sgl-project/sglang · error · RuntimeError

decode_context_parallel_size ({decode_context_parallel_size}

Error message

decode_context_parallel_size ({decode_context_parallel_size}) must be >= 1

What it means

initialize_model_parallel requires decode_context_parallel_size >= 1 (it is a multiplicity factor for splitting decode attention across ranks). Values of 0 or negative are invalid and rejected before any process group is built.

Source

Thrown at python/sglang/srt/distributed/parallel_state.py:2421

    # Get world size and rank. Ensure some consistencies.
    assert torch.distributed.is_initialized()
    backend = backend or torch.distributed.get_backend(get_world_group().device_group)

    # Joiners construct their local TP/PP layout in global rank space.
    world_size: int = (
        tensor_model_parallel_size * pipeline_model_parallel_size
        if recovered_rank
        else torch.distributed.get_world_size()
    )

    if world_size != tensor_model_parallel_size * pipeline_model_parallel_size:
        raise RuntimeError(
            f"world_size ({world_size}) is not equal to "
            f"tensor_model_parallel_size ({tensor_model_parallel_size}) x "
            f"pipeline_model_parallel_size ({pipeline_model_parallel_size})"
        )
    if decode_context_parallel_size < 1:
        raise RuntimeError(
            f"decode_context_parallel_size ({decode_context_parallel_size}) must be >= 1"
        )
    if decode_context_parallel_size > 1 and not (is_hip() or is_cuda()):
        raise RuntimeError(
            "Decode context parallel (decode_context_parallel_size > 1) is "
            "currently only supported on the AMD HIP platform or CUDA platform, but got "
            f"decode_context_parallel_size ({decode_context_parallel_size}) "
            "on a non-HIP or non-CUDA platform."
        )
    if tensor_model_parallel_size % decode_context_parallel_size != 0:
        raise RuntimeError(
            f"tensor_model_parallel_size ({tensor_model_parallel_size}) must be divisible by "
            f"decode_context_parallel_size ({decode_context_parallel_size})"
        )

    # Build the tensor model-parallel groups.
    num_tensor_model_parallel_groups: int = world_size // tensor_model_parallel_size
    global _TP

View on GitHub (pinned to 0132848349)

Solutions

1. Set decode_context_parallel_size to 1 (the no-op default) or a valid divisor of tensor_model_parallel_size
2. Audit how the value is computed in your launcher/config before it reaches initialize_model_parallel
3. Rerun with the corrected flag, e.g. --decode-context-parallel-size 1

Example fix

# before
--decode-context-parallel-size 0
# after
--decode-context-parallel-size 1  # or omit the flag

assert decode_context_parallel_size >= 1, 'decode_context_parallel_size must be >= 1'

When it happens

Trigger: Passing decode_context_parallel_size < 1 to initialize_model_parallel — e.g. via a server arg that computed/decoded to 0 (misconfigured --decode-context-parallel-size or a bad default in a custom launch script).

Common situations: A wrapper script derives the value by division that rounds to 0 (e.g. tp // dp with dp > tp); explicitly setting the flag to 0 thinking it disables DCP; config typos or negative values from templated YAML.

Related errors

• tensor_model_parallel_size ({tensor_model_parallel_size}) mu
• world_size ({world_size}) is not equal to tensor_model_paral
• Decode context parallel (decode_context_parallel_size > 1) i
• num_heads ({self.num_heads}) must be divisible by ulysses_de
• LTX2Attention requires heads divisible by tp_size, got {self

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/445bd9a7f344f088. Report an issue: GitHub.