fix: remove CuTeDSL warmup — OOM with 175GB model loaded

The warmup allocated 1GB of dummy tensors but the model already
uses 175.7GB of the 178.35GB per GPU. No room.

With FULL_AND_PIEWISE CUDA graph mode, the kernel compiles during
the graph capture phase (which manages memory properly). The warmup
was a band-aid for eager mode and is now redundant.

vllm/patches/deepseek_v4.py

@@ -517,20 +517,9 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         self.w2_weight_scale_2 = None
         self.w2_input_scale = None
 
-        # Warm up the CuTeDSL kernel (JIT compiles MLIR→PTX on first call).
-        # CuTeDSL caches by (M, N, K) shape — different shapes trigger new
-        # compiles. Running warmup on every layer ensures all kernels are
-        # compiled before inference, preventing vLLM RPC timeouts.
-        # The MMA tiler needs >= 128 tokens; using 128 for the warmup.
-        device = self._cutedsl_runner.l1_fp4[0].device
-        dummy_hidden = torch.randn(128, self.hidden_size, dtype=torch.bfloat16, device=device)
-        dummy_ids = torch.zeros(128, 1, dtype=torch.int32, device=device)
-        dummy_weights = torch.ones(128, 1, dtype=torch.float32, device=device)
-        try:
-            self._cutedsl_runner.run(dummy_hidden, dummy_weights, dummy_ids, expert_indices=[0])
-        except Exception as exc:
-            print(f"  CuTeDSL warmup FAILED: {exc}", flush=True)
-            print("  The CUDA context may be corrupted. Check kernel alignment/tiling.", flush=True)
+        # Note: No explicit CuTeDSL warmup here. With FULL_AND_PIECEWISE
+        # CUDA graph mode, the kernel compiles during graph capture (startup).
+        # In eager mode, the first inference triggers JIT compilation.
 
     def forward(
         self,