sgl-project/sglang · error · KeyError

f"seq_len {item} not supported for STA"

Error message

f"seq_len {item} not supported for STA"

What it means

A dict subclass mapping supported sequence lengths (exact keys or (low, high) range tuples) to STA tile configs raises KeyError from __getitem__ when the requested length matches no entry. Only lengths with a precomputed sliding-tile mask strategy are supported.

Source

Thrown at python/sglang/multimodal_gen/runtime/layers/attention/backends/sliding_tile_attn.py:47

    st_attn_backend_available = True
except Exception:
    st_attn_backend_available = False

logger = init_logger(__name__)


class RangeDict(dict):

    def __getitem__(self, item: int) -> str:
        for key in self.keys():
            if isinstance(key, tuple):
                low, high = key
                if low <= item <= high:
                    return str(super().__getitem__(key))
            elif key == item:
                return str(super().__getitem__(key))
        raise KeyError(f"seq_len {item} not supported for STA")


class SlidingTileAttentionBackend(AttentionBackend):
    accept_output_buffer: bool = True

    @staticmethod
    def get_supported_head_sizes() -> list[int]:
        # TODO(will-refactor): check this
        return [32, 64, 96, 128, 160, 192, 224, 256]

    @staticmethod
    def get_enum() -> AttentionBackendEnum:
        return AttentionBackendEnum.SLIDING_TILE_ATTN

    @staticmethod
    def get_impl_cls() -> type["SlidingTileAttentionImpl"]:
        return SlidingTileAttentionImpl

View on GitHub (pinned to 0132848349)

Solutions

1. Use a sequence length/resolution listed in the mapping keys (print them to see supported ranges)
2. Pad/round the sequence length up to the nearest supported value
3. Fall back to a dense/other attention backend for unsupported lengths

Example fix

# before
cfg = sta_config[seq_len]  # KeyError: seq_len 8123 not supported for STA
# after
if not any((isinstance(k, tuple) and k[0] <= seq_len <= k[1]) or k == seq_len for k in sta_config):
    seq_len = nearest_supported_length(seq_len)
cfg = sta_config[seq_len]

if not any((isinstance(k, tuple) and k[0] <= seq_len <= k[1]) or k == seq_len for k in sta_cfg):
    raise SystemExit(f"unsupported STA seq_len {seq_len}; supported: {list(sta_cfg.keys())}")

def sta_len_supported(cfg, n: int) -> bool:
    return any((isinstance(k, tuple) and k[0] <= n <= k[1]) or k == n for k in cfg.keys())

try:
    tile_cfg = sta_cfg[seq_len]
except KeyError:
    tile_cfg = sta_cfg[nearest_supported_length(seq_len)]

When it happens

Trigger: Indexing the STA config mapping with a seq_len outside every registered (low, high) range and equal to no exact key.

Common situations: Generating video/images at a resolution whose token count isn't covered by the shipped mask configs; multi-resolution batches mixing supported and unsupported lengths.

Related errors

• st attn not supported
• SGLANG_DIFFUSION_ATTENTION_CONFIG is not set
• mask_strategy cannot be None for SlidingTileAttention
• mask_strategy[0] cannot be None for SlidingTileAttention
• forward_batch cannot be None

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5fe6b615ebe7b931. Report an issue: GitHub.