fix: free per-expert weight lists after stacking in CuTeDSL runner

_ensure_stacked() creates stacked copies of all weights but never freed
the per-expert lists. For 256 experts on a 175GB model, this doubles
weight memory to ~350GB, causing OOM.

Now the per-expert lists (l1_fp4, l1_sf, l1_gs, l2_fp4, l2_sf, l2_gs)
are set to None after stacking, keeping only the single stacked copy.

vllm/nvfp4_cutedsl.py

@@ -45,7 +45,11 @@ class CuTeDSLMoERunner:
         self._l2_gsb = None
 
     def _ensure_stacked(self):
-        """Lazily stack weight tensors into the format the kernel expects."""
+        """Lazily stack weight tensors into the format the kernel expects.
+        
+        After stacking, the per-expert lists are freed to avoid holding
+        two copies of ~175GB of weight data in GPU memory.
+        """
         if self._l1_mat_b is not None:
             return
         self._l1_mat_b = make_b_k_major(torch.stack(self.l1_fp4))
@@ -54,6 +58,13 @@ class CuTeDSLMoERunner:
         self._l2_scale_b = assemble_scales_3d_side(self.l2_sf)
         self._l1_gsb = torch.tensor(self.l1_gs, dtype=torch.float32, device=self.device)
         self._l2_gsb = torch.tensor(self.l2_gs, dtype=torch.float32, device=self.device)
+        # Free per-expert lists — stacked tensors are the only copy now
+        self.l1_fp4 = None
+        self.l1_sf = None
+        self.l1_gs = None
+        self.l2_fp4 = None
+        self.l2_sf = None
+        self.l2_gs = None
 
     def prepare_weights_direct(self, l1_fp4, l1_sf, l1_gs, l2_fp4, l2_sf, l2_gs):
         """Set weights directly from checkpoint (no dequant→requant)."""