sgl-project/sglang · critical · RuntimeError

{transport_name} consumer rank {rank} is outside [0, {total_

Error message

{transport_name} consumer rank {rank} is outside [0, {total_consumer_count})

What it means

After resolving the consumer rank (explicitly or from tp_rank), the transport validates 0 <= rank < total_consumer_count. A rank at or beyond the consumer count would write acknowledgements into another transport's slot space, so it is rejected with this RuntimeError.

Source

Thrown at python/sglang/srt/multimodal/transport/memory_pool.py:73

) -> int:
    if total_consumer_count == 1:
        return 0
    if consumer_rank is None:
        try:
            from sglang.srt.runtime_context import get_parallel

            # Use the global TP rank. An attention/DCP subgroup rank can alias
            # another consumer's acknowledgement slot.
            rank = int(get_parallel().tp_rank)
        except Exception as exc:
            raise RuntimeError(
                f"Cannot resolve the {transport_name} consumer rank before "
                "parallel state initialization"
            ) from exc
    else:
        rank = int(consumer_rank)
    if not 0 <= rank < total_consumer_count:
        raise RuntimeError(
            f"{transport_name} consumer rank {rank} is outside "
            f"[0, {total_consumer_count})"
        )
    return rank


class StreamOrderedPoolConsumerMixin:
    """Ready/wait/ack protocol for stream-ordered GPU feature proxies."""

    def _init_stream_ordered_consumer(
        self,
        *,
        ready_byte_offset: int,
        ack_byte_offset: int,
        generation: int,
        total_consumer_count: int,
        transport_name: str,
    ) -> None:

View on GitHub (pinned to 0132848349)

Solutions

1. Verify consumer_rank < number of consumers; use 0-based ranks
2. Check total_consumer_count matches the actual tokenizer/TP worker count at pool construction
3. Use the rank returned by the runtime rather than a hardcoded value

Example fix

# before
ack(consumer_rank=8)  # total_consumer_count=4
# after
ack(consumer_rank=rank % 4)

assert 0 <= consumer_rank < total_consumer_count, 'rank out of range'

When it happens

Trigger: Passing consumer_rank equal to or above the number of tokenizer workers / consumers, or a TP-rank/DP-size misconfiguration making global tp_rank exceed the transport's consumer count.

Common situations: Mixing up global TP rank vs attention subgroup rank in mixedparallel setups; hardcoding a rank in tests larger than the consumer count; wrong total_consumer_count passed at pool construction.

Related errors

• total_pool_size must be positive
• tokenizer_worker_num must be positive
• Input 'data' must be a torch.Tensor, but got {type}
• Cannot resolve the {transport_name} consumer rank before par
• total_consumer_count must be positive

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c0091cb8d2d1ae55. Report an issue: GitHub.