sgl-project/sglang · error · ValueError

MiniCPM sparse attention does not support PD disaggregation

Error message

MiniCPM sparse attention does not support PD disaggregation

What it means

MiniCPM sparse attention backends (minicpm_flashattn/minicpm_flashinfer selected explicitly or by default) are incompatible with prefill-decode disaggregation. If a sparse backend is in use and disaggregation_mode != 'null', the override hook raises ValueError.

Source

Thrown at python/sglang/srt/arg_groups/overrides.py:1323

        if dense_attention is not None:
            overrides["attention_backend"] = dense_attention
        dense_prefill = dense_backends.get(cfg.prefill_attention_backend)
        if dense_prefill is not None:
            overrides["prefill_attention_backend"] = dense_prefill
        dense_decode = dense_backends.get(cfg.decode_attention_backend)
        if dense_decode is not None:
            overrides["decode_attention_backend"] = dense_decode
    elif has_sparse_attention:
        uses_sparse_backend = cfg.is_attention_backend_not_set() or any(
            backend in ("minicpm_flashattn", "minicpm_flashinfer")
            for backend in (
                cfg.attention_backend,
                cfg.prefill_attention_backend,
                cfg.decode_attention_backend,
            )
        )
        if uses_sparse_backend and cfg.disaggregation_mode != "null":
            raise ValueError(
                "MiniCPM sparse attention does not support PD disaggregation"
            )
        if cfg.is_attention_backend_not_set():
            overrides["attention_backend"] = (
                "minicpm_flashinfer"
                if is_blackwell_supported()
                else "minicpm_flashattn"
            )
    return overrides


@_register_for("MiniCPMV4_6ForConditionalGeneration")
def _minicpm_v4_6_overrides(server_args: Any, hf_config: Any) -> dict:
    cfg = resolving_view(server_args)
    if is_sm100_supported() and cfg.attention_backend is None:
        return {"attention_backend": "triton"}
    return {}

View on GitHub (pinned to 0132848349)

Solutions

Set --disaggregation-mode null (or omit the flag) for MiniCPM
Or force a dense path: set SGLANG_MINICPM_FORCE_DENSE=1 and a non-sparse backend, only if validated for your checkpoint
Use a model that supports PD disaggregation for disaggregated serving

Example fix

# before
--disaggregation-mode prefill
# after
# (no disaggregation; monolithic server)

Defensive patterns

Strategy: validation

Validate before calling

sparse = {'minicpm_flashattn', 'minicpm_flashinfer'} & {attention_backend, prefill_backend, decode_backend}
if model_arch.startswith('MiniCPM') and sparse:
    server_args.disaggregation_mode = 'null'

Try / catch

except ValueError as e:
    if 'PD disaggregation' in str(e): server_args.disaggregation_mode = 'null'; retry()
    raise

Prevention

Keep MiniCPM out of PD-disaggregated fleets or force dense mode deliberately
Validate disaggregation_mode against model backend matrix at submit time

When it happens

Trigger: Running MiniCPM with --disaggregation-mode prefill or decode while a minicpm_* sparse attention backend is set in attention_backend/prefill/decode fields.

Common situations: Joining a PD-disaggregation cluster with a MiniCPM SALA model; PD scripts that always pass --disaggregation-mode.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2f5be992ccae5b30. Report an issue: GitHub.