[alphabetc1]

Motivation

DeepseekV4ForCausalLMDSpark defines shared_experts_fusion_disable_reason
twice. The second definition wins, so the first one, the only one with the NPU
branch, is dead code:

    @classmethod
    def shared_experts_fusion_disable_reason(cls, hf_config, quant_config):
        if _is_npu:
            return (
                "NPU DSpark ModelSlim weight loading does not support mapping "
                "shared experts into fused expert slots."
            )
        return DeepseekV4ForCausalLM.shared_experts_fusion_disable_reason(...)

    @classmethod                                   # <- this one shadows it
    def shared_experts_fusion_disable_reason(cls, hf_config, quant_config):
        return DeepseekV4ForCausalLM.shared_experts_fusion_disable_reason(...)

The effect is that on NPU, shared-experts fusion is not disabled for the DSpark
draft even though the guard was written to disable it. Per the message that
guard carries, NPU ModelSlim weight loading cannot map shared experts into
fused expert slots, so the draft ends up fusing experts on a path that does not
support it.

CUDA is unaffected: both definitions fall through to the same base
implementation there.

The NPU-aware version arrived in #33676 alongside the plain one from #33312; the
overwrite looks like a merge artifact rather than an intentional revert.

Modifications

Delete the second, shadowing definition. No other change.

Accuracy Tests

Not applicable on CUDA, where both definitions were equivalent and behavior is
byte-identical.

I do not have NPU hardware, so the restored branch is not exercised here. It
restores the behavior #33676 intended; an NPU reviewer confirming that intent
would be the right check before merge.

Speed Tests and Profiling

Not applicable. Dead-code removal with no CUDA behavior change.

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.

Found while auditing DSV4 startup; unrelated to that work, so it is on its own.

---

CI States

Latest PR Test (Base): ❌ Run #34808005605
Latest PR Test (Extra): ❌ Run #34808005393
Latest PR Test (AMD ROCm 10): ❌ Run #34808005509

[mmangkad]

/rerun-test test/registered/unit/models/test_deepseek_v4_shared_expert_fusion.py

[github-actions]

⚠️ CI Maintenance Mode is Active

The CI infrastructure is currently under maintenance. All PR CI runs are paused until maintenance is complete. Merging non-CI-fix PRs is prohibited during maintenance mode.

Follow issue #21065 for status updates. Re-run was not dispatched.