[Fix] Fix major performance bug in certain cases (#1563)

Co-authored-by: hnyls2002 <hnyls2002@gmail.com>

.github/workflows/pr-test.yml

@@ -130,6 +130,12 @@ jobs:
           cd test/srt
           python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default
 
+      - name: Benchmark Offline Throughput (Non-streaming, small batch size)
+        timeout-minutes: 10
+        run: |
+          cd test/srt
+          python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_non_stream_small_batch_size
+
   performance-test-1-gpu-part-2:
     if: github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request'
     runs-on: 1-gpu-runner

python/sglang/bench_serving.py

@@ -845,6 +845,7 @@ def run_benchmark(args_: argparse.Namespace):
     tokenizer = get_tokenizer(tokenizer_id)
 
     if args.dataset_name == "sharegpt":
+        assert args.random_input_len is None and args.random_output_len is None
         input_requests = sample_sharegpt_requests(
             dataset_path=args.dataset_path,
             num_requests=args.num_prompts,
@@ -852,6 +853,7 @@ def run_benchmark(args_: argparse.Namespace):
             fixed_output_len=args.sharegpt_output_len,
         )
     elif args.dataset_name == "random":
+        assert args.random_input_len is not None and args.random_output_len is not None
         input_requests = sample_random_requests(
             input_len=args.random_input_len,
             output_len=args.random_output_len,
@@ -964,13 +966,11 @@ if __name__ == "__main__":
     parser.add_argument(
         "--random-input-len",
         type=int,
-        default=1024,
         help="Number of input tokens per request, used only for random dataset.",
     )
     parser.add_argument(
         "--random-output-len",
         type=int,
-        default=128,
         help="Number of output tokens per request, used only for random dataset.",
     )
     parser.add_argument(

python/sglang/srt/managers/scheduler.py

@@ -222,7 +222,7 @@ class Scheduler:
         )
         self.new_token_ratio = self.min_new_token_ratio
         self.new_token_ratio_decay = global_config.new_token_ratio_decay
-        self.do_not_get_new_batch = False
+        self.batch_is_full = False
 
     def event_loop(self):
         while True:
@@ -261,12 +261,10 @@ class Scheduler:
         for recv_req in recv_reqs:
             if isinstance(recv_req, TokenizedGenerateReqInput):
                 self.handle_generate_request(recv_req)
-                self.do_not_get_new_batch = False
             elif isinstance(
                 recv_req, (TokenizedEmbeddingReqInput, TokenizedRewardReqInput)
             ):
                 self.handle_embedding_request(recv_req)
-                self.do_not_get_new_batch = False
             elif isinstance(recv_req, FlushCacheReq):
                 self.flush_cache()
             elif isinstance(recv_req, AbortReq):
@@ -279,11 +277,12 @@ class Scheduler:
 
     @torch.inference_mode()
     def forward_step(self):
-        if self.do_not_get_new_batch and self.current_inflight_req is None:
+        if (
+            self.batch_is_full or len(self.waiting_queue) == 0
+        ) and self.current_inflight_req is None:
             new_batch = None
         else:
             new_batch = self.get_new_prefill_batch()
-        self.do_not_get_new_batch = False
 
         if new_batch is not None:
             # Run a new prefill batch
@@ -447,6 +446,7 @@ class Scheduler:
             len(self.running_batch.reqs) if self.running_batch is not None else 0
         )
         if running_bs >= self.max_running_requests:
+            self.batch_is_full = True
             return None
 
         # Get priority queue
@@ -490,9 +490,11 @@ class Scheduler:
                 )
                 > self.max_loras_per_batch
             ):
+                self.batch_is_full = True
                 break
 
             if adder.no_remaining_tokens():
+                self.batch_is_full = True
                 break
             req.init_next_round_input(None if prefix_computed else self.tree_cache)
             res = adder.add_one_req(req)
@@ -500,6 +502,7 @@ class Scheduler:
                 not res
                 or running_bs + len(adder.can_run_list) >= self.max_running_requests
             ):
+                self.batch_is_full = True
                 break
 
         can_run_list = adder.can_run_list
@@ -810,9 +813,6 @@ class Scheduler:
                 if req.top_logprobs_num > 0:
                     req.output_top_logprobs.append(logits_output.output_top_logprobs[i])
 
-        if not has_finished:
-            self.do_not_get_new_batch = True
-
         self.handle_finished_requests(batch)
 
     def handle_finished_requests(self, batch: ScheduleBatch):
@@ -833,6 +833,8 @@ class Scheduler:
         for i, req in enumerate(batch.reqs):
             if not req.finished() and req is not self.current_inflight_req:
                 unfinished_indices.append(i)
+            else:
+                self.batch_is_full = False
 
             if req.finished() or (
                 req.stream

python/sglang/test/test_utils.py

@@ -514,7 +514,16 @@ def get_similarities(vec1, vec2):
     return F.cosine_similarity(torch.tensor(vec1), torch.tensor(vec2), dim=0)
 
 
-def run_bench_serving(model, num_prompts, request_rate, other_server_args):
+def run_bench_serving(
+    model,
+    num_prompts,
+    request_rate,
+    other_server_args,
+    dataset_name="random",
+    random_input_len=4096,
+    random_output_len=2048,
+    disable_stream=False,
+):
     # Launch the server
     base_url = DEFAULT_URL_FOR_TEST
     process = popen_launch_server(
@@ -530,21 +539,21 @@ def run_bench_serving(model, num_prompts, request_rate, other_server_args):
         base_url=base_url,
         host=None,
         port=None,
-        dataset_name="random",
+        dataset_name=dataset_name,
         dataset_path="",
         model=None,
         tokenizer=None,
         num_prompts=num_prompts,
         sharegpt_output_len=None,
-        random_input_len=4096,
+        random_input_len=random_input_len,
-        random_output_len=2048,
+        random_output_len=random_output_len,
         random_range_ratio=0.0,
         request_rate=request_rate,
         multi=None,
         seed=0,
         output_file=None,
         disable_tqdm=False,
-        disable_stream=False,
+        disable_stream=disable_stream,
         disable_ignore_eos=False,
         extra_request_body=None,
     )

test/srt/test_bench_serving.py

@@ -20,7 +20,22 @@ class TestBenchServing(unittest.TestCase):
         )
 
         if is_in_ci():
-            assert res["output_throughput"] > 2600
+            assert res["output_throughput"] > 2830
+
+    def test_offline_throughput_non_stream_small_batch_size(self):
+        res = run_bench_serving(
+            model=DEFAULT_MODEL_NAME_FOR_TEST,
+            num_prompts=200,
+            request_rate=float("inf"),
+            dataset_name="sharegpt",
+            random_input_len=None,
+            random_output_len=None,
+            disable_stream=True,
+            other_server_args=["--max-running-requests", "10"],
+        )
+
+        if is_in_ci():
+            assert res["output_throughput"] > 1000
 
     def test_offline_throughput_without_radix_cache(self):
         res = run_bench_serving(
@@ -31,7 +46,7 @@ class TestBenchServing(unittest.TestCase):
         )
 
         if is_in_ci():
-            assert res["output_throughput"] > 2800
+            assert res["output_throughput"] > 2880
 
     def test_offline_throughput_without_chunked_prefill(self):
         res = run_bench_serving(
@@ -58,7 +73,7 @@ class TestBenchServing(unittest.TestCase):
         )
 
         if is_in_ci():
-            assert res["output_throughput"] > 2600
+            assert res["output_throughput"] > 2930
 
     def test_offline_throughput_default_fp8(self):
         res = run_bench_serving(