Un-nest tensor_copy_chunk dispatch lambdas (#6074) - #6074
Open
FriedCosey wants to merge 6 commits into
Open
Conversation
Contributor
|
@FriedCosey has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113533821. |
meta-codesync
Bot
changed the title
FriedCosey
force-pushed
the
export-D113533821
branch
from
July 25, 2026 20:38
2be6238 to
62f4bdf
Compare
FriedCosey
pushed a commit
to FriedCosey/FBGEMM
that referenced
this pull request
Jul 25, 2026
Summary: X-link: facebookresearch/FBGEMM#2975 The four tensor copies in tensor_copy_chunk are independent, so dispatch each under its own sibling FBGEMM_DISPATCH_* instead of nesting weights -> indices -> identities -> runtime_meta. Two wins, no behavior change: 1. Readability: the flat structure drops the value_t / index_t / id_t / rm_t aliases that existed only to dodge scalar_t name-shadowing between the nested lambdas -- each copy now just uses scalar_t and reads top-to-bottom instead of 3 levels deep. 2. Fewer template instantiations: nesting stamps each inner copy once per outer type (multiplicative, ~4 x 2); siblings stamp each once per its own type set (additive, 4 + 2) -> smaller binary, faster compile. Differential Revision: D113533821
added 6 commits
July 30, 2026 16:21
Summary: X-link: facebookresearch/FBGEMM#3004 The no-fbcode streamer test compiled the header with FBGEMM_FBCODE off but linked the always-flag-on :raw_embedding_streamer, so the flag-on constructor overran the smaller flag-off object layout → ASan heap-buffer-overflow (pre-existing latent bug). Fix: link the test against a new -UFBGEMM_FBCODE build, :raw_embedding_streamer_no_fbcode, so its header view matches the library layout; switch the test include to the short form so it resolves via fbgemm_gpu's include/ dir instead of the manual pin to the flag-on lib. Differential Revision: D114193888
Summary: X-link: facebookresearch/FBGEMM#2968 Speed up the RES C++ streamer's device->host copy by parallelizing it. Previously stream() copied all updated rows to CPU on a single thread and enqueued one item. This replaces that serial copy with a chunked copy across up to `kNumCopyThreads` threads: tensor_copy_chunk() copies a [start, end) row range, and the per-thread tiling is factored into a pure, unit-testable computeChunkRanges(). Differential Revision: D113594180
Summary: X-link: facebookresearch/FBGEMM#2971 Speed up the RES C++ streamer's ship stage by parallelizing it. Previously a single stream thread drained the queue, so the setEmbeddings RPCs to the PS ran one at a time. This diff replaces it with kNumConsumerThreads consumers on a folly::UMPMCQueue that ship concurrently. Reviewed By: chouxi Differential Revision: D113599253
…#6075) Summary: X-link: meta-pytorch/torchrec#4464 X-link: facebookresearch/FBGEMM#2976 Make the RES streamer's chunk size and ship/copy thread counts tunable from config instead of compile-time constants. Make the RES streamer's chunk size and ship/copy thread counts tunable from config instead of compile-time constants. Previously `kChunkSize` / `kNumConsumerThreads` / `kNumCopyThreads` were constexpr in `raw_embedding_streamer.cpp`, so tuning them meant a base rebuild. This turns them into `RawEmbeddingStreamer` ctor params -- `res_chunk_size` / `res_num_consumers` / `res_num_copy_threads` -- appended at the END of the ctor and torchbind `init<>` lists so existing 7-arg callers are unaffected. Defaults stay 500000/8/4, so behavior is unchanged until overridden. The config plumbing that feeds them (`TableBatchedEmbeddingConfig` -> `fused_params`, mirroring `res_store_shards`) lands in D114283366. Differential Revision: D113532922
Summary: Frontend (Python) half of the C++/Python split of D113532922 (FBGEMM guide: backend and frontend ship in separate FBPKGs, so they land as separate diffs, backend first). Plumbs three RES knobs -- `res_chunk_size`, `res_num_consumers`, `res_num_copy_threads` -- from MVAI config to the `RawEmbeddingStreamer` ctor, mirroring the existing `res_store_shards` path. Args are appended at the end (append-only) and defaulted in C++, and each is only forwarded when explicitly set, so behavior is unchanged until a model opts in. Land after D113532922 has merged into the `training_platform` FBPKG. Differential Revision: D114283366
Summary: X-link: facebookresearch/FBGEMM#2975 The four tensor copies in tensor_copy_chunk are independent, so dispatch each under its own sibling FBGEMM_DISPATCH_* instead of nesting weights -> indices -> identities -> runtime_meta. Two wins, no behavior change: 1. Readability: the flat structure drops the value_t / index_t / id_t / rm_t aliases that existed only to dodge scalar_t name-shadowing between the nested lambdas -- each copy now just uses scalar_t and reads top-to-bottom instead of 3 levels deep. 2. Fewer template instantiations: nesting stamps each inner copy once per outer type (multiplicative, ~4 x 2); siblings stamp each once per its own type set (additive, 4 + 2) -> smaller binary, faster compile. Reviewed By: chouxi Differential Revision: D113533821
FriedCosey
force-pushed
the
export-D113533821
branch
from
July 30, 2026 23:22
62f4bdf to
6ae659a
Compare
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
1 participant
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: https://github.com/facebookresearch/FBGEMM/pull/2975
The four tensor copies in tensor_copy_chunk are independent, so dispatch each under its own sibling FBGEMM_DISPATCH_* instead of nesting weights -> indices -> identities -> runtime_meta.
Two wins, no behavior change:
Reviewed By: chouxi
Differential Revision: D113533821