sgl-project/sglang · error · ValueError

{} has {} entries, but rank_index={}

Error message

{} has {} entries, but rank_index={}

What it means

A per-rank list config has fewer entries than the current rank_index, so no value can be selected for this rank.

Source

Thrown at python/sglang/srt/mem_cache/storage/umbp/umbp_store.py:121

def _select_rank_config_value(
    value: Any,
    rank_index: int,
    field_name: str,
    cast_type,
    auto_increment_scalar: bool = False,
):
    if value is None:
        raise ValueError(f"{field_name} must not be None")

    candidates = value
    if isinstance(value, str) and "," in value:
        candidates = [item.strip() for item in value.split(",") if item.strip()]

    if isinstance(candidates, (list, tuple)):
        if not candidates:
            raise ValueError(f"{field_name} must not be empty")
        if rank_index >= len(candidates):
            raise ValueError(
                f"{field_name} has {len(candidates)} entries, but rank_index={rank_index}"
            )
        return cast_type(candidates[rank_index])

    selected = cast_type(candidates)
    if auto_increment_scalar:
        selected = cast_type(selected + rank_index)
    return selected


# extra_config is an explicit allow-list grouped by the scope in which each key
# takes effect (distributed mode is enabled by master_address). Advanced SPDK
# knobs outside this list go through the "spdk_passthrough" escape hatch.
_COMMON_EXTRA_KEYS = frozenset(
    {
        "dram_capacity_bytes",
        "ssd_enabled",
        "ssd_storage_dir",

View on GitHub (pinned to 0132848349)

Solutions

1. Extend the list to at least rank_index+1 entries (ideally one per rank)
2. Or use a single scalar, which with auto_increment_scalar is broadcast/incremented per rank

Example fix

# before
extra['ssd_paths'] = '/mnt/nvme0,/mnt/nvme1'  # tp=4
# after
extra['ssd_paths'] = '/mnt/nvme0,/mnt/nvme1,/mnt/nvme2,/mnt/nvme3'

vals = rank_list(extra['ssd_paths'])  # as in 4947
assert tp_rank < len(vals) or len(vals) == 1, f'need >= {tp_rank+1} entries for tp={tp_size}'

When it happens

Trigger: Passing e.g. numa list with 2 entries while running TP=4 (rank_index 2 or 3), or a comma string '0,1' with rank_index >= 2.

Common situations: Scaling tensor parallelism up without extending per-rank storage config lists (NUMA nodes, SSD paths).

Understand the failure class

Background: Config validation failed: what "invalid value for {key}" and settings-rejection errors mean across 19 open-source libraries — this error's family across 19 libraries.

Related errors

• extra_config[{!r}] must be a boolean-like value (true/false,
• {} must not be None
• {} must not be empty
• extra_config['ssd_io_backend'] must be one of: posix, io_uri
• extra_config['ssd_durability_mode'] must be one of: strict,

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/dcc84438d3aa9d40. Report an issue: GitHub.