[diffusion] fix: fix wrong validation on 2k resolution (#15478)

python/sglang/multimodal_gen/benchmarks/bench_serving.py

@@ -3,8 +3,18 @@ Benchmark online serving for diffusion models (Image/Video Generation).
 
 
 Usage:
-
+    # Video
     t2v:
+    python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
+         --backend sglang-video --dataset vbench --task t2v --num-prompts 20
+
+    i2v:
+    python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
+         --backend sglang-video --dataset vbench --task i2v --num-prompts 20
+
+
+    # Image
+    t2i:
     python3 -m sglang.multimodal_gen.benchmarks.bench_serving \
          --backend sglang-image --dataset vbench --task t2v --num-prompts 20
 
@@ -13,8 +23,6 @@ Usage:
          --backend sglang-image --dataset vbench --task i2v --num-prompts 20
 
 
-
-
 """
 
 import argparse

python/sglang/multimodal_gen/configs/sample/sampling_params.py

@@ -258,7 +258,7 @@ class SamplingParams:
 
         if pipeline_config.task_type.is_image_gen():
             # settle num_frames
-            logger.debug(f"Setting num_frames to 1 because this is an image-gen model")
+            logger.debug(f"num_frames set to 1 for image generation model")
             self.num_frames = 1
             self.data_type = DataType.IMAGE
         elif self.adjust_frames:

python/sglang/multimodal_gen/runtime/entrypoints/__init__.py

@@ -1 +1,7 @@
 # Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
+from sglang.multimodal_gen.runtime.utils.logging_utils import suppress_loggers
+
+# globally suppress some obsessive loggers
+suppress_loggers(
+    ["imageio", "imageio_ffmpeg", "PIL", "PIL_Image", "multipart", "filelock"]
+)

python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py

@@ -34,11 +34,8 @@ from sglang.multimodal_gen.runtime.utils.logging_utils import (
     init_logger,
     log_batch_completion,
     log_generation_timer,
-    suppress_loggers,
 )
 
-suppress_loggers(["imageio", "imageio_ffmpeg", "PIL", "PIL_Image"])
-
 logger = init_logger(__name__)
 
 # TODO: move to somewhere appropriate
@@ -52,7 +49,6 @@ except RuntimeError:
     pass
 
 
-# TODO: rename
 class DiffGenerator:
     """
     A unified class for generating images/videos using diffusion models.
@@ -413,9 +409,6 @@ class DiffGenerator:
             sync_scheduler_client.close()
             self.owns_scheduler_client = False
 
-    def __enter__(self):
-        return self
-
     def __exit__(self, exc_type, exc_val, exc_tb):
         self.shutdown()
 

python/sglang/multimodal_gen/runtime/pipelines_core/stages/input_validation.py

@@ -294,16 +294,6 @@ class InputValidationStage(PipelineStage):
         result = VerificationResult()
         result.add_check("height", batch.height, V.positive_int)
         result.add_check("width", batch.width, V.positive_int)
-
-        # Validate height and width
-        def check_size(value: int, name: str):
-            if value % (8 * server_args.num_gpus) != 0:
-                raise ValueError(
-                    f"{name} must be divisible by (8 x num_gpus) but {value} % (8 * {server_args.num_gpus}) != 0."
-                )
-
-        check_size(batch.height, "Height")
-        check_size(batch.width, "Width")
         result.add_check("seeds", batch.seeds, V.list_not_empty)
         result.add_check("generator", batch.generator, V.generator_or_list_generators)
         return result