Improve streaming, log_level, memory report, weight loading, and benchmark script (#7632)

Co-authored-by: Kan Wu <wukanustc@gmail.com>

python/sglang/srt/managers/scheduler_output_processor_mixin.py

@@ -521,11 +521,17 @@ class SchedulerOutputProcessorMixin:
                     stream_interval = (
                         req.sampling_params.stream_interval or self.stream_interval
                     )
-                    should_output = len(req.output_ids) % stream_interval == 0
+                    should_output = (
+                        len(req.output_ids) % stream_interval == 1
+                        if not self.model_config.is_multimodal_gen
+                        and stream_interval > 1
+                        else len(req.output_ids) % stream_interval == 0
+                    )
                 else:
                     should_output = (
                         len(req.output_ids) % DEFAULT_FORCE_STREAM_INTERVAL == 0
-                        and not self.model_config.is_multimodal_gen
+                        if not self.model_config.is_multimodal_gen
+                        else False
                     )
 
             if should_output: