[bugfix] support eagle with lora cudagraph specialization (#28318)

Signed-off-by: gnovack <gnovack@amazon.com>

vllm/v1/worker/gpu_model_runner.py

@@ -3602,7 +3602,18 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                     cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                     and not self.speculative_config.enforce_eager
                 )
-                self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
+
+                # Note(gnovack) - We need to disable cudagraphs for one of the two
+                # lora cases when cudagraph_specialize_lora is enabled. This is a
+                # short term mitigation for issue mentioned in
+                # https://github.com/vllm-project/vllm/issues/28334
+                if self.compilation_config.cudagraph_specialize_lora and activate_lora:
+                    use_cudagraphs = False
+
+                self.drafter.dummy_run(
+                    num_tokens,
+                    use_cudagraphs=use_cudagraphs,
+                )
 
         # This is necessary to avoid blocking DP.
         # For dummy runs, we typically skip EPLB since we don't have any real