[IlyasMoutawwakil]

Depends on #49606.

Gemma 4, DiffusionGemma, GraniteMoE-SWA, HY-V4, Inkling and MiMo-V2-Flash still defaulted to the router-masked EP plan: ep_router on the router and moe_tp_experts on the experts. Their experts take (hidden_states, top_k_index, top_k_weights) like every other experts class. So their EP plans move to ep_dispatch_experts.

With ep_dispatch_experts, these models:

• run the all-reduce at inference when ep_size == tp_size, as before (Expert-parallel implementations for ep_dispatch_experts: all-reduce at inference, all-to-all in training #49429);
• use token dispatch in training;
• now also support ep_size > tp_size.

Changes

• The configs of the six models, and the modular files they come from, drop the ep_router rule and use "layers.*...experts": "ep_dispatch_experts".
• Qwen3-MoE, Mellum and Beam drop the plan comment that pointed to the router-masked plan.
• The docs no longer list Gemma 4 among the models that default to masking.

Still on ep_router after this PR:

• MXFP4 GPT-OSS, whose experts route themselves. 🚨 Run MXFP4 GPT-OSS experts through ep_dispatch_experts #49613 moves it.
• Llama 4, whose EP plan never matched an expert-parallel mode, so it was never applied.

#49615 then removes ep_router.

Tests

Test	Result
TP and EP tests of the six models plus Beam and Qwen3-MoE, on CPU	pass, apart from Inkling tests that fail on the parent branch too (scipy is missing in the test env)
tests/expert_parallel, 2 ranks on gloo	8 passed
Modular conversion and copies checks	clean

🤖 Generated with Claude Code

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

Pages changed in this PR:

• en/expert_parallelism

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: beam, diffusion_gemma, gemma4, hy_v4, inkling, mellum, mimo_v2_flash, qwen3_moe