Fix decode OOM caused by retraction

python/sglang/srt/managers/schedule_batch.py

@@ -1609,6 +1609,7 @@ def retract_all(self, server_args: ServerArgs):
     def retract_decode(
         self,
         server_args: ServerArgs,
+        buf_multiplier: int = 1,
     ) -> Tuple[List[Req], float, List[Req]]:
         """Retract the decoding requests when there is not enough memory."""
         sorted_indices = list(range(len(self.reqs)))
@@ -1630,7 +1631,9 @@ def retract_decode(
         retracted_reqs = []
         first_iter = True
         while first_iter or (
-            not self.check_decode_mem(selected_indices=sorted_indices)
+            not self.check_decode_mem(
+                selected_indices=sorted_indices, buf_multiplier=buf_multiplier
+            )
         ):
             if len(sorted_indices) == 1:
                 # Corner case: only one request left

python/sglang/srt/managers/scheduler.py

@@ -1965,7 +1965,7 @@ def update_running_batch(self, batch: ScheduleBatch) -> Optional[ScheduleBatch]:
         ):
             old_ratio = self.new_token_ratio
             retracted_reqs, new_token_ratio, reqs_to_abort = batch.retract_decode(
-                self.server_args
+                self.server_args, self.decode_mem_cache_buf_multiplier
             )
             self.num_retracted_reqs = len(retracted_reqs)
             self.new_token_ratio = new_token_ratio

python/sglang/srt/model_executor/model_runner.py

@@ -1666,7 +1666,9 @@ def init_memory_pool(
         self.max_total_num_tokens = self.profile_max_num_token(total_gpu_memory)
 
         if (small_kv_size := envs.SGLANG_CI_SMALL_KV_SIZE.get()) > 0:
-            # Use a small KV cache pool size for local tests
+            logger.info(
+                f"Use a small KV cache pool size ({small_kv_size}) for local tests"
+            )
             self.max_total_num_tokens = small_kv_size
 
         if max_num_reqs is None:

test/registered/spec/eagle/test_eagle_infer_b.py

@@ -292,7 +292,8 @@ def setUpClass(cls):
         # These config helps find a leak.
         # FIXME(lsyin): use override context manager
         envs.SGLANG_CI_SMALL_KV_SIZE.set(4500)
-        super().setUpClass()
+        with envs.SGLANG_TEST_RETRACT.override(True):
+            super().setUpClass()
 
 
 class TestEAGLEServerTriton(TestEAGLEServerBasic):