Fix: define padded_max_slots before using it in shared buffer allocation

vllm/nvfp4_cutedsl.py

@@ -145,6 +145,8 @@ class CuTeDSLMoERunner:
         # Padded hidden/activated: SHARED across all runners (not per-layer)
         # These are only used during run() which is sequential across layers.
         # Per-layer allocation would be 72 MB × 60 layers = 4.3 GB → OOM.
+        max_rows_per_expert = self._max_chunks_per_expert * 128
+        padded_max_slots = self.num_experts * max_rows_per_expert
         device_key = str(self.device)
         if not hasattr(CuTeDSLMoERunner, '_shared_padded_bufs'):
             CuTeDSLMoERunner._shared_padded_bufs = {}