[diffusion] logging: improve logging (#19312)

python/sglang/cli/utils.py

@@ -47,7 +47,7 @@ def _maybe_download_model(
     from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import hf_hub_download
 
     if os.path.exists(model_name_or_path):
-        logger.info("Model already exists locally")
+        logger.debug("Model already exists locally")
         return model_name_or_path
 
     if not download:
@@ -57,7 +57,7 @@ def _maybe_download_model(
         # Try `model_index.json` first (diffusers models)
         source_hub = "MS Hub" if envs.SGLANG_USE_MODELSCOPE.get() else "HF Hub"
         try:
-            logger.info(
+            logger.debug(
                 "Downloading model_index.json from %s for %s...",
                 source_hub,
                 model_name_or_path,
@@ -67,14 +67,14 @@ def _maybe_download_model(
                 filename="model_index.json",
                 local_dir=local_dir,
             )
-            logger.info("Downloaded to %s", file_path)
+            logger.debug("Downloaded to %s", file_path)
             return os.path.dirname(file_path)
         except Exception as e_index:
             logger.debug("model_index.json not found or failed: %s", e_index)
 
         # Fallback to `config.json`
         try:
-            logger.info(
+            logger.debug(
                 "Downloading config.json from %s for %s...",
                 source_hub,
                 model_name_or_path,
@@ -84,7 +84,7 @@ def _maybe_download_model(
                 filename="config.json",
                 local_dir=local_dir,
             )
-            logger.info("Downloaded to %s", file_path)
+            logger.debug("Downloaded to %s", file_path)
             return os.path.dirname(file_path)
         except Exception as e_config:
             raise ValueError(

python/sglang/multimodal_gen/registry.py

@@ -423,13 +423,13 @@ def get_model_info(
             return None
 
     # 4. Combine and return the complete model info
-    logger.info("Using native sglang backend for model '%s'", model_path)
+    logger.debug("Using native sglang backend for model '%s'", model_path)
     model_info = ModelInfo(
         pipeline_cls=pipeline_cls,
         sampling_param_cls=config_info.sampling_param_cls,
         pipeline_config_cls=config_info.pipeline_config_cls,
     )
-    logger.info(f"Found model info: {model_info}")
+    logger.debug(f"Found model info: {model_info}")
 
     return model_info
 

python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py

@@ -125,35 +125,16 @@ class ComponentLoader(ABC):
             if isinstance(component, nn.Module):
                 component = component.eval()
             current_gpu_mem = current_platform.get_available_gpu_memory()
-            model_size = get_memory_usage_of_component(component)
+            model_size = get_memory_usage_of_component(component) or "NA"
             consumed = gpu_mem_before_loading - current_gpu_mem
-
-            # detect component device
-            try:
-                component_device = str(next(component.parameters()).device)
-                is_on_gpu = "cuda" in component_device
-            except (StopIteration, AttributeError):
-                is_on_gpu = False
-                component_device = "unknown"
-
-            if is_on_gpu:
             logger.info(
-                    f"Loaded %s: %s ({source} version). model size: %.2f GB, consumed GPU: %.2f GB, avail GPU mem: %.2f GB",
+                f"Loaded %s: %s ({source} version). model size: %s GB, consumed GPU mem: %.2f GB, avail GPU mem: %.2f GB",
                 component_name,
                 component.__class__.__name__,
                 model_size,
                 consumed,
                 current_gpu_mem,
             )
-            else:
-                logger.info(
-                    f"Loaded %s: %s ({source} version). model size: %.2f GB, device: %s, avail GPU mem: %.2f GB",
-                    component_name,
-                    component.__class__.__name__,
-                    model_size,
-                    component_device,
-                    current_gpu_mem,
-                )
         return component, consumed
 
     def load_native(

python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py

@@ -282,7 +282,7 @@ class Req:
         return pprint.pformat(asdict(self), indent=2, width=120)
 
     def log(self, server_args: ServerArgs):
-        if self.is_warmup:
+        if self.is_warmup or self.suppress_logs:
             return
         # TODO: in some cases (e.g., TI2I), height and weight might be undecided at this moment
         if self.height:
@@ -300,8 +300,8 @@ class Req:
             self.negative_prompt, key_hint="negative_prompt"
         )
 
-        # log non-sensitive parameters at info level
+        # Log sampling parameters
-        info_str = f"""Sampling params:
+        debug_str = f"""Sampling params:
                        width: {target_width}
                       height: {target_height}
                   num_frames: {self.num_frames}
@@ -319,15 +319,6 @@ class Req:
                  save_output: {self.save_output}
             output_file_path: {self.output_file_path()}
         """  # type: ignore[attr-defined]
-
-        # log full prompts at debug level only (for debugging purposes)
-        debug_str = f"""Full prompts:
-                      prompt: {self.prompt}
-                  neg_prompt: {self.negative_prompt}
-        """
-
-        if not self.suppress_logs:
-            logger.info(info_str)
         logger.debug(debug_str)
 
 

python/sglang/multimodal_gen/runtime/utils/hf_diffusers_utils.py

@@ -605,7 +605,7 @@ def maybe_download_model_index(model_name_or_path: str) -> dict[str, Any]:
             # Add the pipeline name for downstream use
             config["pipeline_name"] = config["_class_name"]
 
-            logger.info(
+            logger.debug(
                 "Downloaded model_index.json for %s, pipeline: %s",
                 model_name_or_path,
                 config["_class_name"],