[diffusion] fix: change default qwen-image vae precision to bf16 (#25411)

python/sglang/multimodal_gen/configs/pipeline_configs/flux.py

@@ -48,6 +48,7 @@ class FluxPipelineConfig(ImagePipelineConfig):
     dit_config: DiTConfig = field(default_factory=FluxConfig)
     # VAE
     vae_config: VAEConfig = field(default_factory=FluxVAEConfig)
+    vae_precision: str = "bf16"
 
     enable_autocast: bool = False
 

python/sglang/multimodal_gen/configs/pipeline_configs/ltx_2.py

@@ -179,8 +179,9 @@ class LTX2PipelineConfig(PipelineConfig):
 
     # Audio VAE configuration
     vae_config: LTXVideoVAEConfig = field(default_factory=LTXVideoVAEConfig)
+    vae_precision: str = "bf16"
     audio_vae_config: LTXAudioVAEConfig = field(default_factory=LTXAudioVAEConfig)
-    audio_vae_precision: str = "fp32"
+    audio_vae_precision: str = "bf16"
 
     @property
     def vae_scale_factor(self):

python/sglang/multimodal_gen/configs/pipeline_configs/mova.py

@@ -38,8 +38,9 @@ class MOVAPipelineConfig(PipelineConfig):
 
     # Video VAE (Wan) + Audio VAE (DAC)
     vae_config: WanVAEConfig = field(default_factory=WanVAEConfig)
+    vae_precision: str = "bf16"
     audio_vae_config: DacVAEConfig = field(default_factory=DacVAEConfig)
-    audio_vae_precision: str = "fp32"
+    audio_vae_precision: str = "bf16"
 
     # Text encoder (UMT5 compatible)
     text_encoder_configs: tuple = field(default_factory=lambda: (T5Config(),))

python/sglang/multimodal_gen/configs/pipeline_configs/qwen_image.py

@@ -148,6 +148,8 @@ class QwenImagePipelineConfig(QwenImageRolloutPipelineMixin, ImagePipelineConfig
 
     vae_sp: bool = False
 
+    vae_precision: str = "bf16"
+
     dit_config: DiTConfig = field(default_factory=QwenImageDitConfig)
     # VAE
     vae_config: VAEConfig = field(default_factory=QwenImageVAEConfig)

python/sglang/multimodal_gen/configs/pipeline_configs/zimage.py

@@ -67,6 +67,7 @@ class ZImagePipelineConfig(ZImageRolloutPipelineMixin, ImagePipelineConfig):
     task_type: ModelTaskType = ModelTaskType.T2I
     dit_config: DiTConfig = field(default_factory=ZImageDitConfig)
     vae_config: VAEConfig = field(default_factory=FluxVAEConfig)
+    vae_precision: str = "bf16"
     text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16",))
     text_encoder_configs: tuple[EncoderConfig, ...] = field(
         default_factory=lambda: (Qwen3TextConfig(),)

python/sglang/multimodal_gen/test/test_utils.py

@@ -33,7 +33,7 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
-SGL_TEST_FILES_CI_DATA_REVISION = "3ca3bad088ecc9ef80947d85c551cd335c75b87f"
+SGL_TEST_FILES_CI_DATA_REVISION = "c8305f1dd8cc82197f36c17d0f503adc94016cc7"
 SGL_TEST_FILES_CONSISTENCY_GT_ROOT = (
     "https://raw.githubusercontent.com/"
     f"sgl-project/ci-data/{SGL_TEST_FILES_CI_DATA_REVISION}/"