sgl-project/sglang · error · ValueError

latent_window_size must be positive, got {latent_window_size

Error message

latent_window_size must be positive, got {latent_window_size}

What it means

latent_window_size_to_pixel_window_size converts a latent-space window to pixel space by multiplying by downsample_factor; it validates both are positive integers first. A non-positive latent window would produce a zero/negative pixel window, which is meaningless for memory slotting, so it fails fast.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/joy_echo/memory.py:43

    LTX2AVDecodingStage,
)
from sglang.multimodal_gen.runtime.server_args import ServerArgs
from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger

logger = init_logger(__name__)


# --- Audio peak-window helpers ---


def latent_window_size_to_pixel_window_size(
    latent_window_size: int,
    *,
    downsample_factor: int,
    is_causal: bool = True,
) -> int:
    if latent_window_size <= 0:
        raise ValueError(
            f"latent_window_size must be positive, got {latent_window_size}"
        )
    if downsample_factor <= 0:
        raise ValueError(f"downsample_factor must be positive, got {downsample_factor}")

    pixel_window_size = int(latent_window_size) * int(downsample_factor)
    if is_causal:
        pixel_window_size = max(pixel_window_size - (int(downsample_factor) - 1), 1)
    return pixel_window_size


def select_max_response_audio_window_with_bounds(
    segment: Tensor,
    window_size: int,
) -> tuple[Tensor, Tensor, Tensor]:
    if segment.dim() != 4:
        raise ValueError(
            f"Expected segment shape [B, C, T, F], got {tuple(segment.shape)}"

View on GitHub (pinned to 0132848349)

Solutions

1. Fix the caller to compute a positive latent window (clamp to at least 1)
2. Correct the config value to a positive integer
3. Sanity-assert window sizes before saving memory slots

Example fix

# before
lws = max(0, num_latent_frames - overlap)
slot = save_memory_slot(latent_window_size_to_pixel_window_size(lws, downsample_factor=8))
# after
lws = max(1, num_latent_frames - overlap)
slot = save_memory_slot(latent_window_size_to_pixel_window_size(lws, downsample_factor=8))

if latent_window_size <= 0:
    latent_window_size = 1  # or reject with a clear message before calling

When it happens

Trigger: Calling the helper with latent_window_size <= 0 (0 or negative), commonly from save_memory_slot with a computed window of 0 — e.g. an empty clip or miscomputed frame chunk.

Common situations: Computing the window as ceil(frames/threshold) which rounds to 0 for short clips; config files with latent_window_size: 0; off-by-one or inversion when deriving the window from pixel sizes.

Related errors

• SP DMD renoise requires packed video `batch.raw_latent_shape
• SP DMD renoise requires `batch.sp_audio_orig_num_frames`.
• v_cache must be provided
• q can only be None when only_qv=True
• q must be provided unless qv is provided with only_qv=True

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2559a84332313635. Report an issue: GitHub.