Support piecewise CUDA graph for embedding models (#13852)

Co-authored-by: qiuxuan.lzw <qiuxuan.lzw@alibaba-inc.com>

python/sglang/srt/model_executor/model_runner.py

@@ -88,6 +88,7 @@ from sglang.srt.layers.dp_attention import (
     initialize_dp_attention,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.pooler import EmbeddingPoolerOutput
 from sglang.srt.layers.sampler import Sampler
 from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
 from sglang.srt.lora.lora_manager import LoRAManager
@@ -2201,7 +2202,7 @@ class ModelRunner:
         forward_batch: ForwardBatch,
         skip_attn_backend_init: bool = False,
         pp_proxy_tensors=None,
-    ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
+    ) -> Union[LogitsProcessorOutput, PPProxyTensors, EmbeddingPoolerOutput]:
 
         if self.is_multimodal and should_use_external_mm_preprocess(self.model):
             data_embedding_funcs = resolve_external_mm_data_embedding_funcs(self.model)