Repository navigation
Conversation
meta-cla
Bot
added
the
CLA Signed
|
@sunguobao has exported this pull request. If you are a Meta employee, you can view the originating Diff in D126865303. |
…esearch#224) Summary: X-link: meta-pytorch/torchrec#4928 Prepare for removing `rccl` and `rcclx` from TorchComms by removing their external build and runtime consumers. Retire `DMPCollection`'s built-in RCCLX sharded-relay integration and its internal PyTorch adapter, utility module, unit tests, and AMD benchmark. The default `replica_pg.allreduce_coalesced` path handles reductions by dtype, preserves `ReduceOp.SUM`/`ReduceOp.AVG` options, and waits for completion. Custom reducer hooks remain available. The `use_sharded_relay` constructor option is removed, and TorchRec no longer reads `NCCL_SHARDED_RELAY_MODE_ENABLE`. Remove PARAM's RCCLX extension dependency on AMD and its advertised `rccl`/`rcclx` TorchComms backend choices. Reject AMD selection of the `pytorch-torchcomms` stack before device or communicator initialization, directing users to `--nw-stack pytorch-dist`. Native PyTorch distributed benchmarks on AMD continue to use their existing underlying RCCL library dependencies. Move the RCCLX PARAM smoke test (`rcclx_comms_test`) off the TorchComms `rcclx` backend, which this change removes from PARAM, onto `--nw-stack pytorch-dist --backend nccl`. On AMD, c10d's `nccl` backend links the RCCLX build that the test's `rcclx-dev` modifier selects, so the test still runs an all-reduce on RCCLX. Remove obsolete TorchComms RCCL/RCCLX build-all entries, Buck auto-mode mappings, and the Infer exclusion. Update the `DperExporter` packaging test's module list to omit the deleted utility. Behavior change: this retires the built-in sharded-relay optimization. Workloads relying on it can lose throughput and need review before landing. Existing packaged copies of the old TorchRec integration load its dependencies dynamically, so they keep loading and fall back to the default all-reduce. This draft covers external consumers and tooling. TorchComms backend implementation and OSS build/packaging removal follow separately. Review `model_parallel.py` and PARAM's backend selection changes first, followed by the build rules and deletion list. Prepared with Codex and Claude Code assistance. Differential Revision: D126865303
meta-codesync
Bot
changed the title
rccl and rcclxrccl and rcclx (#224)
sunguobao
force-pushed
the
export-D126865303
branch
from
October 23, 2026 03:13
449df26 to
3fa1f4f
Compare
sunguobao
added a commit
to sunguobao/torchrec
that referenced
this pull request
Oct 23, 2026
…rch#4928) Summary: X-link: facebookresearch/param#224 Prepare for removing `rccl` and `rcclx` from TorchComms by removing their external build and runtime consumers. Retire `DMPCollection`'s built-in RCCLX sharded-relay integration and its internal PyTorch adapter, utility module, unit tests, and AMD benchmark. The default `replica_pg.allreduce_coalesced` path handles reductions by dtype, preserves `ReduceOp.SUM`/`ReduceOp.AVG` options, and waits for completion. Custom reducer hooks remain available. The `use_sharded_relay` constructor option is removed, and TorchRec no longer reads `NCCL_SHARDED_RELAY_MODE_ENABLE`. Remove PARAM's RCCLX extension dependency on AMD and its advertised `rccl`/`rcclx` TorchComms backend choices. Reject AMD selection of the `pytorch-torchcomms` stack before device or communicator initialization, directing users to `--nw-stack pytorch-dist`. Native PyTorch distributed benchmarks on AMD continue to use their existing underlying RCCL library dependencies. Move the RCCLX PARAM smoke test (`rcclx_comms_test`) off the TorchComms `rcclx` backend, which this change removes from PARAM, onto `--nw-stack pytorch-dist --backend nccl`. On AMD, c10d's `nccl` backend links the RCCLX build that the test's `rcclx-dev` modifier selects, so the test still runs an all-reduce on RCCLX. Remove obsolete TorchComms RCCL/RCCLX build-all entries, Buck auto-mode mappings, and the Infer exclusion. Update the `DperExporter` packaging test's module list to omit the deleted utility. Behavior change: this retires the built-in sharded-relay optimization. Workloads relying on it can lose throughput and need review before landing. Existing packaged copies of the old TorchRec integration load its dependencies dynamically, so they keep loading and fall back to the default all-reduce. This draft covers external consumers and tooling. TorchComms backend implementation and OSS build/packaging removal follow separately. Review `model_parallel.py` and PARAM's backend selection changes first, followed by the build rules and deletion list. Prepared with Codex and Claude Code assistance. Differential Revision: D126865303
