Misc fix for min_p_sampling, --cuda-graph-bs (#2761)

python/sglang/srt/model_executor/model_runner.py

@@ -89,6 +89,7 @@ class ModelRunner:
         self.is_draft_worker = is_draft_worker
         self.is_generation = model_config.is_generation
         self.is_multimodal = model_config.is_multimodal
+        self.should_log = tp_rank == 0
         self.spec_algorithm = SpeculativeAlgorithm.from_string(
             server_args.speculative_algorithm
         )
@@ -117,15 +118,21 @@ class ModelRunner:
 
         if self.is_multimodal:
             self.mem_fraction_static *= 0.95
+            logger.info(
+                f"Automatically reduce --mem-fraction-static to {self.mem_fraction_static:.3f} "
+                f"because this is a multimodal model."
+            )
+
             if self.model_config.hf_config.architectures == [
                 "MllamaForConditionalGeneration"
             ]:
                 logger.info("Automatically turn off --chunked-prefill-size for mllama.")
                 server_args.chunked_prefill_size = -1
-            # TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
+
             if self.model_config.hf_config.architectures == [
                 "Qwen2VLForConditionalGeneration"
             ]:
+                # TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
                 logger.info(
                     "Automatically turn off --chunked-prefill-size and disable radix cache for qwen2-vl."
                 )