sgl-project/sglang · critical · RuntimeError

Cannot resolve the {transport_name} consumer rank before par

Error message

Cannot resolve the {transport_name} consumer rank before parallel state initialization

What it means

Stream-ordered transport acknowledgement needs a consumer rank. If consumer_rank is not supplied, the code queries get_parallel().tp_rank; when the distributed/parallel runtime is not initialized, that lookup throws and is re-raised as this RuntimeError. It signals the transport was used outside a properly initialized sglang runtime.

Source

Thrown at python/sglang/srt/multimodal/transport/memory_pool.py:66

    )


def resolve_consumer_rank(
    total_consumer_count: int,
    consumer_rank: Optional[int] = None,
    transport_name: str = "GPU",
) -> 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,

View on GitHub (pinned to 0132848349)

Solutions

1. Pass consumer_rank explicitly when running outside the sglang runtime
2. Initialize sglang's parallel state (init_distributed / RuntimeContext) before acknowledging
3. Restructure tests to init/destroy parallel state around the transport calls

Example fix

# before
pool.acknowledge_consumption()  # no parallel state
# after
pool.acknowledge_consumption(consumer_rank=0)

try:
    from sglang.srt.runtime_context import get_parallel
    rank = get_parallel().tp_rank
except Exception:
    rank = None  # pass explicitly instead

try:
    pool.acknowledge_consumption()
except RuntimeError as e:
    if 'parallel state initialization' in str(e):
        pool.acknowledge_consumption(consumer_rank=local_rank)

When it happens

Trigger: Calling acknowledge_consumption / _acknowledge_on_stream without an explicit consumer_rank in a standalone script or unit test where torch.distributed and sglang's parallel state were never initialized.

Common situations: Running transport unit tests outside the scheduler; constructing the pool in a notebook or tooling process; init order bugs where the transport outlives or precedes runtime teardown.

Related errors

• {transport_name} consumer rank {rank} is outside [0, {total_
• --mm-feature-transport=cuda_ipc only supports a single node.
• world_size must be positive and divide global_heads
• Group {group_name} is destroyed.
• world_size ({world_size}) is less than tensor_parallel_degre

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/969e19f40c367a44. Report an issue: GitHub.