sgl-project/sglang · error · ValueError

All frames in a batch must have the same resolution

Error message

All frames in a batch must have the same resolution

What it means

Raised by RealESRGANUpscaler.upscale_batch when frames passed in one batch do not all share the same (H, W), since the implementation stacks them into a single numpy array for one batched forward pass.

Source

Thrown at python/sglang/multimodal_gen/runtime/postprocess/realesrgan_upscaler.py:456

            target_h = int(h * outscale)
            target_w = int(w * outscale)
            out = F.interpolate(
                out, size=(target_h, target_w), mode="bicubic", align_corners=False
            )

        out_np = out.squeeze(0).permute(1, 2, 0).clamp(0.0, 1.0).cpu().numpy()
        return (out_np * 255.0).astype(np.uint8)

    def upscale_batch(
        self, frames: list[np.ndarray], outscale: float | None = None
    ) -> list[np.ndarray]:
        """Upscale same-resolution HWC uint8 frames in one batched forward pass."""
        if not frames:
            return []

        h, w = frames[0].shape[:2]
        if any(frame.shape[:2] != (h, w) for frame in frames):
            raise ValueError("All frames in a batch must have the same resolution")

        total_start_time = time.perf_counter()

        start_time = time.perf_counter()
        imgs = np.stack(frames, axis=0)
        stack_duration_s = time.perf_counter() - start_time

        start_time = time.perf_counter()
        h2d_timer = self._start_cuda_timer()
        imgs_t = self._copy_input_to_device(imgs)
        self._stop_cuda_timer(h2d_timer)
        h2d_wall_duration_s = time.perf_counter() - start_time

        start_time = time.perf_counter()
        input_preprocess_timer = self._start_cuda_timer()
        imgs_t = self._preprocess_input_tensor(imgs_t)
        self._stop_cuda_timer(input_preprocess_timer)
        input_preprocess_wall_duration_s = time.perf_counter() - start_time

View on GitHub (pinned to 0132848349)

Solutions

1. Resize/pad all frames to a common resolution before batching
2. Group frames by resolution and call upscale_batch once per group

Example fix

# before
outs = upscaler.upscale_batch([f1080, f720])
# after
from torchvision.transforms import functional as F
frames = [cv2.resize(f, (W, H)) for f in frames]
outs = upscaler.upscale_batch(frames)

from itertools import groupby
frames.sort(key=lambda f: f.shape[:2])  # or group explicitly
shapes = {f.shape[:2] for f in frames}
assert len(shapes) <= 1 or batched_per_group, "mixed resolutions"

def frames_uniform(frames: list) -> bool:
    return all(f.shape[:2] == frames[0].shape[:2] for f in frames)

When it happens

Trigger: Calling upscale_batch (or upscale_batched which groups frames) with a list containing frames of differing resolutions, e.g. mixing 1080p and 720p frames or portrait/landscape variants.

Common situations: Feeding raw video frames without normalization; mixing images from mixed sources; upstream crop/reszie step skipped.

Related errors

• RealESRGAN batch upscale did not produce all frames
• Unsupported activation type: {act_type}
• Unsupported RRDBNet conv_first input channels: {in_channels}
• Failed to load Real-ESRGAN checkpoint from '{resolved_path}'
• Real-ESRGAN weight file '{resolved_path}' is not compatible

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1afdb5e8177ab403. Report an issue: GitHub.