cleaner

Signed-off-by: Reagan <reaganjlee@gmail.com>

vllm/benchmarks/mm_processor.py

@@ -2,12 +2,12 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark multimodal processor latency.
 
-This benchmark measures the latency of the multimodal processor module
+This benchmark measures the latency of the mm processor module
 using randomly generated multimodal prompts with synthetic images.
 MM processor stats are automatically enabled.
 
 Run:
-    vllm bench multimodal-processor \
+    vllm bench mm-processor \
         --model <your_model> \
         --num-prompts 10 \
         --input-len 1024 \
@@ -55,7 +55,6 @@ class MultimodalProcessorBenchmarkMetrics:
 
 def collect_mm_processor_stats(
     llm_engine: Any,
-    debug: bool = False,
 ) -> dict[str, list[float]]:
     """
     Collect multimodal processor timing stats.
@@ -84,12 +83,6 @@ def collect_mm_processor_stats(
         )
         stats_by_stage["total_time"].append(stats_dict.get("total_time", 0.0))
 
-    if debug and not any(stats_by_stage.values()):
-        print(
-            "Warning: No MM processor stats found. "
-            "Ensure --enable-mm-processor-stats is set."
-        )
-
     return stats_by_stage
 
 
@@ -222,8 +215,6 @@ def benchmark_multimodal_processor(
 
     freeze_gc_heap()
 
-    debug = getattr(args, "debug_mm_stats", False)
-
     print(f"Processing {len(prompts)} requests...")
     start_time = time.perf_counter()
 
@@ -236,7 +227,6 @@ def benchmark_multimodal_processor(
 
     mm_stats_by_stage = collect_mm_processor_stats(
         llm.llm_engine,
-        debug=debug,
     )
 
     if not any(mm_stats_by_stage.values()):
@@ -350,11 +340,6 @@ def add_cli_args(parser: argparse.ArgumentParser) -> None:
         default=None,
         help="Path to save the benchmark results in JSON format.",
     )
-    parser.add_argument(
-        "--debug-mm-stats",
-        action="store_true",
-        help="Enable debug logging for MM processor stats collection.",
-    )
     parser.add_argument(
         "--metric-percentiles",
         type=str,
@@ -444,7 +429,7 @@ def main(args: argparse.Namespace) -> None:
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Benchmark multimodal processor latency"
+        description="Benchmark mm processor latency"
     )
     add_cli_args(parser)
     args = parser.parse_args()

vllm/config/observability.py

@@ -66,8 +66,8 @@ class ObservabilityConfig:
 
     enable_mm_processor_stats: bool = False
     """Enable collection of timing statistics for multimodal processor operations.
-    This can be useful for performance analysis and debugging. Defaults to `False`
+    This is for internal use only (e.g., benchmarks) and is not exposed as a CLI
-    (disabled)."""
+    argument. Defaults to `False` (disabled)."""
 
     @cached_property
     def collect_model_forward_time(self) -> bool:

vllm/engine/arg_utils.py

@@ -1041,10 +1041,9 @@ class EngineArgs:
             "--enable-layerwise-nvtx-tracing",
             **observability_kwargs["enable_layerwise_nvtx_tracing"],
         )
-        observability_group.add_argument(
+        # Note: --enable-mm-processor-stats is intentionally not exposed as a CLI
-            "--enable-mm-processor-stats",
+        # argument. It can be set programmatically via parser.set_defaults() for
-            **observability_kwargs["enable_mm_processor_stats"],
+        # internal use (e.g., benchmarks), but is not part of the public API.
-        )
 
         # Scheduler arguments
         scheduler_kwargs = get_kwargs(SchedulerConfig)

vllm/multimodal/registry.py

@@ -158,8 +158,6 @@ class MultiModalRegistry:
         if not model_config.is_multimodal_model:
             return {}
 
-        if observability_config is None:
-            observability_config = ObservabilityConfig()
         processor = self.create_processor(
             model_config, observability_config, cache=cache
         )
@@ -189,8 +187,6 @@ class MultiModalRegistry:
         if not model_config.is_multimodal_model:
             return {}
 
-        if observability_config is None:
-            observability_config = ObservabilityConfig()
         processor = self.create_processor(
             model_config, observability_config, cache=cache
         )
@@ -299,8 +295,6 @@ class MultiModalRegistry:
 
         The model is identified by `model_config`.
         """
-        if observability_config is None:
-            observability_config = ObservabilityConfig()
         processor = self.create_processor(
             model_config, observability_config, cache=cache
         )
@@ -337,8 +331,6 @@ class MultiModalRegistry:
 
         The model is identified by `model_config`.
         """
-        if observability_config is None:
-            observability_config = ObservabilityConfig()
         processor = self.create_processor(
             model_config, observability_config, cache=cache
         )