[Bug][CI] Pin the LTX FP8 quality gate to its established sigma schedule (#5302)
Signed-off-by: mglyn <1203789601@qq.com>
This commit is contained in:
1 parent
24eb4ab89c
commit
39e83e27f0
7 files changed
+134
-19
No files matched your search
@@ -177,6 +177,17 @@ class I2VVideoAudioScheduler:
|
||||
return ((video_out, audio_out),)
|
||||
|
||||
|
||||
def _set_scheduler_sigmas(scheduler: Any, sigmas: torch.Tensor) -> torch.Tensor:
|
||||
sigmas = sigmas.to(torch.float32)
|
||||
timesteps = sigmas[:-1] * scheduler.config.get("num_train_timesteps", 1000)
|
||||
scheduler.sigmas = sigmas
|
||||
scheduler.timesteps = timesteps
|
||||
scheduler.num_inference_steps = len(timesteps)
|
||||
scheduler._step_index = None
|
||||
scheduler._begin_index = None
|
||||
return timesteps
|
||||
|
||||
|
||||
def prepare_scheduler_stage(
|
||||
pipeline: Any,
|
||||
request_inputs: LTXRequestInputs,
|
||||
@@ -188,18 +199,9 @@ def prepare_scheduler_stage(
|
||||
latent_height: int,
|
||||
latent_width: int,
|
||||
) -> tuple[Any, Any, torch.Tensor]:
|
||||
sigmas = (
|
||||
np.linspace(1.0, 1 / request_inputs.num_inference_steps, request_inputs.num_inference_steps)
|
||||
if sigmas is None
|
||||
else sigmas
|
||||
)
|
||||
mu = calculate_shift(
|
||||
pipeline.scheduler.config.get("max_image_seq_len", 4096),
|
||||
pipeline.scheduler.config.get("base_image_seq_len", 1024),
|
||||
pipeline.scheduler.config.get("max_image_seq_len", 4096),
|
||||
pipeline.scheduler.config.get("base_shift", 0.95),
|
||||
pipeline.scheduler.config.get("max_shift", 2.05),
|
||||
)
|
||||
if sigmas is not None and timesteps is not None:
|
||||
raise ValueError("Only one of `sigmas` or `timesteps` may be provided.")
|
||||
|
||||
audio_scheduler = copy.deepcopy(pipeline.scheduler)
|
||||
video_audio_scheduler = pipeline._make_video_audio_scheduler(
|
||||
audio_scheduler,
|
||||
@@ -207,6 +209,27 @@ def prepare_scheduler_stage(
|
||||
latent_height,
|
||||
latent_width,
|
||||
)
|
||||
if sigmas is not None:
|
||||
scheduler_sigmas = torch.as_tensor(sigmas, dtype=torch.float32, device=device)
|
||||
if scheduler_sigmas.ndim != 1 or scheduler_sigmas.numel() < 2:
|
||||
raise ValueError("An LTX custom sigma schedule must contain at least two boundary values.")
|
||||
if scheduler_sigmas[-1] != 0:
|
||||
scheduler_sigmas = torch.cat([scheduler_sigmas, scheduler_sigmas.new_zeros(1)])
|
||||
timesteps_tensor = _set_scheduler_sigmas(pipeline.scheduler, scheduler_sigmas)
|
||||
_set_scheduler_sigmas(audio_scheduler, scheduler_sigmas.clone())
|
||||
return audio_scheduler, video_audio_scheduler, timesteps_tensor
|
||||
|
||||
sigmas = np.linspace(1.0, 1 / request_inputs.num_inference_steps, request_inputs.num_inference_steps)
|
||||
# Official LTX2 one-stage omits the latent, so mu stays at max_shift across resolutions.
|
||||
# Mu varies with latent token count only when the scheduler receives a latent, as in official HQ two-stage;
|
||||
# explicit sigma schedules bypass mu entirely.
|
||||
mu = calculate_shift(
|
||||
pipeline.scheduler.config.get("max_image_seq_len", 4096),
|
||||
pipeline.scheduler.config.get("base_image_seq_len", 1024),
|
||||
pipeline.scheduler.config.get("max_image_seq_len", 4096),
|
||||
pipeline.scheduler.config.get("base_shift", 0.95),
|
||||
pipeline.scheduler.config.get("max_shift", 2.05),
|
||||
)
|
||||
retrieve_timesteps(
|
||||
audio_scheduler,
|
||||
request_inputs.num_inference_steps,
|
||||
|
||||
@@ -71,6 +71,17 @@ class LTXRequestMixin:
|
||||
supports_request_batch = False
|
||||
supports_guidance_rescale = False
|
||||
|
||||
@staticmethod
|
||||
def _resolve_request_sigmas(
|
||||
req: DiffusionRequestBatch,
|
||||
fallback: list[float] | None,
|
||||
) -> list[float] | None:
|
||||
request_sigmas = [sampling.sigmas for sampling in req.sampling_params_list]
|
||||
first = request_sigmas[0] if request_sigmas else None
|
||||
if any(item != first for item in request_sigmas[1:]):
|
||||
raise ValueError("Batched LTX requests must use identical custom sigmas.")
|
||||
return first if first is not None else fallback
|
||||
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt: str | list[str] | None,
|
||||
|
||||
@@ -96,6 +96,7 @@ class LTXOneStagePipeline(LTXPipelineRuntime):
|
||||
image: Any | None = None,
|
||||
) -> DiffusionOutput | list[DiffusionOutput]:
|
||||
del return_dict
|
||||
sigmas = self._resolve_request_sigmas(req, sigmas)
|
||||
request_inputs = self._resolve_request_inputs(
|
||||
req,
|
||||
prompt=prompt,
|
||||
|
||||
Reference in new issue
Block a user