[Core] Whisper support torch.compile (#30385)

Signed-off-by: NickLucche <nlucches@redhat.com>

vllm/compilation/decorators.py

@@ -25,6 +25,7 @@ from vllm.config import (
     set_current_vllm_config,
 )
 from vllm.config.compilation import DynamicShapesType
+from vllm.forward_context import get_forward_context, is_forward_context_available
 from vllm.logger import init_logger
 from vllm.sequence import IntermediateTensors
 from vllm.utils.import_utils import resolve_obj_by_qualname
@@ -388,6 +389,12 @@ def _support_torch_compile(
         if self.do_not_compile or torch.compiler.is_compiling():
             return self.forward(*args, **kwargs)
 
+        # If skip_compiled is set, bypass compiled model call. This is used e.g. for
+        # enc-dec models where tensor shapes/types vary across invocations, preventing
+        # the capture of a single computational graph.
+        if is_forward_context_available() and get_forward_context().skip_compiled:
+            return self.forward(*args, **kwargs)
+
         # if aot_compiled_fn is set, call it with partition wrapper context.
         # The partition wrapper must be active at runtime for CUDA graph
         # capture to work correctly with inductor graph partitioning.