Skip to content

Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)" - #28551

Merged
pwilkin merged 1 commit into
masterfrom
revert-mmq-moe-ncols
Sep 7, 2026
Merged

Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)"#28551
pwilkin merged 1 commit into
masterfrom
revert-mmq-moe-ncols

Conversation

@pwilkin

@pwilkin pwilkin commented Sep 7, 2026

Copy link
Copy Markdown
Member

This reverts commit 0c96345, per discussion in #24546

@pwilkin
pwilkin requested a review from a team as a code owner September 7, 2026 14:20
@pwilkin

pwilkin commented Sep 7, 2026

Copy link
Copy Markdown
Member Author

@JohannesGaessler @CISC

@github-actions github-actions Bot added ggml changes relating to the ggml tensor library for machine learning CUDA Related to the CUDA backend labels Sep 7, 2026
@pwilkin
pwilkin merged commit e71b805 into master Sep 7, 2026
9 of 21 checks passed
Patt92 pushed a commit to Patt92/llama.cpp that referenced this pull request Sep 7, 2026
Drops the bounded multi-backend scheduler split from the feature list: upstream
992cb50 (ggml-org#28387) deleted the input-count check outright, so split boundaries
can no longer drift with the growing inputs_capacity and the local change has
nothing left to do.

Records that upstream independently built and then reverted routed-MoE MMQ tile
sizing (ggml-org#24546, reverted by ggml-org#28551). The revert was about where the logic lives,
not whether it works -- the objection was that it changed kernel configurations
instead of deciding on the host -- and that PR excluded gfx1151 outright, so its
own gfx1151 operator measurements remain the best evidence for the idea here.
x1250 pushed a commit to x1250/llama.cpp that referenced this pull request Sep 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CUDA Related to the CUDA backend ggml changes relating to the ggml tensor library for machine learning

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants