[Refactor] Move Step-3.5, GLM5-Next, Dots3, MiniMax-M3 and Qwen3.5 onto ffn_exit

[ch-wan]

This PR is part of a stack (oldest at bottom):

• [Refactor] Decide an FFN exit's completion once and declare the group it owes #41200
• [Refactor] Take a layer's last-layer fact from its scatter-mode plan #41199
• [Refactor] Build prepare_mlp and the layout moves from named steps #41191
• -> [Refactor] Move Step-3.5, GLM5-Next, Dots3, MiniMax-M3 and Qwen3.5 onto ffn_exit #41198
• [Refactor] Leave the FFN reduction to the next layer under attention DP #41197
• [Refactor] Carry a deferred FFN all-reduce as UnreducedOutput and complete it in the next layer without the fused kernel #41196
• [Fix] Stop counting a deferred FFN sum more than once: replicated TP1 shared expert, dense reduce_scatterv #41195
• [Fix] Plan NextN / MTP draft layers as one-layer models and fix the Bailing V2 NextN draft #41194
• [Fix] Complete the all-reduce when the flashinfer fused norm declines a batch #41193

Motivation

Step-3.5, GLM5-Next, Dots3, MiniMax-M3 and Qwen3.5 decided by hand whether to leave their FFN all-reduce to the next layer, from should_fuse_mlp_allreduce_with_next_layer. They published the flags themselves, then wrapped the output or ran postprocess. Moving them onto ffn_exit puts their decision in one place with the other layer-communicator models. They also gain the deferrals the previous PRs added (without the fused kernel, and under attention DP).

Modifications

One commit per model:

• Step-3.5: the MoE layers use ffn_exit. The MoE still skips its internal all-reduce, so one all-reduce covers the MoE and shared-expert outputs. The dense layers keep reducing their own output.
• GLM5-Next: uses ffn_exit. The MHC layers never defer, as before.
• Dots3: uses ffn_exit and still passes the decisions to the MLP / MoE as arguments. Inside ffn_exit the same decisions are published, so the row-parallel layers agree with the arguments.
• MiniMax-M3: uses ffn_exit, and drops the forced in-layer all-reduce of sparse layers at TP > 1. It worked around a runaway without EOS, which came from the last layer's deferred sum reaching the final norm unreduced. [Fix] Stop deferring the last layer's FFN all-reduce in five models #40868 fixed that, and the layer-stack exit now completes any sum the last layer leaves. This is a behavior change: with the fused kernel on, sparse layers now fuse like the dense ones.
• Qwen3.5: FfnExit decides the deferred MoE finalize and the reduction, and FfnExit.finish passes a handoff on, returns the unreduced output, or runs postprocess. The last-layer fallback is gone, and _finish_mlp_output is removed. The MoE block takes ffn_exit.defer_moe_finalize as is: the old check that the MLP input is a tensor is dropped, because prepare_mlp always returns one here.

Accuracy Tests

4 greedy requests × 32 tokens; text, token logprobs and top-5 logprobs compared, decode CUDA graph on.

• Step-3.5-Flash BF16 TP8, H200: fusion on, fusion off, DP8 and DP2 are all bit-identical to the parent. The audit shows the new deferrals (fusion off: unfused deferral; DP: reduce-scatter moved).
• Qwen3.5-122B-A10B-FP8 TP2, H200: fusion on, fusion off, DP2, and DP2 with a ~3000-token prompt are bit-identical.
• MiniMax-M3-MXFP8 TP4, B200:
  • Fusion off: bit-identical to the parent. A difference in one request came from the parent itself: a second parent run matches this PR's two runs exactly.
  • Fusion on: GSM8K (200 questions, 5-shot) is 0.91 on the parent and 0.91 on this PR. Of 20 raw prompts with EOS honored, the parent runs all 20 to the 1024-token limit; this PR stops at EOS on 8, and no new runaway appears. So this shows the change is not worse, not that the runaway is gone: 12 of the 20 still run to the limit.
• GLM5-Next and Dots3: no checkpoint was available; CPU unit tests only. Both now take new collective paths (the deferral without the fused kernel, and under attention DP the reduce-scatter moved to the next layer), which are GPU-checked on the other models only.
• torch.compile: these models now also defer without the fused kernel, in prefill too, so the tc_piecewise prefill backend cannot trace them at TP > 1 (see [Refactor] Carry a deferred FFN all-reduce as UnreducedOutput and complete it in the next layer without the fused kernel #41196). A MiniMax-M3 multimodal launch selects that backend by default.
• CPU unit tests: CuteDSL fusion, the Qwen3.5 loaders and pipeline parallel, the FFN exit, the layer-stack exit, aux capture and deepstack all pass.

Rebased onto main 9dc4c5d891: every commit's patch is unchanged, and the related unit tests pass on the new base. The GPU results above were measured on the previous base, 5bb850d536.

Speed Tests and Profiling

Not measured separately. The collectives are those of the previous PRs; MiniMax-M3's sparse layers now take the fused kernel when fusion is on.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): 🚫 Run #36114852375
Latest PR Test (Extra): 🚫 Run #36114852328
Latest PR Test (AMD ROCm 10): ⏳ Run #36114852454