sgl-project/sglang · error · ValueError

--enable-dsa-cache-layer-split currently only supports the m

Error message

--enable-dsa-cache-layer-split currently only supports the mooncake transfer backend (mooncake / mooncake_tcp). Got --disaggregation-transfer-backend {cfg.disaggregation_transfer_backend!r}. mori/nixl support will be added later by the community.

What it means

DSA cache-layer-split relies on the mooncake all-CP-rank KV/indexer transfer path, so it currently only works with the mooncake transfer backend (mooncake / mooncake_tcp). mori and nixl backends are rejected until support is added.

Source

Thrown at python/sglang/srt/server_args.py:5955

                        "require ordinary local decode cache semantics."
                    )
                if cfg.enable_dsa_cache_layer_split and (
                    not cfg.enable_prefill_cp or cfg.cp_strategy != "interleave"
                ):
                    raise ValueError(
                        "--enable-dsa-cache-layer-split requires "
                        "--enable-prefill-cp and --cp-strategy interleave "
                        "(or legacy --enable-nsa-prefill-context-parallel with "
                        "--nsa-prefill-cp-mode round-robin-split)."
                    )
                # Layer split relies on the mooncake all-CP-rank KV/indexer
                # transfer path. mori/nixl support is a temporary limitation
                # and will be added later by the community.
                if (
                    cfg.enable_dsa_cache_layer_split
                    and cfg.disaggregation_transfer_backend != "mooncake"
                ):
                    raise ValueError(
                        "--enable-dsa-cache-layer-split currently only supports "
                        "the mooncake transfer backend (mooncake / mooncake_tcp). "
                        f"Got --disaggregation-transfer-backend "
                        f"{cfg.disaggregation_transfer_backend!r}. mori/nixl "
                        "support will be added later by the community."
                    )
                if cfg.enable_dsa_cache_layer_split and cfg.pp_size > 1:
                    raise ValueError(
                        "--enable-dsa-cache-layer-split is not supported with "
                        "pipeline parallelism (pp_size > 1) yet. It requires "
                        "prefill context parallelism, and CP + PP has not been "
                        "validated for this feature."
                    )

            else:
                # DeepSeek V3/R1/V3.1
                if cfg.cuda_graph_config.prefill.backend != Backend.DISABLED:
                    logger.info("Piecewise CUDA graph is enabled, use MLA for prefill.")

View on GitHub (pinned to 0132848349)

Solutions

Set --disaggregation-transfer-backend mooncake (or mooncake_tcp)
Remove --enable-dsa-cache-layer-split if you must stay on mori/nixl
Watch for community-added mori/nixl support in later releases

Example fix

# before
--enable-dsa-cache-layer-split --disaggregation-transfer-backend nixl
# after
--enable-dsa-cache-layer-split --disaggregation-transfer-backend mooncake

Defensive patterns

Strategy: validation

Validate before calling

if args.get("enable_dsa_cache_layer_split") and args.get("disaggregation_transfer_backend", "mooncake") != "mooncake":
    args["disaggregation_transfer_backend"] = "mooncake"  # or fail fast

Prevention

Pin the transfer backend explicitly rather than relying on cluster defaults
Document backend limitations per optimization flag in your deployment runbook

When it happens

Trigger: Running a PD prefill worker with --enable-dsa-cache-layer-split and --disaggregation-transfer-backend set to 'mori', 'nixl', or anything other than 'mooncake'.

Common situations: Clusters standardized on nixl/mori transfer backends trying the DSA optimization; forgetting to switch the backend when adopting the flag.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/82f33f900579de6b. Report an issue: GitHub.