[REFACT]Refactor diffusion outputs to payload metadata (#4922)

Signed-off-by: bjf-frz <frz123db@gmail.com>
This commit is contained in:
bjf-frz authored and GitHub committed 2026-07-14 17:24:03 +08:00
1 parent 4b9d09cb87
commit 1a4a042a83
44 files changed
+1373 -478

No files matched your search

+12 -11
View File
@@ -1198,8 +1198,10 @@ class DiffusionOutput:
Final output (after pipeline completion)
"""
# Fields may be replaced with SHM handle dicts by ipc.pack_diffusion_output_shm
output: torch.Tensor | tuple[Any, ...] | dict[str, Any] | None = None
# Legacy compatibility fields. New pipeline-specific payloads should be
# carried by output["payload"] instead.
trajectory_timesteps: torch.Tensor | dict[str, Any] | None = None
trajectory_latents: torch.Tensor | dict[str, Any] | None = None
trajectory_log_probs: torch.Tensor | dict[str, Any] | None = None
@@ -1212,10 +1214,6 @@ class DiffusionOutput:
post_process_func: Callable[..., Any] | None = None
# Extra custom output data (e.g. latent trajectories, prompt embeds)
# passed through to OmniRequestOutput.custom_output
custom_output: dict[str, Any] = field(default_factory=dict)
# logged timings info, directly from Req.timings
# timings: Optional["RequestTimings"] = None
@@ -1230,10 +1228,9 @@ class DiffusionOutput:
# memory usage info
peak_memory_mb: float = 0.0
# When True, move all tensor fields (including tensors inside
# ``custom_output``) to CPU at construction time. Useful when the output
# is shipped across process boundaries (e.g. step-execution mode) and the
# receiving side must not initialise a stray CUDA context.
# When True, move tensor fields to CPU at construction time. Useful when
# the output is shipped across process boundaries (e.g. step-execution
# mode) and the receiving side must not initialise a stray CUDA context.
to_cpu: bool = False
def __post_init__(self) -> None:
@@ -1243,14 +1240,18 @@ class DiffusionOutput:
def _maybe_to_cpu(value: Any) -> Any:
if isinstance(value, torch.Tensor):
return value.detach().cpu()
if isinstance(value, dict):
return {key: _maybe_to_cpu(item) for key, item in value.items()}
if isinstance(value, list):
return [_maybe_to_cpu(item) for item in value]
if isinstance(value, tuple):
return tuple(_maybe_to_cpu(item) for item in value)
return value
self.output = _maybe_to_cpu(self.output)
self.trajectory_timesteps = _maybe_to_cpu(self.trajectory_timesteps)
self.trajectory_latents = _maybe_to_cpu(self.trajectory_latents)
self.trajectory_log_probs = _maybe_to_cpu(self.trajectory_log_probs)
if self.custom_output:
self.custom_output = {k: _maybe_to_cpu(v) for k, v in self.custom_output.items()}
@classmethod
def from_exception(cls, exc: BaseException) -> "DiffusionOutput":