[Tracking Issue]: Qwen3-next performance optimisations

[vadiklyutiy]

Proposal to improve performance

This is a tracking issue for performance optimization for Qwen3-next to keep all necessary things in one place.

• torch.compile for GDN attn [Perf][Qwen3-next]: torch.compile GDN attn #27152
• not optimal linear for small batch sizes [Performance]: non-optimal performance of linear for small batches #27173
• GDN attn decrease CPU overhead [Performance][Qwen3-next] Decrease huge CPU overhead #27222
• Full CudaGraph for TRT-LLM Gen attn (for MTP only) [Perf] Enable full CUDA graphs for spec decoding with FlashInfer #26937
• Enable TRT-LLM Gen MoE.
  FP16:
  FP8: [Performance] Support FP8 flashinfer TRTLLM MOE on Qwen3 and Qwen-3next #27492
  FP4:
  DONE:Update the routing for TRTLLMGEN to support kimi k2 and qwen flashinfer-ai/flashinfer#1831
• Moving of the gate / router op to be after the shared_experts execution. [perf] Enable concurrent execution of "shared_experts" and "selected_experts" #27578
  [Performance] Dual stream execution of "shared_experts" and "selected_experts" inside FusedMoE #26440 for reference.
• async-sched + spec-decoding (not Qwen3-next specific feature, but required) [Core] Async Scheduling X Spec Decoding Compatibility #24799
• GDN prefix cache [V1][Hybrid] GatedDeltaNet Automatic Prefix Caching #26807

[ZJY0516]

GDN prefix cache #26807

[ZJY0516]

FP8 flashinfer TRTLLM MOE on Qwen3 and Qwen-3next: #27492