fix: skip w13 swap in FlashInfer CUTLASS path for non-gated MoE

[djmmoss]

Summary

convert_to_unquantized_kernel_format() in vllm/model_executor/layers/fused_moe/oracle/unquantized.py unconditionally calls swap_w13_to_w31(layer.w13_weight) when the FLASHINFER_CUTLASS backend is selected. That helper assumes the second-to-last dim of w13 is [w1; w3] (the gate/up projections in a SwiGLU-style gated MoE) and flips the two halves so the kernel sees [w3; w1].

For non-gated MoE (is_act_and_mul=False), w13_weight has shape [num_experts, intermediate_size, hidden_size] — there is only one logical weight, no halves to swap. swap_w13_to_w31 reshapes with // 2 and flips, silently splitting the single weight matrix down the middle and swapping the halves, which scrambles the weights.

The model still runs, but produces nonsense output: at temperature=0, generation degenerates into prompt repetition, finish_reason is always length, and content is empty.

This patch gates the swap on layer.moe_config.is_act_and_mul. For non-gated layouts, the weight is left untouched.

Test plan

• Reproduced the issue with a non-gated MoE model under VLLM_USE_FLASHINFER_MOE_FP16=1 VLLM_FLASHINFER_MOE_BACKEND=throughput (CUTLASS path) — 0/10 prompts produced coherent output before the fix.
• After the fix, CUTLASS produces correct, coherent reasoning that matches the Triton baseline on math/algebra/sequence/deduction prompts at temperature=0. Numerical drift causes minor wording differences but final answers match.
• Throughput benefit preserved: +19.7% output throughput / -17.9% mean TPOT vs Triton on a 500-prompt input=8000/output=1000 benchmark (TP=8, multi-node).

🤖 Generated with Claude Code