|
This pull request has been merged in b05be06. |
meta-codesync Bot
pushed a commit
to meta-pytorch/torchrec
that referenced
this pull request
Oct 23, 2026
Summary: Pull Request resolved: #4928 X-link: facebookresearch/param#224 Prepare for removing `rccl` and `rcclx` from TorchComms by removing their external build and runtime consumers. Retire `DMPCollection`'s built-in RCCLX sharded-relay integration and its internal PyTorch adapter, utility module, unit tests, and AMD benchmark. The default `replica_pg.allreduce_coalesced` path handles reductions by dtype, preserves `ReduceOp.SUM`/`ReduceOp.AVG` options, and waits for completion. Custom reducer hooks remain available. The `use_sharded_relay` constructor option is removed, and TorchRec no longer reads `NCCL_SHARDED_RELAY_MODE_ENABLE`. Remove PARAM's RCCLX extension dependency on AMD and its advertised `rccl`/`rcclx` TorchComms backend choices. Reject AMD selection of the `pytorch-torchcomms` stack before device or communicator initialization, directing users to `--nw-stack pytorch-dist`. Native PyTorch distributed benchmarks on AMD continue to use their existing underlying RCCL library dependencies. Move the RCCLX PARAM smoke test (`rcclx_comms_test`) off the TorchComms `rcclx` backend, which this change removes from PARAM, onto `--nw-stack pytorch-dist --backend nccl`. On AMD, c10d's `nccl` backend links the RCCLX build that the test's `rcclx-dev` modifier selects, so the test still runs an all-reduce on RCCLX. Remove obsolete TorchComms RCCL/RCCLX build-all entries, Buck auto-mode mappings, and the Infer exclusion. Update the `DperExporter` packaging test's module list to omit the deleted utility. Behavior change: this retires the built-in sharded-relay optimization. Workloads relying on it can lose throughput and need review before landing. Existing packaged copies of the old TorchRec integration load its dependencies dynamically, so they keep loading and fall back to the default all-reduce. This draft covers external consumers and tooling. TorchComms backend implementation and OSS build/packaging removal follow separately. Review `model_parallel.py` and PARAM's backend selection changes first, followed by the build rules and deletion list. Prepared with Codex and Claude Code assistance. Reviewed By: srinathb-meta Differential Revision: D126865303 fbshipit-source-id: 632c206381e3f035dd49486e1e6ed2a6a645e0ee
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
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:
X-link: meta-pytorch/torchrec#4928
Prepare for removing
rcclandrcclxfrom TorchComms by removing their external build and runtime consumers.Retire
DMPCollection's built-in RCCLX sharded-relay integration and its internal PyTorch adapter, utility module, unit tests, and AMD benchmark. The defaultreplica_pg.allreduce_coalescedpath handles reductions by dtype, preservesReduceOp.SUM/ReduceOp.AVGoptions, and waits for completion. Custom reducer hooks remain available. Theuse_sharded_relayconstructor option is removed, and TorchRec no longer readsNCCL_SHARDED_RELAY_MODE_ENABLE.Remove PARAM's RCCLX extension dependency on AMD and its advertised
rccl/rcclxTorchComms backend choices. Reject AMD selection of thepytorch-torchcommsstack before device or communicator initialization, directing users to--nw-stack pytorch-dist. Native PyTorch distributed benchmarks on AMD continue to use their existing underlying RCCL library dependencies.Move the RCCLX PARAM smoke test (
rcclx_comms_test) off the TorchCommsrcclxbackend, which this change removes from PARAM, onto--nw-stack pytorch-dist --backend nccl. On AMD, c10d'sncclbackend links the RCCLX build that the test'srcclx-devmodifier selects, so the test still runs an all-reduce on RCCLX.Remove obsolete TorchComms RCCL/RCCLX build-all entries, Buck auto-mode mappings, and the Infer exclusion. Update the
DperExporterpackaging test's module list to omit the deleted utility.Behavior change: this retires the built-in sharded-relay optimization. Workloads relying on it can lose throughput and need review before landing. Existing packaged copies of the old TorchRec integration load its dependencies dynamically, so they keep loading and fall back to the default all-reduce. This draft covers external consumers and tooling. TorchComms backend implementation and OSS build/packaging removal follow separately.
Review
model_parallel.pyand PARAM's backend selection changes first, followed by the build rules and deletion list.Prepared with Codex and Claude Code assistance.
Differential Revision: D126865303