sgl-project/sglang · critical · RuntimeError

The size of ({name}) is ({self.name_to_size[name]}), but you

Error message

The size of ({name}) is ({self.name_to_size[name]}), but you haven't specified the order ({self.order}).

What it means

A distributed initialization (TP/DP/EP-style group) validates that every parallel dimension with size > 1 appears in the user-supplied `order` string. If, e.g. dp=2 but 'dp' is not in order='tp-ep', it raises this RuntimeError because the rank-mesh layout would be ambiguous.

Source

Thrown at python/sglang/multimodal_gen/runtime/utils/distributed.py:195

        self.sp = sp
        self.pp = pp
        self.cfg = cfg
        self.dp = dp
        self.rank_offset = rank_offset
        self.world_size = tp * sp * pp * cfg * dp

        self.name_to_size = {
            "tp": self.tp,
            "sp": self.sp,
            "pp": self.pp,
            "cfg": self.cfg,
            "dp": self.dp,
        }
        order = order.lower()

        for name in self.name_to_size.keys():
            if name not in order and self.name_to_size[name] != 1:
                raise RuntimeError(
                    f"The size of ({name}) is ({self.name_to_size[name]}), but you haven't specified the order ({self.order})."
                )
            elif name not in order:
                order = order + "-" + name

        self.order = order
        self.ordered_size = []

        for token in order.split("-"):
            self.ordered_size.append(self.name_to_size[token])

    def get_mask(self, order: str, token: str):
        ordered_token = order.split("-")
        token = token.split("-")
        mask = [False] * len(ordered_token)
        for t in token:
            mask[ordered_token.index(t)] = True
        return mask

View on GitHub (pinned to 0132848349)

Solutions

1. Add the missing dimension to the order string, e.g. order='tp-dp' or 'tp-ep-dp', matching every dimension with size > 1
2. Or reduce the missing dimension's size back to 1
3. Compute the order programmatically from the active parallel sizes instead of hardcoding

Example fix

# before
order="tp"
# dp=2 -> error
# after
order="tp-dp"

missing = [k for k, v in name_to_size.items() if v > 1 and k not in order.lower()]
assert not missing, f'add {missing} to order'

When it happens

Trigger: Constructing the distributed helper with name_to_size containing a dimension >1 (e.g. dp=2) while the `order` argument omits that dimension name. Size-1 dimensions are tolerated and get appended implicitly.

Common situations: Enabling data parallel or expert parallel via server args without updating a hardcoded parallel mesh order string; changing --dp-size from 1 to >1 in configs that only list 'tp'.

Related errors

• world_size ({world_size}) is less than tensor_parallel_degre
• Invalid {tp_size=}. Expected tp_size >= 1.
• kv_gather_degree does not compose with ulysses_degree or rin
• launch_local_runtime requires --dp-size 1; got dp_size={get_
• Encoder DP mode requires --dp-size > 1 and --tp-size 1; got

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/ee5d4174b13f93df. Report an issue: GitHub.