fix concurrency > 1

Author: thecodingwizard

python/sglang/srt/managers/schedule_batch.py

@@ -1663,6 +1663,14 @@ def filter_batch(
         if self.multimodal_inputs is not None:
             self.multimodal_inputs = [self.multimodal_inputs[i] for i in keep_indices]
         self.req_pool_indices = self.req_pool_indices[keep_indices_device]
+
+        if self.spec_algorithm.is_eagle() and self.enable_overlap:
+            # In eagle overlap mode, seq_lens is mutated in the EagleWorkerClient's forward_stream,
+            # but we copy seq_lens in the scheduler's stream. This is a problem because seq_lens may
+            # not have been mutated by EagleWorkerClient before the scheduler stream starts making
+            # a copy of it. To avoid this, we synchronize all streams before copying seq_lens.
+            torch.cuda.synchronize()
+
         self.seq_lens = self.seq_lens[keep_indices_device]
         self.orig_seq_lens = self.orig_seq_lens[keep_indices_device]
         self.out_cache_loc = None