[Core] Force PIECEWISE CUDAGraph mode for encoder-decoder

Whisper does not work with full cudagraphs. That is being worked on in
PR vllm-project#25208.

The failure can be reproduced reliably via
`tests/models/multimodal/generation/test_whisper.py`, at least in my
H100 development environment. The tests passed on the PR and I'm not
sure why.

Regardless, this seems like the right change to make until vllm-project#25208
sorts out exactly what changes are needed.

Signed-off-by: Russell Bryant <[email protected]>

Author: russellb

vllm/config/__init__.py

@@ -364,9 +364,11 @@ def __post_init__(self):
                     self.compilation_config.cudagraph_mode = \
                         CUDAGraphMode.FULL_AND_PIECEWISE
 
-                    # pooling model does not support full cudagraphs
+                    # pooling models and encoder-decoder models
+                    # do not support full cudagraphs
                     if self.model_config is not None and \
-                        self.model_config.pooler_config is not None:
+                        (self.model_config.pooler_config is not None
+                         or self.model_config.is_encoder_decoder):
                         self.compilation_config.cudagraph_mode = \
                             CUDAGraphMode.PIECEWISE
                 else: