[diffusion] CI: add performance tracking job to nightly (#21091)

python/sglang/multimodal_gen/runtime/entrypoints/openai/image_api.py

@@ -136,6 +136,7 @@ async def generations(
             enable_upscaling=request.enable_upscaling,
             upscaling_model_path=request.upscaling_model_path,
             upscaling_scale=request.upscaling_scale,
+            perf_dump_path=request.perf_dump_path,
         )
         batch = prepare_request(
             server_args=server_args,

python/sglang/multimodal_gen/runtime/entrypoints/openai/protocol.py

@@ -53,6 +53,8 @@ class ImageGenerationsRequest(BaseModel):
     upscaling_model_path: Optional[str] = None
     upscaling_scale: Optional[int] = 4
     diffusers_kwargs: Optional[Dict[str, Any]] = None  # kwargs for diffusers backend
+    # Performance profiling
+    perf_dump_path: Optional[str] = None
 
 
 # Video API protocol models
@@ -109,6 +111,8 @@ class VideoGenerationsRequest(BaseModel):
     output_compression: Optional[int] = None
     output_path: Optional[str] = None
     diffusers_kwargs: Optional[Dict[str, Any]] = None  # kwargs for diffusers backend
+    # Performance profiling
+    perf_dump_path: Optional[str] = None
 
 
 class VideoListResponse(BaseModel):

python/sglang/multimodal_gen/runtime/entrypoints/openai/video_api.py

@@ -81,6 +81,7 @@ def _build_video_sampling_params(request_id: str, request: VideoGenerationsReque
         output_path=request.output_path,
         output_compression=request.output_compression,
         output_quality=request.output_quality,
+        perf_dump_path=request.perf_dump_path,
     )
 
 

python/sglang/multimodal_gen/runtime/layers/lora/linear.py

@@ -33,7 +33,7 @@ from sglang.multimodal_gen.runtime.layers.vocab_parallel_embedding import (
 )
 from sglang.multimodal_gen.utils import get_mixed_precision_state
 
-torch._dynamo.config.recompile_limit = 16
+torch._dynamo.config.recompile_limit = 64
 
 
 class BaseLayerWithLoRA(nn.Module):

python/sglang/multimodal_gen/runtime/managers/gpu_worker.py

@@ -297,6 +297,19 @@ class GPUWorker:
                 # Avoid logging warmup perf records that share the same request_id.
                 if not req.is_warmup:
                     PerformanceLogger.log_request_summary(metrics=output_batch.metrics)
+
+            # dump per-request perf report to specified file (server mode)
+            if (
+                req.perf_dump_path is not None
+                and not req.is_warmup
+                and output_batch.metrics is not None
+            ):
+                PerformanceLogger.dump_benchmark_report(
+                    file_path=req.perf_dump_path,
+                    metrics=output_batch.metrics,
+                    meta={"model": self.server_args.model_path},
+                    tag="server_perf_dump",
+                )
         except Exception as e:
             logger.error(
                 f"Error executing request {req.request_id}: {e}", exc_info=True