[ByronHsu]

Why?

• Kimi-K3's routed experts use the SiTU activation. Two checks only allowed silu/relu2: the CuTe DSL MoE runner and the ModelOpt NVFP4 activation guard. That blocked K3 from FlashInfer's NVFP4 W4A16 CuTe DSL kernel (NVFP4 weights, BF16 activations).
• FlashInfer already supports SiTU in this kernel (feat(moe): support GeGLU and SiTU in W4A16 flashinfer-ai/flashinfer#4394, situ_beta / situ_linear_beta).

What?

• modelopt_quant.py: allow situ on the CuTe DSL v2 NVFP4 path, as already done for TRT-LLM.
• flashinfer_cutedsl.py:
  • accept situ (FlashInfer selects SiTU through the gated Swiglu path);
  • pass situ_beta / situ_linear_beta (K3's gemm1_alpha / gemm1_clamp_limit) to CuteDslMoEWrapper. Without them the kernel would compute plain SwiGLU.
• New GPU test test/registered/unit/layers/quantization/test_nvfp4_cutedsl_w4a16_situ_moe.py.
• Usage is the same as other W4A16 models: --moe-runner-backend flashinfer_cutedsl with SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16=1.
• Checkpoint requirement (documented, not enforced): gate (w1) and up (w3) must share one weight_scale_2 per expert, because the kernel applies one GEMM1 scale before SiTU's tanh.

Verification?

Before:  K3 NVFP4 + SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16=1 + --moe-runner-backend flashinfer_cutedsl
         -> AssertionError: activation='situ' is unsupported by flashinfer_cutedsl
After:   same command
         -> routed experts run CuTe DSL W4A16 with SiTU fused into GEMM1

Observed on GB300 with the same kernel path on a development branch (5-layer truncated K3 converted to ModelOpt NVFP4, TP4, CUDA graphs), CuTe DSL W4A16 vs Marlin on the same NVFP4 weights:

mean |prompt logprob diff| = 0.026, max = 0.20, first token equal 7/8   (observed)

Reproduce:

SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16=1 \
python -m sglang.launch_server --model-path <Kimi-K3 ModelOpt NVFP4 checkpoint> \
  --quantization modelopt_fp4 --moe-runner-backend flashinfer_cutedsl \
  --tp 4 --trust-remote-code

Test?

• test_nvfp4_cutedsl_w4a16_situ_moe.py on GB300 with this branch and the upstream pins (flashinfer 0.7.1, torch 2.14.1+cu130, GB300): passed (observed). The test checks two things:
  • CuTe DSL W4A16 SiTU output matches a dequantized torch reference;
  • a SiLU reference does not match (cosine < 0.99), so the test proves SiTU is applied.
• test_nvfp4_moe_backends.py::TestNvFp4MoeBackends::test_flashinfer_cutedsl (existing SiLU path), same run: passed (observed; 2 passed in 415.95s together with the test above).
• pre-commit run on the changed files: all hooks passed (observed).
• CI: pending.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ⏳ Run #41449060368
Latest PR Test (Extra): ❌ Run #41449060314
Latest PR Test (AMD ROCm 10): ⏳ Run #41449060148

Tip

Base CI only covers the core cases, and its workload may not exercise your change. Use /rerun-test to run the tests your change touches, or copy this prompt for an agent:

Find the registered tests this PR affects and run them with /rerun-test, following https://github.com/sgl-project/sglang/blob/main/.agents/skills/rerun-test/SKILL.md

[ByronHsu]

/tag-and-rerun-ci

[ByronHsu]

/rerun-failed-ci

[ByronHsu]

/rerun-failed-ci

[ByronHsu]

/rerun-failed-ci

[ByronHsu]

/rerun-failed-ci

[ByronHsu]

/rerun-failed-ci