Skip to content

Gate MoE fusion passes to XeLP+ and align pipeline comments#5

Draft
Copilot wants to merge 2 commits into
oom/fixingfrom
copilot/add-condition-for-code-snippet
Draft

Gate MoE fusion passes to XeLP+ and align pipeline comments#5
Copilot wants to merge 2 commits into
oom/fixingfrom
copilot/add-condition-for-code-snippet

Conversation

Copy link
Copy Markdown

Copilot AI commented Jun 2, 2026

Details:

  • Problem summary: The MoE fusion block in transformations_pipeline.cpp needed to execute conditionally by GPU architecture, and nearby comments no longer matched the gated behavior.
  • Behavior change (MoE fusion gating): Restrict MoeOpFusion, FuseMOESharedExpert, and FuseMOE3GemmCompressed registration to xe_lp and newer.
  • Documentation-in-code update: Clarified comments to distinguish always-on MoE conversion passes from XeLP+-only MoE fusion passes.
  • Snippet:
    if (!disable_moe_opt && device_info.arch >= cldnn::gpu_arch::xe_lp) {
        const bool has_batch_dim = !is_pa;
        manager.register_pass<ov::pass::MoeOpFusion>(has_batch_dim);
        manager.register_pass<ov::intel_gpu::FuseMOESharedExpert>();
        manager.register_pass<ov::intel_gpu::FuseMOE3GemmCompressed>();
    }

Tickets:

  • N/A

AI Assistance:

  • AI assistance used: yes
  • AI was used to implement the conditional guard and update surrounding comments; human validation was performed via code review of the affected section and consistency checks against nearby transformation-pipeline logic.

Copilot AI changed the title Gate MoE fusion registrations to XeLP+ GPUs Gate MoE fusion passes to XeLP+ and align pipeline comments Jun 2, 2026
Copilot AI requested a review from peterchen-intel June 2, 2026 07:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants