[diffusion] CI: apply diffusers backend in lora case (#22157)

Co-authored-by: daiweitao <dwti614707404@163.com>

python/sglang/multimodal_gen/test/server/test_server_common.py

@@ -43,6 +43,14 @@ from sglang.multimodal_gen.test.test_utils import (
 logger = init_logger(__name__)
 
 
+def _is_lora_case(case: DiffusionTestCase) -> bool:
+    return bool(
+        case.server_args.lora_path
+        or case.server_args.dynamic_lora_path
+        or case.server_args.second_lora_path
+    )
+
+
 @pytest.fixture
 def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     """Start a diffusion server for a single case and tear it down afterwards."""
@@ -65,9 +73,9 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     sampling_params = case.sampling_params
     extra_args = os.environ.get("SGLANG_TEST_SERVE_ARGS", "")
 
-    # In GT generation mode, force --backend diffusers
+    # Keep LoRA GT on the normal backend path so adapter state matches CI.
     if os.environ.get("SGLANG_GEN_GT", "0") == "1":
-        if "--backend" not in extra_args:
+        if not _is_lora_case(case) and "--backend" not in extra_args:
             extra_args = "--backend diffusers " + extra_args.strip()
 
     extra_args += f" --num-gpus {server_args.num_gpus}"
@@ -853,9 +861,8 @@ Consider updating perf_baselines.json with the snippets below:
         # Check if we're in GT generation mode
         is_gt_gen_mode = os.environ.get("SGLANG_GEN_GT", "0") == "1"
 
-        # Dynamic LoRA loading test - tests LayerwiseOffload + set_lora interaction
+        # GT generation also needs the dynamic set_lora step before generation.
-        # Server starts WITHOUT lora_path, then set_lora is called after startup
+        if case.run_lora_dynamic_load_check:
-        if case.run_lora_dynamic_load_check and not is_gt_gen_mode:
             self._test_dynamic_lora_loading(diffusion_server, case)
 
         generate_fn = get_generate_fn(