Add swiglu_limit=10.0 activation clamping (was missing)

DeepSeek-V4 uses SiluAndMulWithClamp(10.0) which clamps:
- silu(gate) to max 10.0
- up to [-10.0, 10.0]

Our runner was doing plain F.silu(gate) * up without clamping.
Large gate values could produce unbounded SiLU output, causing
numerical issues in the L2 GEMM. This is likely contributing to
garbage model output.

vllm/nvfp4_cutedsl.py

@@ -46,6 +46,7 @@ class CuTeDSLMoERunner:
         self.top_k = top_k
         self.device = device
         self.experts_start_idx = experts_start_idx
+        self._swiglu_limit = None  # Set via set_swiglu_limit()
         
         # Weight storage (set before _ensure_stacked)
         self.l1_fp4 = None
@@ -92,6 +93,10 @@ class CuTeDSLMoERunner:
               f"pid={os.getpid()}", flush=True)
         self._buffers_allocated = False
 
+    def set_swiglu_limit(self, limit: float | None):
+        """Set the swiglu_limit for activation clamping."""
+        self._swiglu_limit = limit
+
     def _fill_token_indices(self):
         """Fill _token_indices with [0,0,..0, 1,1,..1, ...] (each token repeated top_k times).
         
@@ -438,10 +443,15 @@ class CuTeDSLMoERunner:
         # Extract real token outputs from padded GEMM output
         l1_out_real = l1_out[padded_dst]
         
-        # === SiLU(gate) * up ===
+        # === SiLU(gate) * up (with swiglu_limit clamp) ===
         gate = l1_out_real[:, :self.intermediate_size]
         up = l1_out_real[:, self.intermediate_size:]
-        activated = torch.nn.functional.silu(gate) * up
+        gate_silu = torch.nn.functional.silu(gate)
+        # Apply DeepSeek-V4 swiglu_limit: clamp both silu(gate) and up
+        if self._swiglu_limit is not None:
+            gate_silu = gate_silu.clamp(max=self._swiglu_limit)
+            up = up.clamp(min=-self._swiglu_limit, max=self._swiglu_limit)
+        activated = gate_silu * up
         
         # === L2: down ===
         padded_activated = self._shared_bufs['activated']

vllm/patches/deepseek_v4.py

@@ -551,6 +551,10 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         # the actual amax and compute correct gs values.
         self._warmup_activation_global_scales()
 
+        # Set swiglu_limit for activation clamping in the runner
+        if self.swiglu_limit is not None:
+            self._cutedsl_runner.set_swiglu_limit(float(self.swiglu_limit))
+
     def _warmup_activation_global_scales(self) -> None:
         """Run a warmup forward pass to compute correct activation global scales.