sgl-project/sglang · critical · RuntimeError

video_vae became unavailable during decode

Error message

video_vae became unavailable during decode

What it means

Inside the use_declared_component('video_vae') context the selected module must not be None; if component resolution yields None mid-decode (component released/unregistered concurrently) this RuntimeError fires.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/decoding.py:376

    def forward(self, batch: Req, server_args: ServerArgs) -> OutputBatch:
        _minimax_h3_decoder_task(batch)
        visual_latent = _required_tensor(batch.latents, "batch.latents")
        audio_latent = _required_tensor(batch.audio_latents, "batch.audio_latents")
        if visual_latent.ndim != 5:
            raise ValueError("batch.latents must be [B, C, T, H, W]")
        if audio_latent.ndim != 3:
            raise ValueError(
                "batch.audio_latents must be [audio_channel, latent_dim, T]"
            )

        if self.video_vae is None:
            raise RuntimeError("MiniMax H3 tasks require the video_vae output decoder")
        with self.use_declared_component(
            component_name="video_vae",
            module=self.video_vae,
        ) as selected_video_vae:
            if selected_video_vae is None:
                raise RuntimeError("video_vae became unavailable during decode")
            self.video_vae = selected_video_vae
            if selected_video_vae.training:
                selected_video_vae.eval()
            visual_arch_config = server_args.pipeline_config.vae_config.arch_config
            visual_decode_latent = _reverse_normalize_latents(
                visual_latent,
                mean_values=visual_arch_config.latents_mean,
                std_values=visual_arch_config.latents_std,
                name="video_vae",
            )
            video_vae_dtype = resolve_decode_precision(server_args, "video_vae")
            visual_autocast_enabled = autocast_enabled_for_device(
                visual_latent, video_vae_dtype, server_args.disable_autocast
            )
            if visual_autocast_enabled:
                selected_video_vae.prepare_decoder_autocast_weights(video_vae_dtype)
            with autocast_context(
                video_vae_dtype,

View on GitHub (pinned to 0132848349)

Solutions

1. Keep the video_vae component registered/alive for the lifetime of the decode call
2. Avoid tearing down or mutating the component registry concurrently with inference
3. File/inspect a race where the component lease is released early

assert stage.video_vae is not None  # and keep module referenced for the call duration

When it happens

Trigger: The declared video_vae component being unset or garbage-collected/released by another thread while decode is in flight.

Common situations: Concurrent pipeline teardown or component hot-swap during generation; a bug where the component registry entry for video_vae was deleted before decoding finished.

Related errors

• MiniMax H3 tasks require the video_vae output decoder
• unsupported input for wan_rmsnorm_silu
• fl2va requires first_frame, last_frame, or both
• ref2va requires at least one of reference_image, reference_v
• t2va takes no conditioning inputs; pick another task

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/4b195404fb2500e0. Report an issue: GitHub.