[spmd_types] decoder sharding configs in spmd.* - #3471
Merged
Conversation
This was referenced May 31, 2026
meta-cla
Bot
added
the
CLA Signed
This was referenced Jun 1, 2026
pianpwk
changed the title
pianpwk
marked this pull request as ready for review
June 3, 2026 21:31
pianpwk
added a commit
that referenced
this pull request
Jun 9, 2026
Adding this to allow both spmd types specs, and DTensor-style placements. We're adding something like D107576760 in spmd_types, but in the meantime, because non-llama3 models are still in DTensor placement form, I think we have 2 options: 1) keep this as a hybrid spmd/DTensor container (e.g. moe_sharding.py calls `dense_activation_placement(tp=Replicate())`, and migrate to SpmdLayout once all models are off full_dtensor. 2) since we're adding a spmd->DTensor translation shim anyways, migrate all models from `Replicate/Shard/Partial` to `spmd.R/S/P` immediately, use SpmdLayout, and go deal with I/R distinctions later. This should be ok because non-llama3 models won't support spmd backend yet. Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * #3278 * #3472 * #3468 * #3471 * #3253 * #3587 * #3586 * __->__ #3501
pianpwk
added a commit
that referenced
this pull request
Jun 9, 2026
pianpwk
added a commit
that referenced
this pull request
Jun 9, 2026
Splitting up infra PR. This adds: - _shard_spmd_state: util for distributing param/buffer onto mesh in SpmdLayout format, supporting PartitionSpec shard order - redistribute_spmd_per_axis: module boundary, SpmdLayout-based redistribution handler - _validate_spmd_redistributions: helper for validating it resolves to just one spmd.redistribute call - that's all titan will support Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * #3278 * #3472 * #3468 * #3471 * #3253 * __->__ #3587
wwwjn
pushed a commit
that referenced
this pull request
Jun 10, 2026
Adding this to allow both spmd types specs, and DTensor-style placements. We're adding something like D107576760 in spmd_types, but in the meantime, because non-llama3 models are still in DTensor placement form, I think we have 2 options: 1) keep this as a hybrid spmd/DTensor container (e.g. moe_sharding.py calls `dense_activation_placement(tp=Replicate())`, and migrate to SpmdLayout once all models are off full_dtensor. 2) since we're adding a spmd->DTensor translation shim anyways, migrate all models from `Replicate/Shard/Partial` to `spmd.R/S/P` immediately, use SpmdLayout, and go deal with I/R distinctions later. This should be ok because non-llama3 models won't support spmd backend yet. Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * #3278 * #3472 * #3468 * #3471 * #3253 * #3587 * #3586 * __->__ #3501
wwwjn
pushed a commit
that referenced
this pull request
Jun 10, 2026
wwwjn
pushed a commit
that referenced
this pull request
Jun 10, 2026
Splitting up infra PR. This adds: - _shard_spmd_state: util for distributing param/buffer onto mesh in SpmdLayout format, supporting PartitionSpec shard order - redistribute_spmd_per_axis: module boundary, SpmdLayout-based redistribution handler - _validate_spmd_redistributions: helper for validating it resolves to just one spmd.redistribute call - that's all titan will support Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * #3278 * #3472 * #3468 * #3471 * #3253 * __->__ #3587
saforem2
pushed a commit
to saforem2/torchtitan
that referenced
this pull request
Jun 10, 2026
Splitting up infra PR. This adds: - _shard_spmd_state: util for distributing param/buffer onto mesh in SpmdLayout format, supporting PartitionSpec shard order - redistribute_spmd_per_axis: module boundary, SpmdLayout-based redistribution handler - _validate_spmd_redistributions: helper for validating it resolves to just one spmd.redistribute call - that's all titan will support Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * pytorch#3278 * pytorch#3472 * pytorch#3468 * pytorch#3471 * pytorch#3253 * __->__ pytorch#3587
tianyu-l
approved these changes
Jun 11, 2026
| """ | ||
| q_placements: SpmdLayout = dense_activation_placement(tp=spmd.S(2)) | ||
| kv_src_placements: SpmdLayout = dense_activation_placement(tp=spmd.S(2)) | ||
| kv_placements: SpmdLayout = dense_activation_placement(tp=spmd.S(2), cp=spmd.R) |
Contributor
There was a problem hiding this comment.
Suggested change
| kv_placements: SpmdLayout = dense_activation_placement(tp=spmd.S(2), cp=spmd.R) | |
| kv_dst_placements: SpmdLayout = dense_activation_placement(tp=spmd.S(2), cp=spmd.R) |
| }, | ||
| in_dst_shardings={ | ||
| "q": q_placements, | ||
| "k": kv_placements, |
| @@ -19,8 +19,8 @@ | |||
| def dense_param_placement(*, tp: spmd.PerMeshAxisSpmdType) -> SpmdLayout: | |||
Contributor
There was a problem hiding this comment.
all these should be renamed to _layout, can be its own PR
This was referenced Jun 11, 2026
pianpwk
added a commit
that referenced
this pull request
Jun 11, 2026
Main infra PR for spmd_types + titan. `spmd_types.py`: - Adds a thread-local DeviceMesh stack, and current_mesh(), set_current_mesh() helpers, to be set and accessed at model init (weight parallelize) and runtime (PG access for collectives) - Various helpers for spmd_types : module-boundary redistributions, spmd -> DTensor placement translation (for full_dtensor backend), `mesh_size(axis_name)` helper that returns > 1 when mesh is set & axis is active. Some of this will be moved to spmd_types in near-term, see comments. `parallel_dims.py`: - For spmd backend, we need 2 views over the world mesh: [pp, dp, cp, tp] for typechecking, and [pp, dp_replicate, dp_shard, cp, tp] to pass the unfolded fsdp axes to fully_shard, via DataParallelMeshDims. So we hold both the full-DTensor-style dense mesh, as well as the "typechecking" mesh. `module.py`: module.parallelize() paths for spmd backend: weight init, local SPMD drop in, input/output redistribution. `trainer.py/utils.py`: typechecking context, input annotation, spmd set_current_spmd_mesh, PP + typechecking raises a hard error. Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.12.0) (oldest at bottom): * #3278 * #3472 * #3632 * #3631 * #3468 * #3471 * __->__ #3253
pianpwk
added a commit
that referenced
this pull request
Jun 11, 2026
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
switches decoder_sharding.py and llama3/sharding.py to spmd.* types.
LocalMapConfig(in_grad_placements=...)carries info only used by default/full_dtensor backends; spmd_types backend just checks for presence of config, to switch to local SPMDStack from ghstack (oldest at bottom):