sgl-project/sglang · error · ValueError

auto_duration was requested but this checkpoint has no durat

Error message

auto_duration was requested but this checkpoint has no duration head. It ships from LTX-2.5 onward.

What it means

The LTX-2 duration stage short-circuits unless batch.auto_duration is set; when it is set but the loaded checkpoint has no duration head (self.duration_head is None), automatic duration prediction is impossible and it raises, noting duration heads ship from LTX-2.5 onward.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/ltx_2/duration.py:55

        if self.duration_head is None:
            return []
        dtype = resolve_precision(
            server_args, "duration_head", precision_attr="dit_precision"
        )
        return [
            ComponentUse(
                self._component_stage_name(stage_name),
                "duration_head",
                target_dtype=dtype,
            )
        ]

    def forward(self, batch: Req, server_args: ServerArgs) -> Req:
        if not batch.auto_duration:
            return batch

        if self.duration_head is None:
            raise ValueError(
                "auto_duration was requested but this checkpoint has no duration "
                "head. It ships from LTX-2.5 onward."
            )

        video_tokens = batch.prompt_embeds
        audio_tokens = batch.audio_prompt_embeds
        if isinstance(video_tokens, list):
            video_tokens = video_tokens[0]
        if isinstance(audio_tokens, list):
            audio_tokens = audio_tokens[0]

        # A CFG batch carries [negative, positive] with duplicated rows, so
        # predict from the first positive row only.
        with (
            self.use_declared_component(
                component_name="duration_head", module=self.duration_head
            ) as duration_head,
            torch.no_grad(),

View on GitHub (pinned to 0132848349)

Solutions

1. Disable auto_duration for this checkpoint (unset the flag / set False)
2. Upgrade to an LTX-2.5+ checkpoint that includes the duration head
3. Verify the checkpoint actually loaded the duration head module (self.duration_head is not None) if you believe it should have one

Example fix

# before
req.auto_duration = True  # LTX-2.0 checkpoint -> raises
# after
req.auto_duration = False
# or serve an LTX-2.5+ checkpoint that has a duration head

if batch.auto_duration and stage.duration_head is None:
    batch.auto_duration = False  # or raise with a clear message

def supports_auto_duration(stage) -> bool:
    return getattr(stage, "duration_head", None) is not None

When it happens

Trigger: Setting auto_duration=True on a request while running a pre-LTX-2.5 checkpoint that lacks the duration head weights.

Common situations: Using a newer client/flag against an older LTX-2 checkpoint; enabling auto_duration globally while serving multiple checkpoints; checkpoint download missing the duration head module.

Related errors

• Serve backend {name!r} uses API version {backend.api_version
• scalar_type_id {scalar_type_id} doesn't exists.
• padding_side must be 'left' or 'right', got {padding_side}
• Unsupported text encoder output: expected `hidden_states`.
• num_inference_steps must be positive, got {steps}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1bf354ef6713a8ba. Report an issue: GitHub.