[Frontend] Add omni benchmark support for image and video endpoints (#4728)

This commit is contained in:
ZacheryAU authored and GitHub committed 2026-09-08 09:28:21 +08:00
1 parent e1c59edec6
commit b1ec136bd7
21 files changed
+1626 -72

No files matched your search

+13 -3
View File
@@ -739,10 +739,16 @@ class StagePool:
def _infer_output_unit_type(self, request_outputs: list[Any], *, token_count: int) -> str:
final_output_type = getattr(self.stage_client, "final_output_type", None)
if self._has_image_output(request_outputs) or final_output_type in {"image", "images"}:
return "image"
if self._has_video_output(request_outputs) or final_output_type in {"video", "videos"}:
# Prefer declared modality over payload heuristics: video diffusion often
# stores frames in ``images`` (see serving_video / output_formatter).
if final_output_type in {"video", "videos"}:
return "video"
if final_output_type in {"image", "images"}:
return "image"
if self._has_video_output(request_outputs):
return "video"
if self._has_image_output(request_outputs):
return "image"
if self._has_audio_output(request_outputs) or final_output_type == "audio":
return "audio"
if self._has_trajectory_latent_output(request_outputs) or self._has_latent_output(request_outputs):
@@ -778,6 +784,10 @@ class StagePool:
total_videos = sum(self._count_videos(ro) for ro in request_outputs)
if total_videos > 0:
return total_videos
# Video payloads are commonly carried on ``images`` for diffusion.
total_images = sum(self._count_images(ro) for ro in request_outputs)
if total_images > 0:
return total_images
if unit_type == "latent":
total_latents = sum(
self._count_value_units(getattr(ro, "trajectory_latents", None))