]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
Tensor-parallel: Fix delayed AllReduce on Gemma-4 MoE (llama/22129)
authorGaurav Garg <redacted>
Mon, 20 Apr 2026 16:25:39 +0000 (21:55 +0530)
committerGeorgi Gerganov <redacted>
Thu, 30 Apr 2026 08:29:14 +0000 (11:29 +0300)
commit239c5c86c30d36249e3479459914c4eb24958f19
treeecb4894f2b95c555d493543761a2c610316278c6
parent6429023e5f48c37b03e4903bf2bab8ef875b244f
Tensor-parallel: Fix delayed AllReduce on Gemma-4 MoE (llama/22129)

* Fix delayed AllReduce on Gemma-4 MoE

Skip forward past nodes that don't consume the current one, and allow a chain of MULs.

* Check for all sources before skipping nodes

* Address review comments
ggml/src/ggml-backend-meta.cpp