Tune max segment length per cta in triton table batched embeddings, and expose the param via cli #5270
+28
−10
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
OmarPavel
pushed a commit
to OmarPavel/FBGEMM
that referenced
this pull request
Dec 22, 2025
…nd expose the param via cli (pytorch#5270) Summary: see previous diff in the stack Reviewed By: stashuk-olek Differential Revision: D89558538
OmarPavel
force-pushed
the
export-D89558538
branch
from
327c182 to
b0af65b
Compare
…nd expose the param via cli Summary: Tune max segment length per cta in triton table batched embeddings, and expose the param via cli. Improves performance by ~2% on B200 We were using hardcoded 1024, tuned for H100. This change sets the default for B200 to 4096 (after testing smaller and larger values). It also exposes this flag via command line, if used together with --no-deterministic, you can add --max-cta-segment-length 4096 to change this via cmdline (for whatever cuda device you're building for). Tested at 512/1024/2048/4096/8192 values, 4096 outperforms at the two tested batch sizes of 256 and 512 ``` Embedding Dim: 256 ┌─────────────────────┬───────────────────┬──────────────┐ │ CTA Segment Length │ Backward Time (μs)│ vs 1024 │ ├─────────────────────┼───────────────────┼──────────────┤ │ 512 │ 24,651 │ -4.1% slower │ │ 1024 (default) │ 23,680 │ baseline │ │ 4096 │ 23,118 │ +2.4% faster │ │ 8192 │ 28,698 │ -21.2% slower│ └─────────────────────┴───────────────────┴──────────────┘ Embedding Dim: 512 ┌─────────────────────┬───────────────────┬──────────────┐ │ CTA Segment Length │ Backward Time (μs)│ vs 1024 │ ├─────────────────────┼───────────────────┼──────────────┤ │ 1024 (default) │ 42,067 │ baseline │ │ 2048 │ 41,471 │ +1.4% faster │ │ 4096 │ 41,235 │ +2.0% faster │ │ 8192 │ 41,414 │ +1.6% faster │ └─────────────────────┴───────────────────┴──────────────┘ Reviewed By: stashuk-olek Differential Revision: D89558538
OmarPavel
force-pushed
the
export-D89558538
branch
from
b0af65b to
d543064
Compare
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.
Summary: see previous diff in the stack
Reviewed By: stashuk-olek
Differential Revision: D89558538