openvino
[GPU] MoE 3 GeMM: separate Router Subgraph and MoE body kernels
#35691
Open
Overview
Commits
14
Changes
View On GitHub

[GPU] MoE 3 GeMM: separate Router Subgraph and MoE body kernels #35691

v-Golubev wants to merge 14 commits into openvinotoolkit:master from v-Golubev:vg/transformations/moe_extract_router_to_separate_node
v-Golubev
v-Golubev Tests extending
f00f4c5c
v-Golubev style
7411b115
v-Golubev [Transformations] ConvertTiledMoeBlockTo3GatherMatmuls: Gelu activati…
172ba935
v-Golubev Add MoERouterFused op/primitive infrastructure (no behavior change)
d060d71a
v-Golubev Add standalone MoERouterFused accuracy tests (8 tests pass)
f5c773df
v-Golubev Split routing logic from MOE3Gemm into MoERouterFused
2f136d71
v-Golubev Update accuracy tests for two-primitive router+MOE flow
9f7c122d
v-Golubev Remove dead routing kernel classes and fix functional test assertion
584f907c
v-Golubev Remove wasted topk internal buffer allocations from MOE3Gemm
3bce0a91
v-Golubev Update primitive header comments to reflect new input layout
881a54a7
v-Golubev Unify MOE3GemmFusedCompressed into MOECompressed op
5a328a82
v-Golubev Replace FuseMOE3GemmCompressed with FuseMoERouter
325a937c
v-Golubev cleanup
8be5d029
v-Golubev cleanup
5a2b056a
v-Golubev v-Golubev requested a review 1 day ago
v-Golubev v-Golubev requested a review 1 day ago
v-Golubev v-Golubev requested a review 1 day ago
v-Golubev v-Golubev added WIP
github-actions github-actions added category: GPU
github-actions github-actions added category: transformations

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
No one assigned
Labels
category: GPU category: transformations WIP
Milestone
No milestone
Terms Of Use Privacy Policy Refund Policy Impressum