[diffusion] CI: improve diffusion comparison benchmark setting for realistic perf and auto-discover ut (#22086)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

python/sglang/cli/utils.py

@@ -78,6 +78,9 @@ def get_is_diffusion_model(model_path: str) -> bool:
     if is_known_non_diffusers_diffusion_model(model_path):
         return True
 
+    if _is_registered_diffusion_model(model_path):
+        return True
+
     try:
         if envs.SGLANG_USE_MODELSCOPE.get():
             from modelscope import model_file_download

python/sglang/multimodal_gen/runtime/launch_server.py

@@ -88,7 +88,7 @@ def launch_server(server_args: ServerArgs, launch_http_server: bool = True):
         result_pipes_from_slaves_w.append(w)
 
     # Launch all worker processes
-    master_port = server_args.master_port or (server_args.master_port + 100)
+    master_port = server_args.master_port
     scheduler_pipe_readers = []
     scheduler_pipe_writers = []
 

python/sglang/multimodal_gen/runtime/server_args.py

@@ -188,8 +188,7 @@ class ServerArgs:
     )
 
     # Master port for distributed inference
-    # TODO: do not hard code
-    master_port: int | None = None
+    master_port: int = 30005
 
     # http server endpoint config
     host: str | None = "127.0.0.1"
@@ -386,36 +385,27 @@ class ServerArgs:
                 "Warmup enabled, the launch time is expected to be longer than usual"
             )
 
+    @staticmethod
+    def _require_port(port: int, name: str) -> None:
+        """Raise if *port* is occupied (used under ``--strict-ports``)."""
+        if not is_port_available(port):
+            raise RuntimeError(
+                f"{name} port {port} is unavailable and --strict-ports is enabled. "
+                f"Either use a different port or disable --strict-ports."
+            )
+
     def _adjust_network_ports(self):
         if self.strict_ports:
-            # Strict mode: fail if port is unavailable
-            if not is_port_available(self.port):
-                raise RuntimeError(
-                    f"Port {self.port} is unavailable and --strict-ports is enabled. "
-                    f"Either use a different port or remove --strict-ports to allow auto-selection."
-                )
-            if not is_port_available(self.scheduler_port):
-                raise RuntimeError(
-                    f"Scheduler port {self.scheduler_port} is unavailable and --strict-ports is enabled. "
-                    f"Either use a different port or remove --strict-ports to allow auto-selection."
-                )
-            if self.master_port is not None and not is_port_available(self.master_port):
-                raise RuntimeError(
-                    f"Master port {self.master_port} is unavailable and --strict-ports is enabled. "
-                    f"Either use a different port or remove --strict-ports to allow auto-selection."
-                )
+            self._require_port(self.port, "HTTP")
+            self._require_port(self.scheduler_port, "Scheduler")
+            self._require_port(self.master_port, "Master")
         else:
             self.port = self.settle_port(self.port)
             initial_scheduler_port = self.scheduler_port + (
                 random.randint(0, 100) if self.scheduler_port == 5555 else 0
             )
             self.scheduler_port = self.settle_port(initial_scheduler_port)
-            initial_master_port = (
-                self.master_port
-                if self.master_port is not None
-                else (30005 + random.randint(0, 100))
-            )
-            self.master_port = self.settle_port(initial_master_port, 37)
+            self.master_port = self.settle_port(self.master_port, 37)
 
     def _adjust_parallelism(self):
         if self.tp_size is None:

python/sglang/multimodal_gen/test/run_suite.py

@@ -28,17 +28,21 @@ _UPDATE_WEIGHTS_MODEL_PAIR_IDS = (
     "Qwen-Image",
 )
 
+
+def _discover_unit_tests() -> list[str]:
+    """Auto-discover all test_*.py files in the unit/ directory."""
+    unit_dir = Path(__file__).resolve().parent / "unit"
+    if not unit_dir.is_dir():
+        return []
+    return sorted(
+        f"../unit/{f.name}" for f in unit_dir.glob("test_*.py") if f.is_file()
+    )
+
+
 SUITES = {
     # no GPU required; safe to run on any CPU-only runner
-    "unit": [
-        "../unit/test_sampling_params.py",
-        "../unit/test_storage.py",
-        "../unit/test_lora_format_adapter.py",
-        "../unit/test_server_args.py",
-        "../unit/test_input_validation.py",
-        "../unit/test_resolve_prompts.py",
-        # add new unit tests here
-    ],
+    # Auto-discovered from test/unit/test_*.py
+    "unit": _discover_unit_tests(),
     "1-gpu": [
         "test_server_a.py",
         "test_server_b.py",

python/sglang/multimodal_gen/test/server/test_server_common.py

@@ -102,6 +102,10 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     if server_args.enable_warmup:
         extra_args += " --warmup"
 
+    # Strict ports: fail immediately if port is occupied instead of silently
+    # picking another one (which causes the test client to connect to the wrong server).
+    extra_args += " --strict-ports"
+
     for arg in server_args.extras:
         extra_args += f" {arg}"
 

python/sglang/multimodal_gen/test/server/test_server_utils.py

@@ -375,8 +375,10 @@ class ServerManager:
         # Apply custom environment variables
         env.update(self.env_vars)
 
-        # TODO: unify with run_command
-        logger.info(f"Running command: {shlex.join(command)}")
+        cmd_str = shlex.join(command)
+        # Use print (not logger) so the command always appears in CI output
+        # regardless of log-level configuration.
+        print(f"[server-test] Running command: {cmd_str}", flush=True)
 
         process = subprocess.Popen(
             command,
@@ -412,11 +414,10 @@ class ServerManager:
             log_thread.daemon = True
             log_thread.start()
 
-        logger.info(
-            "[server-test] Starting server pid=%s, model=%s, log=%s",
-            process.pid,
-            self.model,
-            stdout_path,
+        print(
+            f"[server-test] Starting server pid={process.pid}, "
+            f"model={self.model}, log={stdout_path}",
+            flush=True,
         )
 
         self._wait_for_ready(process, stdout_path)

python/sglang/srt/multimodal/processors/lfm2_vl.py

@@ -12,9 +12,9 @@
 # limitations under the License.
 """Multimodal processor for LFM2-VL models with SigLip2 NaFlex support."""
 
-from typing import Any, Dict, List, Optional, Union
+from typing import List, Union
 
-from sglang.srt.managers.schedule_batch import Modality
+from sglang.srt.managers.schedule_batch import Modality, MultimodalProcessorOutput
 from sglang.srt.models.lfm2_vl import Lfm2VlForConditionalGeneration
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor as SGLangBaseProcessor,
@@ -56,7 +56,7 @@ class Lfm2VlImageProcessor(SGLangBaseProcessor):
         input_text: str,
         request_obj,
         **kwargs,
-    ) -> Optional[Dict[str, Any]]:
+    ):
         if not image_data:
             input_ids = self._tokenizer(
                 input_text, return_tensors="pt", add_special_tokens=False
@@ -77,8 +77,8 @@ class Lfm2VlImageProcessor(SGLangBaseProcessor):
             base_output, self.mm_tokens
         )
 
-        return {
-            "input_ids": input_ids.tolist(),
-            "mm_items": mm_items,
-            "im_token_id": self.IMAGE_TOKEN_ID,
-        }
+        return MultimodalProcessorOutput(
+            input_ids=input_ids.tolist(),
+            mm_items=mm_items,
+            im_token_id=self.IMAGE_TOKEN_ID,
+        )