CRITICAL FIX: SE swizzled buffers were allocated then overwritten with None — graph capture would fall through to broken Python path

dsv4/layers/shared_expert.py

@@ -188,8 +188,8 @@ class Nvfp4SharedExpert:
         self._l2_gsa_buf = torch.zeros(1, dtype=torch.float32, device=self.device)
         
         # Pre-allocated swizzled scale output buffers (for CUDA graph capture)
-        self._padded_x_sf_swizzled_buf_l1 = None
-        self._padded_x_sf_swizzled_buf_l2 = None
+        # NOTE: _padded_x_sf_swizzled_buf_l1/l2 are allocated above (line 183-184)
+        # Do NOT set to None — they are required for CUDA graph capture swizzle path
         
         # Pre-allocated L1 output buffer for graph capture
         # L1 produces gate+up combined: 2 * intermediate_size BF16 columns