Fix cudagraph: static loop for per-expert scale swizzle

The while loop had variable trip count (GPU scalar in condition),
requiring CPU-GPU sync. Replaced with fixed max_chunks_per_expert
iterations. Unused chunks are zero buffers (harmless for GEMM).

vllm/nvfp4_cutedsl.py

@@ -83,6 +83,9 @@ class CuTeDSLMoERunner:
         self._padded_hidden_buf = None
         self._padded_activated_buf = None
         self._padded_expert_offsets_buf = None
+        self._max_chunks_per_expert = cutedsl_ceil_div(
+            self.max_num_tokens * self.top_k, self.num_experts * 128
+        )
         self._buffers_allocated = False
 
     def _fill_token_indices(self):
@@ -249,24 +252,17 @@ class CuTeDSLMoERunner:
         padded_x_sf[dst_rows, :K_sf] = x_sf
         
         # Phase 2: Per-expert swizzle and concatenate
+        # Fixed loop: max_chunks_per_expert iterations per expert (cudagraph-safe).
+        # Unused chunks are zero buffers that contribute nothing to the GEMM.
+        max_chunks = self._max_chunks_per_expert
         swizzled_parts = []
         for e in range(num_experts):
-            n_padded = padded_rows_per_expert[e]
-            start = padded_expert_offsets[e]
             buf = per_expert_bufs[e]
-            # Process in 128-row chunks
-            offset = start
-            remaining = n_padded
-            while remaining > 0:
-                buf.zero_()
-                chunk = min(remaining, 128)
-                buf[:chunk, :K_sf] = padded_x_sf[offset:offset + chunk]
-                swizzled = pad_and_swizzle_single(buf)
-                swizzled_parts.append(swizzled)
-                offset += 128
-                remaining -= 128
-            if n_padded == 0:
+            for c in range(max_chunks):
                 buf.zero_()
+                src_offset = padded_expert_offsets[e] + c * 128
+                # Copy 128 rows — rows beyond n_padded are already zero
+                buf[:, :K_sf] = padded_x_sf[src_offset:src_offset + 128]
                 swizzled = pad_and_swizzle_single(buf)
                 swizzled_parts.append(swizzled)