sgl-project/sglang · error · ValueError

cutedsl_mla backend can only be used with MLA models.

Error message

cutedsl_mla backend can only be used with MLA models.

What it means

The cutedsl_mla attention backend factory only supports MLA models; runner.use_mla_backend must be true. Non-MLA (GQA/MHA) models use different KV layouts and kernels that CuteDslMLABackend cannot execute.

Source

Thrown at python/sglang/srt/layers/attention/attention_registry.py:104

    return TRTLLMMLABackend(runner)


@register_attention_backend("tokenspeed_mla")
def create_tokenspeed_mla_backend(runner):
    if not runner.use_mla_backend:
        raise ValueError("tokenspeed_mla backend can only be used with MLA models.")
    from sglang.srt.layers.attention.tokenspeed_mla_backend import (
        TokenspeedMLABackend,
    )

    return TokenspeedMLABackend(runner)


@register_attention_backend("cutedsl_mla")
def create_cutedsl_mla_backend(runner):
    if not runner.use_mla_backend:
        raise ValueError("cutedsl_mla backend can only be used with MLA models.")
    from sglang.srt.layers.attention.cutedsl_mla_backend import CuteDslMLABackend

    return CuteDslMLABackend(runner)


@register_attention_backend("aiter")
def create_aiter_backend(runner):
    from sglang.srt.layers.attention.aiter_backend import AiterAttnBackend

    return AiterAttnBackend(runner)


@register_attention_backend("wave")
def create_wave_backend(runner):
    from sglang.srt.layers.attention.wave_backend import WaveAttnBackend

    return WaveAttnBackend(runner)

View on GitHub (pinned to 0132848349)

Solutions

1. Drop the explicit backend flag and use auto-selection
2. Use an MLA-based checkpoint (DeepSeek family) with cutedsl_mla

Example fix

# before
--attention-backend cutedsl_mla --model llama-3
# after
--model llama-3  # auto backend selection

if not model_runner.use_mla_backend and server_args.attention_backend == "cutedsl_mla":
    raise SystemExit("cutedsl_mla requires an MLA model")

def is_mla_model(runner) -> bool:
    return bool(getattr(runner, "use_mla_backend", False))

When it happens

Trigger: Launching a server with --attention-backend cutedsl_mla (or resolving that backend string) for a non-MLA model architecture, where use_mla_backend is False.

Common situations: Forcing the CuTe DSL MLA backend on Llama/Qwen-class models, or a stale launch script from a DeepSeek deployment reused with a different checkpoint.

Related errors

• tokenspeed_mla backend can only be used with MLA models.
• trtllm_mha backend can only be used with non-MLA models.
• hpc_ops backend can only be used with non-MLA models.
• trtllm_mla cannot serve decode context parallelism with spec
• trtllm_mla does not forward the cyclic DCP metadata to its d

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/954e76c91b6be4af. Report an issue: GitHub.