[diffusion] CI: restore public Qwen-Image 2.1 TP2 E2E coverage (#40507)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>

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

@@ -49,6 +49,7 @@ from sglang.multimodal_gen.test.test_utils import (
     DEFAULT_FLUX_2_KLEIN_BASE_4B_MODEL_NAME_FOR_TEST,
     DEFAULT_JOYAI_IMAGE_EDIT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOVA_360P_MODEL_NAME_FOR_TEST,
+    DEFAULT_QWEN_IMAGE_21_MODEL_NAME_FOR_TEST,
     DEFAULT_QWEN_IMAGE_EDIT_2509_MODEL_NAME_FOR_TEST,
     DEFAULT_QWEN_IMAGE_EDIT_2511_MODEL_NAME_FOR_TEST,
     DEFAULT_QWEN_IMAGE_EDIT_MODEL_NAME_FOR_TEST,
@@ -1141,26 +1142,25 @@ TWO_GPU_CASES = [
             ring_degree=2,
         ),
     ),
-    # TODO: re-enable when the checkpoint is accessible to fork PR CI
-    # DiffusionTestCase(
-    #     "qwen_image21_t2i_tp2",
-    #     DiffusionServerArgs(
-    #         model_path="Qwen/Qwen-Image-2.1",
-    #         tp_size=2,
-    #         ulysses_degree=1,
-    #         ring_degree=1,
-    #     ),
-    #     replace(
-    #         T2I_sampling_params,
-    #         output_size="1024x1024",
-    #         output_format="png",
-    #         extras={"num_inference_steps": 40, "guidance_scale": 1, "seed": 42},
-    #     ),
-    #     perf_repeat_requests=2,
-    #     run_perf_check=False,
-    #     run_component_accuracy_check=False,
-    #     run_t2v_input_reference_check=False,
-    # ),
+    DiffusionTestCase(
+        "qwen_image21_t2i_tp2",
+        DiffusionServerArgs(
+            model_path=DEFAULT_QWEN_IMAGE_21_MODEL_NAME_FOR_TEST,
+            tp_size=2,
+            ulysses_degree=1,
+            ring_degree=1,
+        ),
+        replace(
+            T2I_sampling_params,
+            output_size="1024x1024",
+            output_format="png",
+            extras={"num_inference_steps": 40, "guidance_scale": 1, "seed": 42},
+        ),
+        perf_repeat_requests=2,
+        run_perf_check=False,
+        run_component_accuracy_check=False,
+        run_t2v_input_reference_check=False,
+    ),
     DiffusionTestCase(
         "qwen_image_t2i_2_gpus_extra_high",
         DiffusionServerArgs(

python/sglang/multimodal_gen/test/server/perf_baselines/h100.json

@@ -206,9 +206,11 @@
             "estimated_full_test_time_s": 65.6
         },
         "qwen_image21_t2i_tp2": {
+            "expected_load_ms": 45596.77,
+            "notes": "H100 CI run 35549625911 attempt 3, job 106188864502, PR head 044e71f5ecb507ffdf66729f934b51c3dc68b59f; request 1 after server warmup, complete cached public model snapshot b3179ad355be050328e483a9dfdd9e60cd62adfa. Existing tolerances are unchanged.",
             "stages_ms": {},
             "denoise_step_ms": {},
-            "expected_e2e_ms": 0.0,
+            "expected_e2e_ms": 2914.82,
             "expected_avg_denoise_ms": 0.0,
             "expected_median_denoise_ms": 0.0,
             "estimated_full_test_time_s": 300.0

python/sglang/multimodal_gen/test/test_utils.py

@@ -40,7 +40,7 @@ logger = init_logger(__name__)
 # NPU/ascend) is read from sgl-project/ci-data-diffusion, where the GT-gen workflows
 # publish.
 SGL_TEST_FILES_CI_DATA_REPO = "sgl-project/ci-data-diffusion"
-SGL_TEST_FILES_CI_DATA_REVISION = "252710158cd4c74b7604808a385e93f7bced6d28"
+SGL_TEST_FILES_CI_DATA_REVISION = "38ba32bd812b2dfb0eccc83ef063096c089e3389"
 
 # The NPU pin is kept as a separate branch so ascend GT can be bumped independently
 # when it's regenerated on its own cadence.