Using FlashInfer CUTLASS Backend for vLLM is Slow on SM120/121

[bkryu]

vLLM has reported that SM121 NVFP4 inference is generally slow on SM12x, especially on DGX Spark (SM121)

Current issue tracks ongoing efforts to improve performance on SM120 & 121 for NVFP4 inference.

Benchmarking:

• bench: Enable microbenchmarking on SM121 #3002

Kernel Improvements:

• feat: add PDL support to rmsnorm_fp4quant and add_rmsnorm_fp4quant CuTe DSL kernels #3008
• perf: Optimize CUTLASS MoE helper kernels for small-batch decode workloads #3014
• perf: Port TRT-LLM SM120/SM121 FP4 CUTLASS GEMM optimizations. Add PDL #3026
• feat: Add backend="b12x" for mm_fp4 on SM120 #3051
• feat: Add b12x CuTe DSL fused MoE for SM120 #3066
• feat: Add b12x_fused_moe / B12xMoEWrapper SM120 APIs with micro kernel and ReLU2 #3080
• perf(moe): optimize SM120 b12x MoE short decode #3193

Misc. fixes

• Integrate CUTLASS Small Tile N Blockscaled GEMMs/Grouped GEMMs for SM120 and SM121 #3152
• fix(sm12x): fix micro-kernel workspace sizing when routed_rows > num_local_experts #3191