[diffusion] feat: support multiple LoRA adapters loading and application (#16667)

Co-authored-by: niehen6174 <niehen.6174@gmail.com>
Co-authored-by: Mick <mickjagger19@icloud.com>

python/sglang/multimodal_gen/test/server/perf_baselines.json

@@ -529,6 +529,31 @@
             "expected_avg_denoise_ms": 94.15,
             "expected_median_denoise_ms": 102.03
         },
+        "zimage_image_t2i_multi_lora": {
+            "stages_ms": {
+                "InputValidationStage": 0.04,
+                "TextEncodingStage": 103.81,
+                "ConditioningStage": 0.01,
+                "TimestepPreparationStage": 1.3,
+                "LatentPreparationStage": 0.11,
+                "DenoisingStage": 813.7,
+                "DecodingStage": 34.51
+            },
+            "denoise_step_ms": {
+                "0": 30.35,
+                "1": 74.53,
+                "2": 99.34,
+                "3": 100.92,
+                "4": 99.46,
+                "5": 100.57,
+                "6": 99.72,
+                "7": 100.86,
+                "8": 103.87
+            },
+            "expected_e2e_ms": 955.14,
+            "expected_avg_denoise_ms": 89.96,
+            "expected_median_denoise_ms": 99.72
+        },
         "zimage_image_t2i_2_gpus": {
             "stages_ms": {
                 "InputValidationStage": 0.08,

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

@@ -556,6 +556,76 @@ Consider updating perf_baselines.json with the snippets below:
         assert resp.status_code == 200, f"Dynamic set_lora failed: {resp.text}"
         logger.info("[Dynamic LoRA] set_lora succeeded for %s", case.id)
 
+    def _test_multi_lora_e2e(
+        self,
+        ctx: ServerContext,
+        case: DiffusionTestCase,
+        generate_fn: Callable[[str, openai.Client], str],
+        first_lora_path: str,
+        second_lora_path: str,
+    ) -> None:
+        """
+        Test multiple LoRA adapters with different set_lora input scenarios.
+        Tests: basic multi-LoRA, different strengths, cached adapters, switch back to single.
+        """
+        base_url = f"http://localhost:{ctx.port}/v1"
+        client = OpenAI(base_url=base_url, api_key="dummy")
+
+        # Test 1: Basic multi-LoRA with list format
+        resp = requests.post(
+            f"{base_url}/set_lora",
+            json={
+                "lora_nickname": ["default", "lora2"],
+                "lora_path": [first_lora_path, second_lora_path],
+                "target": "all",
+                "strength": [1.0, 1.0],
+            },
+        )
+        assert (
+            resp.status_code == 200
+        ), f"set_lora with multiple adapters failed: {resp.text}"
+        assert generate_fn(case.id, client) is not None
+
+        # Test 2: Different strengths
+        resp = requests.post(
+            f"{base_url}/set_lora",
+            json={
+                "lora_nickname": ["default", "lora2"],
+                "lora_path": [first_lora_path, second_lora_path],
+                "target": "all",
+                "strength": [0.8, 0.5],
+            },
+        )
+        assert (
+            resp.status_code == 200
+        ), f"set_lora with different strengths failed: {resp.text}"
+        assert generate_fn(case.id, client) is not None
+
+        # Test 3: Different targets
+        requests.post(f"{base_url}/set_lora", json={"lora_nickname": "default"})
+        resp = requests.post(
+            f"{base_url}/set_lora",
+            json={
+                "lora_nickname": ["default", "lora2"],
+                "lora_path": [first_lora_path, second_lora_path],
+                "target": ["transformer", "transformer_2"],
+                "strength": [0.8, 0.5],
+            },
+        )
+        assert (
+            resp.status_code == 200
+        ), f"set_lora with cached adapters failed: {resp.text}"
+        assert generate_fn(case.id, client) is not None
+
+        # Test 4: Switch back to single LoRA
+        resp = requests.post(f"{base_url}/set_lora", json={"lora_nickname": "default"})
+        assert (
+            resp.status_code == 200
+        ), f"set_lora back to single adapter failed: {resp.text}"
+        assert generate_fn(case.id, client) is not None
+
+        logger.info("[Multi-LoRA] All multi-LoRA tests passed for %s", case.id)
+
     def _test_v1_models_endpoint(
         self, ctx: ServerContext, case: DiffusionTestCase
     ) -> None:
@@ -681,3 +751,21 @@ Consider updating perf_baselines.json with the snippets below:
         # LoRA API functionality test with E2E validation (only for LoRA-enabled cases)
         if case.server_args.lora_path or case.server_args.dynamic_lora_path:
             self._test_lora_api_functionality(diffusion_server, case, generate_fn)
+
+            # Test dynamic LoRA switching (requires a second LoRA adapter)
+            if case.server_args.second_lora_path:
+                self._test_lora_dynamic_switch_e2e(
+                    diffusion_server,
+                    case,
+                    generate_fn,
+                    case.server_args.second_lora_path,
+                )
+
+                # Test multi-LoRA functionality
+                self._test_multi_lora_e2e(
+                    diffusion_server,
+                    case,
+                    generate_fn,
+                    case.server_args.lora_path,
+                    case.server_args.second_lora_path,
+                )

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

@@ -155,6 +155,9 @@ class DiffusionServerArgs:
     dynamic_lora_path: str | None = (
         None  # LoRA path for dynamic loading test (loaded via set_lora after startup)
     )
+    second_lora_path: str | None = (
+        None  # Second LoRA adapter path for multi-LoRA testing
+    )
     # misc
     enable_warmup: bool = False
 
@@ -361,6 +364,17 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         ),
         T2I_sampling_params,
     ),
+    # Multi-LoRA test case for Z-Image-Turbo
+    DiffusionTestCase(
+        "zimage_image_t2i_multi_lora",
+        DiffusionServerArgs(
+            model_path="Tongyi-MAI/Z-Image-Turbo",
+            modality="image",
+            lora_path="reverentelusarca/elusarca-anime-style-lora-z-image-turbo",
+            second_lora_path="tarn59/pixel_art_style_lora_z_image_turbo",
+        ),
+        T2I_sampling_params,
+    ),
     # === Text and Image to Image (TI2I) ===
     DiffusionTestCase(
         "qwen_image_edit_ti2i",