sgl-project/sglang · error · ValueError

--enable-dsa-cache-layer-split is not supported on decode wo

Error message

--enable-dsa-cache-layer-split is not supported on decode workers. This flag is a prefill-CP optimization; decode receives full cache shards through PD transfer.

What it means

--enable-dsa-cache-layer-split is a prefill context-parallel optimization for PD (prefill/decode disaggregated) deployments; decode workers receive full cache shards via PD transfer, so the flag is rejected on decode workers.

Source

Thrown at python/sglang/srt/server_args.py:5928

                    import torch

                    major, _ = torch.cuda.get_device_capability()
                    self._set_default_dsa_kv_cache_dtype(
                        major, resolved_view(self).quantization
                    )
                    self._set_default_dsa_backends(major)

                if cfg.enable_prefill_cp:
                    assert (
                        cfg.disaggregation_mode != "decode"
                    ), "CP is only supported for prefill when PD disaggregation, please remove --enable-prefill-cp."
                if (
                    cfg.enable_dsa_cache_layer_split
                    and cfg.disaggregation_mode != "prefill"
                ):
                    if cfg.disaggregation_mode == "decode":
                        raise ValueError(
                            "--enable-dsa-cache-layer-split is not supported on "
                            "decode workers. This flag is a prefill-CP "
                            "optimization; decode receives full cache shards "
                            "through PD transfer."
                        )
                    raise ValueError(
                        "--enable-dsa-cache-layer-split is only supported on PD "
                        "prefill workers. Non-PD workers also run decode and "
                        "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)."

View on GitHub (pinned to 0132848349)

Solutions

Remove --enable-dsa-cache-layer-split from decode worker launch args
Keep the flag only on PD prefill workers
Split your prefill/decode launch configs so flags are role-specific

Example fix

# before (decode worker)
--disaggregation-mode decode --enable-dsa-cache-layer-split
# after
--disaggregation-mode decode

Defensive patterns

Strategy: validation

Validate before calling

mode = args.get("disaggregation_mode")
if args.get("enable_dsa_cache_layer_split") and mode == "decode":
    args["enable_dsa_cache_layer_split"] = False  # decode workers must not set it

Prevention

Render launch args per PD role (prefill vs decode) from separate templates
Never apply prefill-only optimization flags cluster-wide

When it happens

Trigger: Launching with --disaggregation-mode decode together with --enable-dsa-cache-layer-split.

Common situations: Reusing the same flag set for prefill and decode pods in a PD cluster; deployment templates that inject DSA flags cluster-wide.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/afbb410d01d64a5e. Report an issue: GitHub.