sgl-project/sglang · error · ValueError

--enable-linear-replayssm-spec with DSPARK/DFLASH requires a

Error message

--enable-linear-replayssm-spec with DSPARK/DFLASH requires a KDA (kimi_linear) model; got a non-KDA model.

What it means

When building the hybrid request pool, the linear-replay SSM speculative path (--enable-linear-replayssm-spec with DSPARK or DFLASH) is gated on the model actually being a KDA (kimi_linear) model; kimi_linear_config() returning None for any other architecture triggers this error.

Source

Thrown at python/sglang/srt/mem_cache/kv_cache_configurator.py:873

            pre_alloc_size=pre_alloc_size,
        )
        return req_to_token_pool

    def _build_hybrid_req_pool(
        self,
        *,
        max_num_reqs: int,
        extra_max_context_len: int,
    ) -> ReqToTokenPool:
        # DSPARK/DFLASH commit routes through the backend fold (KDA-only); a
        # non-KDA model there would scatter a None intermediate_ssm and crash.
        _algo = (get_spec().speculative_algorithm or "").upper()
        if (
            get_exec().mamba.enable_linear_replayssm_spec
            and _algo in ("DSPARK", "DFLASH")
            and kimi_linear_config(self.model_config) is None
        ):
            raise ValueError(
                "--enable-linear-replayssm-spec with DSPARK/DFLASH requires a KDA "
                "(kimi_linear) model; got a non-KDA model."
            )
        req_to_token_pool = HybridReqToTokenPool(
            size=max_num_reqs,
            mamba_size=get_schedule().max_mamba_cache_size,
            mamba_spec_state_size=max_num_reqs,
            max_context_len=self.model_config.context_len + extra_max_context_len,
            device=self.device,
            enable_memory_saver=get_exec().features.enable_memory_saver,
            cache_params=self.mambaish_config.mamba2_cache_params,
            mamba_layer_ids=(
                [
                    i
                    for i in self.mambaish_config.mamba2_cache_params.layers
                    if self.layer_info.start_layer <= i < self.layer_info.end_layer
                ]
            ),

View on GitHub (pinned to 0132848349)

Solutions

1. Drop --enable-linear-replayssm-spec when running non-KDA models
2. Use a kimi_linear (KDA) checkpoint with this combination
3. Switch the speculative algorithm to one not named DSPARK/DFLASH

Example fix

# before
python -m sglang.launch_server --model qwen3-... --speculative-algorithm DFLASH --enable-linear-replayssm-spec
# after
python -m sglang.launch_server --model qwen3-... --speculative-algorithm DFLASH

is_kda = any("Kimi" in a or "kimi_linear" in a.lower() for a in model_config.hf_config.architectures)
if exec_cfg.mamba.enable_linear_replayssm_spec and (server_args.speculative_algorithm or "").upper() in ("DSPARK", "DFLASH") and not is_kda:
    exec_cfg.mamba.enable_linear_replayssm_spec = False

When it happens

Trigger: Enable the exec flag get_exec().mamba.enable_linear_replayssm_spec together with a DSPARK/DFLASH speculative algorithm on a non-kimi_linear checkpoint.

Common situations: Reusing a KDA-specific spec-decode flag on a different hybrid model, or a mis-set speculative_algorithm string.

Related errors

• expected q/k shape {expected_shape}
• expected v/g shape {expected_shape}
• Kimi-K3 DCP + DSPARK currently requires SGLANG_RAGGED_VERIFY
• DFLASH speculative decoding only supports CUDA and NPU devic
• Currently DFLASH speculative decoding does not support dp at

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a78d8bae92fe4090. Report an issue: GitHub.