[ROCm][Bugfix] Bring back fallback to eager mode removed in vllm-project#14917, but for ROCm only (vllm-project#15413)

Signed-off-by: Gregory Shtrasberg <[email protected]>

Author: gshtras

vllm/config.py

@@ -29,7 +29,7 @@
 from vllm.model_executor.layers.quantization import (QUANTIZATION_METHODS,
                                                      get_quantization_config)
 from vllm.model_executor.models import ModelRegistry
-from vllm.platforms import CpuArchEnum
+from vllm.platforms import CpuArchEnum, current_platform
 from vllm.sampling_params import GuidedDecodingParams
 from vllm.tracing import is_otel_available, otel_import_error_traceback
 from vllm.transformers_utils.config import (
@@ -684,6 +684,13 @@ def _verify_cuda_graph(self) -> None:
             self.max_seq_len_to_capture = self.max_model_len
         self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                           self.max_model_len)
+        ROCM_UNSUPPORTED_MODELS = ['mllama']
+        if (self.hf_config.model_type in ROCM_UNSUPPORTED_MODELS
+                and not self.enforce_eager and current_platform.is_rocm()):
+            logger.warning(
+                "CUDA graph is not supported for %s on ROCm yet, fallback "
+                "to the eager mode.", self.hf_config.model_type)
+            self.enforce_eager = True
 
     def _verify_bnb_config(self) -> None:
         """