TP: fix delayed AllReduce + zero-sized slices#22489
Merged
JohannesGaessler merged 1 commit intoApr 29, 2026
Merged
Conversation
github-actions
Bot
added
the
ggml
am17an
approved these changes
Apr 28, 2026
Contributor
|
Tested and verified. The issue is resolved. |
gaugarg-nv
approved these changes
Apr 29, 2026
rsenthilkumar6
pushed a commit
to rsenthilkumar6/llama.cpp
that referenced
this pull request
May 1, 2026
samuraieng
pushed a commit
to samuraieng/llama.cpp
that referenced
this pull request
May 6, 2026
ljubomirj
pushed a commit
to ljubomirj/llama.cpp
that referenced
this pull request
May 6, 2026
meh
pushed a commit
to meh/llama.cpp
that referenced
this pull request
May 10, 2026
baramofme
pushed a commit
to baramofme/llama-cpp-turboquant
that referenced
this pull request
May 23, 2026
carlosfundora
pushed a commit
to carlosfundora/llama.cpp-1-bit-turbo
that referenced
this pull request
May 24, 2026
(cherry picked from commit 739393b)
winstonma
pushed a commit
to winstonma/llama.cpp
that referenced
this pull request
May 27, 2026
fewtarius
pushed a commit
to fewtarius/llama.cpp
that referenced
this pull request
May 30, 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
4 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #22391 .
The problem is that k-quants have a block size of 256 vs. the size of a single expert at 512. So for 3+ GPUs one of them ends up with a zero-sized slice. This would normally not be an issue since a zero-sized slice is supported; the corresponding nodes are disabled and the backend participates in the following AllReduce with a zeroed out buffer in order to receive the results of other backends. However, the interaction of a zero-sized slice and a delayed AllReduce for better MoE performance does not work correctly. For those the range of disabled nodes needs to be extended, otherwise one of the backends will have garbage data prior to the AllReduce.
Using 3x RTX 4090 the Qwen 3.6 q4_K_M PPL on the first 512 tokens of Wikitext is 4.1590 for
-sm layer, for-sm tensoron master it's 8.3604, for-sm tensorwith this PR it's4.1554.Requirements