[diffusion] feat: support default 4-step inference for Flux2-Klein distilled models (#17225)

Signed-off-by: Lancer <maruixiang6688@gmail.com>

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

@@ -304,6 +304,27 @@
             "expected_avg_denoise_ms": 520.09,
             "expected_median_denoise_ms": 528.0
         },
+        "flux_2_klein_image_t2i": {
+            "stages_ms": {
+                "InputValidationStage": 0.05,
+                "TextEncodingStage": 530.93,
+                "ImageVAEEncodingStage": 0.0,
+                "ConditioningStage": 0.02,
+                "LatentPreparationStage": 12.71,
+                "TimestepPreparationStage": 2.91,
+                "DenoisingStage": 2112.24,
+                "DecodingStage": 489.8
+            },
+            "denoise_step_ms": {
+                "0": 511.3,
+                "1": 541.19,
+                "2": 518.93,
+                "3": 541.2
+            },
+            "expected_e2e_ms": 3148.75,
+            "expected_avg_denoise_ms": 528.06,
+            "expected_median_denoise_ms": 526.56
+        },
         "flux_2_image_t2i_layerwise_offload": {
             "stages_ms": {
                 "InputValidationStage": 0.06,

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

@@ -349,6 +349,14 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         ),
         T2I_sampling_params,
     ),
+    DiffusionTestCase(
+        "flux_2_klein_image_t2i",
+        DiffusionServerArgs(
+            model_path="black-forest-labs/FLUX.2-klein-4B",
+            modality="image",
+        ),
+        T2I_sampling_params,
+    ),
     # TODO: replace with a faster model to test the --dit-layerwise-offload
     # TODO: currently, we don't support sending more than one request in test, and setting `num_outputs_per_prompt` to 2 doesn't guarantee the denoising be executed twice,
     # so we do one warmup and send one request instead