[DP] change device mesh dim naming convention to make it more consistent #720
Conversation
This file contains 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
[ghstack-poisoned]
This was referenced Dec 5, 2024
Merged
facebook-github-bot
added
the
CLA Signed
…ore consistent" [ghstack-poisoned]
…ore consistent" **Summary** This PR improves the design of DeviceMesh hierarchy in torchtitan. Now, we define all device meshes except `world_mesh` into 2 categories: 1. Basic mesh: those meshes defined in job `.toml` file by users. This include `pp` (`pipeline_parallel_degree`), `dp_replicate` (`data_parallel_replicate_degree`), `dp_shard` (`data_parallel_shard_degree`), `tp` (`tensor_parallel_degree`), and `cp`(`context_parallel_degree`). 2. Synthesized mesh (or called "derived mesh"): meshes that are synthesized from basic mesh by `_flatten()`. If the mesh in synthesized from a single mesh, then it is equivalent to aliasing. So far we utilize 2 synthesized meshes: `dp` and `dp_shard_cp`. The `dp` mesh is used for data loading and the `dp_shard_cp` mesh is used for model params sharding. **Test** CI [ghstack-poisoned]
fegin
approved these changes
Dec 5, 2024
…ion to make it more consistent" **Summary** This PR improves the design of DeviceMesh hierarchy in torchtitan. Now, we define all device meshes except `world_mesh` into 2 categories: 1. Basic mesh: those meshes defined in job `.toml` file by users. This include `pp` (`pipeline_parallel_degree`), `dp_replicate` (`data_parallel_replicate_degree`), `dp_shard` (`data_parallel_shard_degree`), `tp` (`tensor_parallel_degree`), and `cp`(`context_parallel_degree`). 2. Synthesized mesh (or called "derived mesh"): meshes that are synthesized from basic mesh by `_flatten()`. If the mesh in synthesized from a single mesh, then it is equivalent to aliasing. So far we utilize 2 synthesized meshes: `dp` and `dp_shard_cp`. The `dp` mesh is used for data loading and the `dp_shard_cp` mesh is used for model params sharding. **Test** CI [ghstack-poisoned]
…ore consistent" **Summary** This PR improves the design of DeviceMesh hierarchy in torchtitan. Now, we define all device meshes except `world_mesh` into 2 categories: 1. Basic mesh: those meshes defined in job `.toml` file by users. This include `pp` (`pipeline_parallel_degree`), `dp_replicate` (`data_parallel_replicate_degree`), `dp_shard` (`data_parallel_shard_degree`), `tp` (`tensor_parallel_degree`), and `cp`(`context_parallel_degree`). 2. Synthesized mesh (or called "derived mesh"): meshes that are synthesized from basic mesh by `_flatten()`. If the mesh in synthesized from a single mesh, then it is equivalent to aliasing. So far we utilize 2 synthesized meshes: `dp` and `dp_shard_cp`. The `dp` mesh is used for data loading and the `dp_shard_cp` mesh is used for model params sharding. **Test** CI [ghstack-poisoned]
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.
Stack from ghstack (oldest at bottom):
Summary
This PR improves the design of DeviceMesh hierarchy in torchtitan. Now, we define all device meshes except
world_meshinto 2 categories:.tomlfile by users. This includepp(pipeline_parallel_degree),dp_replicate(data_parallel_replicate_degree),dp_shard(data_parallel_shard_degree),tp(tensor_parallel_degree), andcp(context_parallel_degree)._flatten(). If the mesh in synthesized from a single mesh, then it is equivalent to aliasing. So far we utilize 2 synthesized meshes:dpanddp_shard_cp. Thedpmesh is used for data loading and thedp_shard_cpmesh is used for model params sharding.Test
CI