Fix garbled imports in cutedsl/runner.py

cutedsl/runner.py

@@ -18,13 +18,6 @@ import torch
 from cutedsl.bridge import (
     quantize_activation_nvfp4,
     quantize_weight_to_nvfp4,
-
-
-class _MoEApply(torch.autograd.Function):
-    """Custom autograd function to make CuTeDSL MoE runner opaque to torch.compile."""
-    @staticmethod
-    def forward(ctx, runner, hidden_states, topk_weights, topk_ids, expert_indices):
-        return runner._run_impl(hidden_states, topk_weights, topk_ids, expert_indices)
     quantize_to_nvfp4,
     make_b_k_major,
     assemble_scales_3d_side,
@@ -36,6 +29,13 @@ from cutedsl.kernel.moe.torch_scaled_grouped_mm import (
 )
 
 
+class _MoEApply(torch.autograd.Function):
+    """Custom autograd function to make CuTeDSL MoE runner opaque to torch.compile."""
+    @staticmethod
+    def forward(ctx, runner, hidden_states, topk_weights, topk_ids, expert_indices):
+        return runner._run_impl(hidden_states, topk_weights, topk_ids, expert_indices)
+
+
 class CuTeDSLMoERunner:
     """Manages NVFP4 MoE execution via the CuTeDSL kernel.