diff --git a/dsv4/ops/gemm_runner.py b/dsv4/ops/gemm_runner.py
index a6dbe50d..e9ec40e9 100644
--- a/dsv4/ops/gemm_runner.py
+++ b/dsv4/ops/gemm_runner.py
@@ -175,8 +175,17 @@ def run_nvfp4_grouped_gemm(
 
     out = torch.zeros(tokens_sum, n_dim, dtype=torch.bfloat16, device=mat_a.device)
 
-    cache_key = (num_experts, str(mat_a.device), mma_tiler_mn, cluster_shape_mn,
-                 mat_a.shape[1], mat_b.shape[2])
+    # NVFP4-3: use 2-CTA UMMA for M>=256 (1.7-1.9× throughput at prefill)
+    use_2cta = tokens_sum >= 256 and cluster_shape_mn[0] % 2 == 0
+    effective_mma_tiler_mn = (
+        (mma_tiler_mn[0] * 2, mma_tiler_mn[1]) if use_2cta else mma_tiler_mn
+    )
+    effective_cluster_shape_mn = (
+        (cluster_shape_mn[0] * 2, cluster_shape_mn[1]) if use_2cta else cluster_shape_mn
+    )
+
+    cache_key = (num_experts, str(mat_a.device), effective_mma_tiler_mn, effective_cluster_shape_mn,
+                 mat_a.shape[1], mat_b.shape[2], use_2cta)
     
     if cache_key not in _compiled_kernel_cache:
         # Lazy compilation — safety net if warmup_compilation wasn't called.
@@ -187,8 +196,9 @@ def run_nvfp4_grouped_gemm(
             accumulate_on_output=False,
             separate_tensormap_init=True,
             consistent_token_padding=False,
-            mma_tiler_mnk=(*mma_tiler_mn, 256),
-            cluster_shape_mnk=(*cluster_shape_mn, 1),
+            mma_tiler_mnk=(*effective_mma_tiler_mn, 256),
+            cluster_shape_mnk=(*effective_cluster_shape_mn, 1),
+            use_2cta_instrs=use_2cta,
         )
 
         def to_cute(t):
@@ -385,8 +395,18 @@ def run_fused_swiglu_grouped_gemm(
 
     out = torch.zeros(tokens_sum, n_dim, dtype=torch.bfloat16, device=mat_a.device)
 
-    cache_key = ('fused', num_experts, str(mat_a.device), mma_tiler_mn, cluster_shape_mn,
-                 mat_a.shape[1], mat_b.shape[2], swiglu_limit)
+    # NVFP4-3: use 2-CTA UMMA for M>=256 (1.7-1.9× throughput at prefill)
+    # At decode (M<256), 1-CTA is correct (2-CTA wastes hardware)
+    use_2cta = tokens_sum >= 256 and cluster_shape_mn[0] % 2 == 0
+    effective_mma_tiler_mn = (
+        (mma_tiler_mn[0] * 2, mma_tiler_mn[1]) if use_2cta else mma_tiler_mn
+    )
+    effective_cluster_shape_mn = (
+        (cluster_shape_mn[0] * 2, cluster_shape_mn[1]) if use_2cta else cluster_shape_mn
+    )
+
+    cache_key = ('fused', num_experts, str(mat_a.device), effective_mma_tiler_mn, effective_cluster_shape_mn,
+                 mat_a.shape[1], mat_b.shape[2], swiglu_limit, use_2cta)
     
     if cache_key not in _fused_kernel_cache:
         # Lazy compilation
@@ -396,8 +416,9 @@ def run_fused_swiglu_grouped_gemm(
             accumulate_on_output=False,
             separate_tensormap_init=True,
             consistent_token_padding=False,
-            mma_tiler_mnk=(*mma_tiler_mn, 256),
-            cluster_shape_mnk=(*cluster_shape_mn, 1),
+            mma_tiler_mnk=(*effective_mma_tiler_mn, 256),
+            cluster_shape_mnk=(*effective_cluster_shape_mn, 1),
+            use_2cta_instrs=use_2cta,
             fused_swiglu=True,
             swiglu_limit=swiglu_limit,
         )