[diffusion] UX: reduce per-request log noise (#35614)

python/sglang/multimodal_gen/runtime/models/adapter/ltx_2_duration_head.py

@@ -180,13 +180,6 @@ class LTX2DurationHead(nn.Module):
                 num_frames,
                 frame_rate,
             )
-        else:
-            logger.info(
-                "Predicted duration %.2fs (%d frames @ %.2f fps)",
-                seconds,
-                num_frames,
-                frame_rate,
-            )
         return num_frames
 
 

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

@@ -544,7 +544,7 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
             apply_attention_backend_override(layer, target)
         self.attn_backend = stage_backend
         self._attention_backend_active_override = target
-        logger.info(
+        logger.debug(
             "Attention backend for this batch: %s (%d layers switched)",
             target.name.lower() if target else "server default",
             len(layers),

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/longcat_image.py

@@ -285,7 +285,7 @@ class LongCatPromptRewriteStage(PipelineStage):
             self.text_encoder = text_encoder
 
             if enable_prompt_rewrite:
-                logger.info(
+                logger.debug(
                     "Prompt rewriting is enabled (enable_prompt_rewrite=True). "
                     "This runs autoregressive decoding on the Qwen2.5-VL text "
                     "encoder (up to %d tokens). Pass --enable-prompt-rewrite "

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/ltx_2/duration.py

@@ -84,7 +84,7 @@ class LTX2DurationStage(PipelineStage):
                 max_seconds=float(batch.auto_duration_max_seconds),
             )
 
-        logger.info(
+        logger.debug(
             "Auto-duration: %d frames (requested %d) @ %.2f fps",
             num_frames,
             int(batch.num_frames),