sgl-project/sglang · critical · RuntimeError

world_size ({world_size}) is less than tensor_parallel_degre

Error message

world_size ({world_size}) is less than tensor_parallel_degree ({tensor_parallel_degree}) x pipeline_parallel_degree ({pipeline_parallel_degree}) xsequence_parallel_degree ({sequence_parallel_degree}) xclassifier_free_guidance_degree ({classifier_free_guidance_degree}) xdata_parallel_degree ({data_parallel_size})

What it means

Raised by initialize_model_parallel when the total requested parallelism (tensor x pipeline x sequence x classifier-free-guidance x data parallel degrees) exceeds the initialized world_size. The product of all parallel degrees must fit within the number of distributed ranks.

Source

Thrown at python/sglang/multimodal_gen/runtime/distributed/parallel_state.py:443

    if backend is None:
        from sglang.multimodal_gen.runtime.platforms import current_platform

        backend = current_platform.get_torch_distributed_backend_str()
    # Get world size and rank. Ensure some consistencies.
    assert torch.distributed.is_initialized()
    world_size: int = torch.distributed.get_world_size()
    backend = backend or torch.distributed.get_backend(get_world_group().device_group)

    dit_parallel_size = (
        data_parallel_size
        * classifier_free_guidance_degree
        * sequence_parallel_degree
        * pipeline_parallel_degree
        * tensor_parallel_degree
    )

    if world_size < dit_parallel_size:
        raise RuntimeError(
            f"world_size ({world_size}) is less than "
            f"tensor_parallel_degree ({tensor_parallel_degree}) x "
            f"pipeline_parallel_degree ({pipeline_parallel_degree}) x"
            f"sequence_parallel_degree ({sequence_parallel_degree}) x"
            f"classifier_free_guidance_degree "
            f"({classifier_free_guidance_degree}) x"
            f"data_parallel_degree ({data_parallel_size})"
        )

    rank_generator: RankGenerator = RankGenerator(
        tensor_parallel_degree,
        sequence_parallel_degree,
        pipeline_parallel_degree,
        classifier_free_guidance_degree,
        data_parallel_size,
        "tp-sp-pp-cfg-dp",
    )
    global _DP

View on GitHub (pinned to 0132848349)

Solutions

1. Reduce the parallel degrees (tp/pp/sp/cfg) so their product times data_parallel_size is <= world_size
2. Or increase world_size (more ranks/GPUs) to at least the product of the degrees
3. Verify data_parallel_size is not implicitly inflating the requirement; often world_size/tp should equal dp, not require dp extra ranks

Example fix

# before
--tensor-parallel-degree 8 --data-parallel-size 2   # world_size=8

# after
--tensor-parallel-degree 4 --data-parallel-size 2   # 4*2=8 <= 8

def check_degrees(world_size, tp, pp, sp, cfg, dp):
    need = tp * pp * sp * cfg * dp
    assert world_size >= need, f"need {need} ranks, have {world_size}"
    return True

When it happens

Trigger: Calling maybe_init_distributed_environment_and_model_parallel with e.g. tp=8, pp=2 on a world_size of 8, or setting data_parallel_size such that tp*pp*sp*cfg*dp > world_size.

Common situations: Misconfigured CLI/server args (e.g. --tp 8 on 4 GPUs), forgetting that data_parallel_size multiplies the constraint, or changing GPU count without updating parallel degrees in config files.

Related errors

• Invalid {tp_size=}. Expected tp_size >= 1.
• The size of ({name}) is ({self.name_to_size[name]}), but you
• --enable-strict-thinking requires a grammar backend that sup
• Invalid grammar backend: {name}
• launch_local_runtime requires --dp-size 1; got dp_size={get_

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/798e1b5d3c8bc5d4. Report an issue: GitHub.