Fix torch.compile crash: @torch.compiler.disable on all CuTeDSL run()

CuTeDSL internals (Path.cwd, threading, JIT) are incompatible with
torch.dynamo tracing. Marking run() as compiler-disabled makes the
runners opaque to torch.compile — they execute eagerly while the
rest of the model gets compiled.

cutedsl/nvfp4_linear.py

@@ -110,6 +110,7 @@ class CuTeDSLNvfp4Linear:
             self._activation_global_scale = gs
 
 
+    @torch.compiler.disable
     def run(self, hidden_states: torch.Tensor) -> torch.Tensor:
         """Forward: BF16 input → NVFP4 GEMM → BF16 output."""
         self._ensure_initialized()

cutedsl/shared_expert_pipeline.py

@@ -274,6 +274,7 @@ class CuTeDSLSharedExpertRunner:
 
         return out[:num_tokens]
 
+    @torch.compiler.disable
     def run(self, hidden_states: torch.Tensor) -> torch.Tensor:
         """Full shared expert forward: L1 → SiLU → L2 → output."""
         self._ensure_initialized()

vllm/nvfp4_cutedsl.py

@@ -374,6 +374,7 @@ class CuTeDSLMoERunner:
         
 
 
+    @torch.compiler.disable
     def run(self, hidden_states, topk_weights, topk_ids, expert_indices=None):
         """Run the NVFP4 MoE forward pass.