[Fix] Slice input_embeds to extend_input_len in prepare_for_extend (#20376)

python/sglang/srt/managers/schedule_batch.py

@@ -1528,8 +1528,11 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
             # If input_embeds are available, store them
             if req.input_embeds is not None:
-                # If req.input_embeds is already a list, append its content directly
-                input_embeds.extend(req.input_embeds)  # Use extend to avoid nesting
+                # Slice to match extend_input_len — PrefillAdder truncates
+                # fill_ids/extend_input_len on chunk overflow but not input_embeds.
+                input_embeds.extend(
+                    req.input_embeds[pre_len : pre_len + req.extend_input_len]
+                )
 
             multimodal_inputs.append(req.multimodal_inputs)
 
@@ -1973,6 +1976,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def prepare_for_decode(self):
         self.forward_mode = ForwardMode.DECODE
         bs = len(self.reqs)
+        # Decode embeds the last output token via embed_tokens; clear the stale
+        # prefill-time tensor so it doesn't leak into ForwardBatch.
+        self.input_embeds = None
 
         if self.is_spec_v2:
             # TODO(spec-v2): all spec v2 should go through this path

python/sglang/srt/managers/scheduler.py

@@ -2336,6 +2336,8 @@ class Scheduler(
             self.is_mixed_chunk
             and not self.running_batch.is_empty()
             and not (new_batch.return_logprob or self.running_batch.return_logprob)
+            # mix_with_running cats input_ids but not input_embeds — shapes would mismatch
+            and new_batch.input_embeds is None
         ):
             # TODO (lianmin): support return_logprob + mixed chunked prefill
             self.running_batch.filter_batch(v1_spec_info_filtered=True)