[CI][Perf] migrate diffusion DFX benches to vllm bench serve --omni (#7737)

Signed-off-by: wangyu <410167048@qq.com>
This commit is contained in:
wangyu authored and GitHub committed 2026-09-21 19:16:39 -07:00
1 parent a4ad73e08b
commit 0570b0da7e
30 files changed
+2040 -729

No files matched your search

@@ -1730,7 +1730,10 @@ class Cosmos3OmniDiffusersPipeline(
},
},
}
return DiffusionOutput(output=action_output)
return DiffusionOutput(
output=action_output,
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
@staticmethod
def _truthy(value) -> bool:
@@ -3610,6 +3613,7 @@ class Cosmos3OmniDiffusersPipeline(
"payload": {"video": output_video},
"metadata": {"video": {"fps": frame_rate}},
},
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
full_output = torch.cat(output_chunks, dim=2)[:, :, :total_frames]
@@ -3636,6 +3640,7 @@ class Cosmos3OmniDiffusersPipeline(
"video": {"fps": frame_rate},
},
},
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
# -- Forward (main generation entry point) -------------------------------
@@ -4072,7 +4077,10 @@ class Cosmos3OmniDiffusersPipeline(
if _is_rank_zero():
logger.info("Sound tokenizer decoded in %.2fs", time.time() - sound_decode_start)
logger.info("Total pipeline time: %.2fs", time.time() - pipeline_start)
return DiffusionOutput(output={"video": video, "audio": audio, "audio_sample_rate": sound_sample_rate})
return DiffusionOutput(
output={"video": video, "audio": audio, "audio_sample_rate": sound_sample_rate},
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
if action_enabled:
if action_latents is None or raw_action_dim is None or domain_id is None:
@@ -4092,6 +4100,10 @@ class Cosmos3OmniDiffusersPipeline(
},
},
},
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
return DiffusionOutput(output={"image": video} if is_t2i else {"video": video})
return DiffusionOutput(
output={"image": video} if is_t2i else {"video": video},
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
from __future__ import annotations
@@ -548,7 +548,10 @@ class HunyuanVideo15Pipeline(
latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
output = self.vae.decode(latents, return_dict=False)[0]
return DiffusionOutput(output=output)
return DiffusionOutput(
output=output,
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
from __future__ import annotations
@@ -643,7 +643,10 @@ class HunyuanVideo15I2VPipeline(
latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
output = self.vae.decode(latents, return_dict=False)[0]
return DiffusionOutput(output=output)
return DiffusionOutput(
output=output,
stage_durations=self.stage_durations if hasattr(self, "stage_durations") else None,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)