sgl-project/sglang · critical · ValueError

LTX-2 audio scheduler was not prepared.

Error message

LTX-2 audio scheduler was not prepared.

What it means

_before_denoising_loop asserts that the LTX-2 audio scheduler (ctx.audio_scheduler) was constructed during preparation. LTX-2 does joint video+audio denoising, so a missing audio scheduler means the prep stage never ran or failed silently.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/ltx_2/denoising.py:1690

        # cfg-parallel batches are broadcast from rank 0; scheduler state must be local
        for name in ("sigmas", "timesteps"):
            value = getattr(scheduler, name, None)
            if isinstance(value, torch.Tensor):
                setattr(scheduler, name, value.to(device))

    def _before_denoising_loop(
        self, ctx: LTX2DenoisingContext, batch: Req, server_args: ServerArgs
    ) -> None:
        """Reset the mirrored audio scheduler before the shared loop begins."""
        if is_ltx2_two_stage_pipeline_name(
            server_args.pipeline_class_name
        ) and ctx.stage in ("stage1", "stage2"):
            pipeline = self.pipeline() if self.pipeline else None
            if pipeline is not None:
                pipeline.switch_lora_phase(ctx.stage, batch=batch)
        super()._before_denoising_loop(ctx, batch, server_args)
        if ctx.audio_scheduler is None:
            raise ValueError("LTX-2 audio scheduler was not prepared.")
        ctx.audio_scheduler.set_begin_index(0)
        if self.sampler_name == "res2s" and ctx.use_native_hq_res2s_sde_noise:
            self._ltx2_init_res2s_noise_generators(ctx)

    def _prepare_step_attn_metadata(
        self,
        ctx: LTX2DenoisingContext,
        batch: Req,
        server_args: ServerArgs,
        step_index: int,
        t_int: int,
        timesteps_cpu: torch.Tensor,
    ):
        """Preserve the legacy LTX-2 attention-metadata contract."""
        # Legacy LTX-2 paths used the plain attention-metadata builder call here.
        return self._build_attn_metadata(step_index, batch, server_args)

    def _run_denoising_step(

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the LTX-2 prepare path that builds ctx.audio_scheduler runs before _before_denoising_loop
2. Check for config flags that disable audio-scheduler init and disable them or route to a non-LTX-2 stage
3. Grep for where audio_scheduler is assigned and confirm it isn't conditionally skipped for your request type

Example fix

// before
ctx.audio_scheduler = build_audio_scheduler(cfg) if enable_audio else None
// after
ctx.audio_scheduler = build_audio_scheduler(cfg)  # required for LTX-2 stage
assert ctx.audio_scheduler is not None

assert ctx.audio_scheduler is not None, "audio scheduler missing; check prepare stage"

When it happens

Trigger: ctx.audio_scheduler is None at loop start — e.g. the prepare stage skipped audio scheduler construction (audio disabled path, exception swallowed, or stage ordering changed).

Common situations: Running an LTX-2 pipeline variant with audio disabled but still entering the joint loop; a refactor reordering prepare vs loop stages; a config flag that skips audio init without skipping the LTX-2 stage.

Related errors

• 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}
• LTX-2 SP time-sharding for packed token latents currently re
• Expected {seq_len=} > 0 for packed token latents.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/bdb27557704caace. Report an issue: GitHub.