[sgl] copy mm_input in piecewise cuda graph when eagle3 is on (#23613)

python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py

@@ -798,6 +798,18 @@ class PiecewiseCudaGraphRunner:
                     **kwargs,
                 )
                 if isinstance(output, LogitsProcessorOutput):
+                    # Preserve mm_input_embeds when speculative decoding is
+                    # enabled. The speculative draft's prefill path
+                    # (eagle_worker_v2._draft_extend_for_prefill) reads
+                    # mm_input_embeds off this LogitsProcessorOutput to reuse
+                    # the target's encoder embeddings instead of re-embedding
+                    # multimodal placeholder token ids.
+                    mm_input_embeds = None
+                    if (
+                        self.model_runner.spec_algorithm.is_speculative()
+                        and output.mm_input_embeds is not None
+                    ):
+                        mm_input_embeds = output.mm_input_embeds[: self.raw_num_tokens]
                     return LogitsProcessorOutput(
                         next_token_logits=output.next_token_logits[
                             : self.raw_num_tokens
@@ -807,6 +819,7 @@ class PiecewiseCudaGraphRunner:
                             if output.hidden_states is not None
                             else None
                         ),
+                        mm_input_embeds=mm_input_embeds,
                     )
                 elif isinstance(output, EmbeddingPoolerOutput):
                     return output