Enable More General Comms Bandwidth Estimates #2905
Closed
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D73229318 |
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
8f6e4a4 to
e0330a5
Compare
|
This pull request was exported from Phabricator. Differential Revision: D73229318 |
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Pull Request resolved: pytorch#2905 Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
e0330a5 to
fe13806
Compare
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
fe13806 to
4d0da7d
Compare
|
This pull request was exported from Phabricator. Differential Revision: D73229318 |
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Pull Request resolved: pytorch#2905 Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
4d0da7d to
44bcc3d
Compare
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
44bcc3d to
62d0c53
Compare
|
This pull request was exported from Phabricator. Differential Revision: D73229318 |
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 22, 2025
Summary: Pull Request resolved: pytorch#2905 Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
62d0c53 to
a8d6855
Compare
gregmacnamara
added a commit
to gregmacnamara/torchrec
that referenced
this pull request
Apr 23, 2025
Summary: Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
gregmacnamara
force-pushed
the
export-D73229318
branch
from
a8d6855 to
b1d43ca
Compare
Summary: Pull Request resolved: pytorch#2905 Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates. Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware. In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size. For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth. Reviewed By: iamzainhuda Differential Revision: D73229318
|
This pull request was exported from Phabricator. Differential Revision: D73229318 |
gregmacnamara
force-pushed
the
export-D73229318
branch
from
b1d43ca to
c63b22d
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:
Optimizing sharding plans relies on good perf estimates which in turn relies on good bandwidth estimates.
Bandwidth of a collective communication depends on many factors including the world size, type of collective (all to all, reduce scatter, ...), and hardware.
In this change, we introduce an abstract class GeneralizedCommsBandwidths that implements a get_bw function. Users can subclass this and define the get_bw appropriately for the hardware or network setup being used. If passed to the Topology, it will be called from perf estimators and will enable setting correct configurations of bandwidth depending on hardware, collective, world_size. We prefer to define this as a class and function as opposed to a dictionary or other similar data structure because we want to enable users to specify bandwidth as precisely as is useful to them. For example, specifying a different bandwidth for every different worldsize rather than only <local_world_size or >=local_world_size.
For Topology objects, the comms_bandwidths argument is optional so we do not change existing behavior and perf estimates do not change. If not passed, we create a GeneralizedCommsBandwidth object with the passed values for inter and intra node comms bw and that replicates the existing behavior of perf estimators. We re-define the signature of perf estimators downstream to ensure consistency and clarity for users so that there is a consistent comms bandwidth.
Reviewed By: iamzainhuda
Differential Revision: D73229318