[AMD] Copy decode result on forward_stream instead of copy_stream (#29642)

python/sglang/srt/managers/scheduler.py

@@ -253,6 +253,7 @@ from sglang.srt.utils import (
     get_bool_env_var,
     get_int_env_var,
     is_cuda,
+    is_hip,
     is_mps,
     kill_itself_when_parent_died,
     require_mlp_sync,
@@ -291,6 +292,7 @@ TEST_RETRACT_INTERVAL = envs.SGLANG_TEST_RETRACT_INTERVAL.get()
 TEST_RETRACT_NO_PREFILL_BS = envs.SGLANG_TEST_RETRACT_NO_PREFILL_BS.get()
 
 _is_npu = is_npu()
+_is_hip = is_hip()
 
 
 class Scheduler(
@@ -3239,6 +3241,14 @@ class Scheduler(
                         batch_result.copy_done = self.device_module.Event()
                         if batch_result.delay_sample_func is None:
                             self._relay_forward_payload(future_indices, batch_result)
+                            if _is_hip:
+                                # Cross-stream sync costs more than the tiny D2H it
+                                # overlaps.
+                                batch_result.copy_to_cpu(
+                                    return_logprob=batch.return_logprob,
+                                    return_hidden_states=batch.return_hidden_states,
+                                )
+                            else:
                                 # Result D2H on copy_stream overlaps the next forward
                                 # instead of serializing on forward_stream; it's a leaf
                                 # gated by copy_done, so nothing on forward_stream waits.