NVFP4-3: enable 2-CTA UMMA when MMA tile M >= 256 (1.7-1.9x throughput)

dsv4/ops/gemm_runner.py

@@ -298,6 +298,11 @@ def warmup_fused_swiglu_compilation(num_experts, K_packed, N_packed, device,
     global_scale_a = torch.ones(1, dtype=torch.float32, device=device)
     global_scale_b = torch.ones(1, dtype=torch.float32, device=device)
     
+    # NVFP4-3: Enable 2-CTA UMMA when MMA tile M >= 256 and cluster M is even.
+    # 2-CTA gives 1.7-1.9× throughput at prefill shapes by doubling effective MMA tile M.
+    # Not beneficial at decode (M < 256) — would waste hardware.
+    use_2cta = mma_tiler_mn[0] >= 256 and cluster_shape_mn[0] % 2 == 0
+    
     kernel = FusedSwiGLUScaledGroupedGemmKernel(
         scenario="2Dx3D",
         sf_vec_size=SF_VEC_SIZE,
@@ -308,6 +313,7 @@ def warmup_fused_swiglu_compilation(num_experts, K_packed, N_packed, device,
         cluster_shape_mnk=(*cluster_shape_mn, 1),
         fused_swiglu=True,
         swiglu_limit=swiglu_limit,
+        use_2cta_instrs=use_2cta,
     )
     
     def to_cute(t):