sgl-project/sglang · error · ValueError

Explicit vision TP cannot be combined with data parallel

Error message

Explicit vision TP cannot be combined with data parallel

What it means

Qwen3-VL vision modules can either run data-parallel (vision TP resolved to 1,0) or with an explicit TP size/rank — not both. Passing tp_size/tp_rank while use_data_parallel=True is contradictory and rejected.

Source

Thrown at python/sglang/srt/models/qwen3_vl.py:113

logger = logging.getLogger(__name__)

_is_cpu_amx_available = cpu_has_amx_support()
_is_cpu = is_cpu()

# Below this image count the per-image loop beats the vectorized path (which has a
# fixed setup cost; measured crossover ~6 on H20); both give the same result.
_VECTORIZED_VL_POS_EMBED_MIN_IMAGES = 6


def _resolve_vision_tp(
    *,
    use_data_parallel: bool,
    tp_size: Optional[int],
    tp_rank: Optional[int],
) -> tuple[int, int]:
    if use_data_parallel:
        if tp_size is not None or tp_rank is not None:
            raise ValueError("Explicit vision TP cannot be combined with data parallel")
        return 1, 0
    if (tp_size is None) != (tp_rank is None):
        raise ValueError("Vision tp_size and tp_rank must be set together")
    if tp_size is None:
        parallel = get_parallel()
        return parallel.attn_tp_size, parallel.attn_tp_rank
    assert tp_rank is not None
    return tp_size, tp_rank


class Qwen3_VisionMLP(nn.Module):

    def __init__(
        self,
        in_features: int,
        hidden_features: int,
        bias: bool = True,
        hidden_act="silu",

View on GitHub (pinned to 0132848349)

Solutions

Remove the explicit vision tp_size/tp_rank arguments when using data parallel
Or drop --dp-size / disable vision data parallel if you need explicit vision TP

Example fix

# before (dp + explicit vision tp)
Qwen3VLModel(..., use_data_parallel=True, tp_size=2, tp_rank=0)
# after
Qwen3VLModel(..., use_data_parallel=True)

Defensive patterns

Strategy: validation

Validate before calling

if use_data_parallel:
    assert tp_size is None and tp_rank is None, "drop vision TP args under DP"

Prevention

Choose either DP or explicit vision TP, never both
Document which parallelism mode your launcher sets

When it happens

Trigger: Constructing the Qwen3-VL visual tower with use_data_parallel=True and a non-None tp_size or tp_rank (e.g. via --dp-size with explicit vision TP args).

Common situations: Serving Qwen3-VL with --dp-size N while also passing vision TP overrides in server args.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/835314b5cec7cfe6. Report an issue: GitHub.