[CI] Fix stage-b-test-1-gpu-large (0) timeout by reordering LoRA tests and using tokenizer from cache (#22292)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

python/sglang/test/test_utils.py

@@ -1129,12 +1129,26 @@ def run_bench_serving(
         other_args=other_server_args,
     )
 
+    # Resolve tokenizer to local snapshot path when available, so the benchmark
+    # client's AutoTokenizer.from_pretrained uses the local path directly instead
+    # of calling the HF Hub API (which can stall for minutes in CI).
+    bench_tokenizer = tokenizer
+    if bench_tokenizer is None:
+        try:
+            from sglang.srt.utils import find_local_repo_dir
+
+            local_dir = find_local_repo_dir(model, revision=None)
+            if local_dir and os.path.isdir(local_dir):
+                bench_tokenizer = local_dir
+        except Exception:
+            pass
+
     # Run benchmark
     args = get_benchmark_args(
         base_url=base_url,
         dataset_name=dataset_name,
         dataset_path=dataset_path,
-        tokenizer=tokenizer,
+        tokenizer=bench_tokenizer,
         num_prompts=num_prompts,
         random_input_len=random_input_len,
         random_output_len=random_output_len,

test/registered/perf/test_bench_serving_1gpu_part1.py

@@ -141,7 +141,7 @@ class TestBenchServing1GPUPart1(CustomTestCase):
                 self.assertLess(res["median_ttft_ms"], 86)
             self.assertLess(res["median_itl_ms"], 10)
 
-    def test_lora_online_latency(self):
+    def test_online_lora_latency(self):
         if is_in_amd_ci():
             pass
 
@@ -149,14 +149,14 @@ class TestBenchServing1GPUPart1(CustomTestCase):
 
         if is_in_ci():
             write_github_step_summary(
-                f"### test_lora_online_latency\n"
+                f"### test_online_lora_latency\n"
                 f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
                 f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
             )
             self.assertLess(res["median_e2e_latency_ms"], 2400)
             self.assertLess(res["median_ttft_ms"], 58)
 
-    def test_lora_online_latency_with_concurrent_adapter_updates(self):
+    def test_online_lora_latency_with_concurrent_adapter_updates(self):
         if is_in_amd_ci():
             pass
 
@@ -164,7 +164,7 @@ class TestBenchServing1GPUPart1(CustomTestCase):
 
         if is_in_ci():
             write_github_step_summary(
-                f"### test_lora_online_latency\n"
+                f"### test_online_lora_latency_with_concurrent_adapter_updates\n"
                 f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
                 f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
             )