace (e.g. passing `workspace=None` from the runner argument) can still build a workspace if the role's workspace attribute is not `None`. NOTE: we need to do further optimization for cases where multiple roles have the same "image" and "workspace". In this case we only need to build the image+workspace once. But as it stands we end up building a separate ephemeral per role (even if the ephemeral is the SAME across all the roles). This isn't an issue practically since image builders like Docker are content addressed and caches layers. Reviewed By: AbishekS Differential Revision: D83793199
kiukchung
force-pushed
the
export-D83793199
branch
from
b748dcd to
bedb24a
Compare
kiukchung
added a commit
to kiukchung/torchx-1
that referenced
this pull request
Oct 4, 2025
…rch#1139) Summary: So far you can only specify the workspace in the runner's API: ``` runner.dryrun(appdef, cfg, workspace=...) ``` in which case the workspace applies to ONLY `role[0]`. This behavior was intentional since multi-role usecases of TorchX typically had a single "main" role that the application owner actually owned and the other roles were prepackaged apps (not part of your project). This is no longer the case with applications such as reenforcement learning where the project encompasses multiple applications (e.g. trainer, generator, etc) therefore we need a more flexible way to specify a workspace per Role. For BC this I'm maintaining the following behavior: 1. If `workspace` is specified as a runner argument then it takes precedence over `role[0].workspace` 2. Non-zero roles (e.g. `role[1], role[2], ...`) are unaffected by the workspace argument. That is their workspace attributes (e.g. `role[1].workspace`) are respected as is. 3. "disabling" workspace (e.g. passing `workspace=None` from the runner argument) can still build a workspace if the role's workspace attribute is not `None`. NOTE: we need to do further optimization for cases where multiple roles have the same "image" and "workspace". In this case we only need to build the image+workspace once. But as it stands we end up building a separate ephemeral per role (even if the ephemeral is the SAME across all the roles). This isn't an issue practically since image builders like Docker are content addressed and caches layers. Reviewed By: AbishekS Differential Revision: D83793199
kiukchung
force-pushed
the
export-D83793199
branch
from
bedb24a to
12526d4
Compare
…rch#1139) Summary: So far you can only specify the workspace in the runner's API: ``` runner.dryrun(appdef, cfg, workspace=...) ``` in which case the workspace applies to ONLY `role[0]`. This behavior was intentional since multi-role usecases of TorchX typically had a single "main" role that the application owner actually owned and the other roles were prepackaged apps (not part of your project). This is no longer the case with applications such as reenforcement learning where the project encompasses multiple applications (e.g. trainer, generator, etc) therefore we need a more flexible way to specify a workspace per Role. For BC this I'm maintaining the following behavior: 1. If `workspace` is specified as a runner argument then it takes precedence over `role[0].workspace` 2. Non-zero roles (e.g. `role[1], role[2], ...`) are unaffected by the workspace argument. That is their workspace attributes (e.g. `role[1].workspace`) are respected as is. 3. "disabling" workspace (e.g. passing `workspace=None` from the runner argument) can still build a workspace if the role's workspace attribute is not `None`. NOTE: we need to do further optimization for cases where multiple roles have the same "image" and "workspace". In this case we only need to build the image+workspace once. But as it stands we end up building a separate ephemeral per role (even if the ephemeral is the SAME across all the roles). This isn't an issue practically since image builders like Docker are content addressed and caches layers. Reviewed By: AbishekS Differential Revision: D83793199
kiukchung
force-pushed
the
export-D83793199
branch
from
12526d4 to
0c92e58
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
So far you can only specify the workspace in the runner's API:
in which case the workspace applies to ONLY
role[0].This behavior was intentional since multi-role usecases of TorchX typically had a single "main" role that the application owner actually owned and the other roles were prepackaged apps (not part of your project).
This is no longer the case with applications such as reenforcement learning where the project encompasses multiple applications (e.g. trainer, generator, etc) therefore we need a more flexible way to specify a workspace per Role.
For BC this I'm maintaining the following behavior:
If
workspaceis specified as a runner argument then it takes precedence overrole[0].workspaceNon-zero roles (e.g.
role[1], role[2], ...) are unaffected by the workspace argument. That is their workspace attributes (e.g.role[1].workspace) are respected as is."disabling" workspace (e.g. passing
workspace=Nonefrom the runner argument) can still build a workspace if the role's workspace attribute is notNone.NOTE: we need to do further optimization for cases where multiple roles have the same "image" and "workspace". In this case we only need to build the image+workspace once. But as it stands we end up building a separate ephemeral per role (even if the ephemeral is the SAME across all the roles). This isn't an issue practically since image builders like Docker are content addressed and caches layers.
Reviewed By: AbishekS
Differential Revision: D83793199