[diffusion] quant: support quant for all dits (#19156)

Co-authored-by: zyzshishui <zyzshishui@gmail.com>

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

@@ -316,11 +316,10 @@
             "expected_avg_denoise_ms": 38.25,
             "expected_median_denoise_ms": 35.95
         },
-        "flux_2_image_t2i_layerwise_offload": {
+        "layerwise_offload": {
             "stages_ms": {
                 "InputValidationStage": 0.06,
                 "TextEncodingStage": 513.58,
-                "ImageVAEEncodingStage": 0.0,
                 "LatentPreparationStage": 0.46,
                 "TimestepPreparationStage": 2.38,
                 "DenoisingStage": 52187.62,
@@ -537,6 +536,30 @@
             "expected_avg_denoise_ms": 83.75,
             "expected_median_denoise_ms": 93.58
         },
+        "zimage_image_t2i_fp8": {
+            "stages_ms": {
+                "InputValidationStage": 0.04,
+                "TextEncodingStage": 428.59,
+                "LatentPreparationStage": 0.14,
+                "TimestepPreparationStage": 47.26,
+                "DenoisingStage": 778.56,
+                "DecodingStage": 10.39
+            },
+            "denoise_step_ms": {
+                "0": 40.9,
+                "1": 61.08,
+                "2": 95.65,
+                "3": 95.83,
+                "4": 95.65,
+                "5": 96.09,
+                "6": 96.23,
+                "7": 96.04,
+                "8": 96.29
+            },
+            "expected_e2e_ms": 1370.28,
+            "expected_avg_denoise_ms": 85.97,
+            "expected_median_denoise_ms": 95.83
+        },
         "zimage_image_t2i_multi_lora": {
             "stages_ms": {
                 "InputValidationStage": 0.04,

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

@@ -295,6 +295,8 @@ class PerformanceSummary:
         )
 
 
+SMALL_T2I_MODEL = "Tongyi-MAI/Z-Image-Turbo"
+
 T2I_sampling_params = DiffusionSamplingParams(
     prompt="Doraemon is eating dorayaki",
     output_size="1024x1024",
@@ -397,9 +399,9 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
     # 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
     DiffusionTestCase(
-        "flux_2_image_t2i_layerwise_offload",
+        "layerwise_offload",
         DiffusionServerArgs(
-            model_path="black-forest-labs/FLUX.2-dev",
+            model_path=SMALL_T2I_MODEL,
             modality="image",
             dit_layerwise_offload=True,
             dit_offload_prefetch_size=2,
@@ -411,6 +413,15 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         DiffusionServerArgs(model_path="Tongyi-MAI/Z-Image-Turbo", modality="image"),
         T2I_sampling_params,
     ),
+    DiffusionTestCase(
+        "zimage_image_t2i_fp8",
+        DiffusionServerArgs(
+            model_path="Tongyi-MAI/Z-Image-Turbo",
+            modality="image",
+            extras=["--transformer-path MickJ/Z-Image-Turbo-fp8"],
+        ),
+        T2I_sampling_params,
+    ),
     # Multi-LoRA test case for Z-Image-Turbo
     DiffusionTestCase(
         "zimage_image_t2i_multi_lora",