Skip to content

Conversation

shawngu-quic
Copy link
Contributor

Added redesigned moe-mxfp4 kernels optimized for Adreno:

  • Added pref-transpose for experts, fused with SOA convert kernel.
  • Preprocess router table on CPU side.
  • Separated decoding and prefill new kernels for MoE-mxfp4.

Achieved large perf uplift for prefill, especially for long prompts.

@github-actions github-actions bot added ggml changes relating to the ggml tensor library for machine learning OpenCL Issues specific to the OpenCL backend labels Oct 15, 2025
@shawngu-quic shawngu-quic force-pushed the moe-mxfp4-trans-reorder branch 2 times, most recently from 0fb80c7 to 7e84cc9 Compare October 17, 2025 19:13
@shawngu-quic shawngu-quic force-pushed the moe-mxfp4-trans-reorder branch from 7e84cc9 to 61dedfa Compare October 17, 2025 20:34
Copy link
Collaborator

@max-krasnyansky max-krasnyansky left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice bump in perf on Gen5 devices!

@max-krasnyansky
Copy link
Collaborator

@lhez
Copy link
Collaborator

lhez commented Oct 17, 2025

@shawngu-quic There are also some compilation warnings that have to be fixed.

@max-krasnyansky max-krasnyansky merged commit 8138785 into ggml-org:master Oct 18, 2025
69 of 70 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ggml changes relating to the ggml tensor library for machine learning OpenCL Issues specific to the OpenCL backend

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants