sgl-project/sglang · error · ValueError

--enable-linear-replayssm is not supported under PD disaggre

Error message

--enable-linear-replayssm is not supported under PD disaggregation yet (follow-up). Got --disaggregation-mode={cfg.disaggregation_mode!r}.

What it means

ReplaySSM is not yet supported under prefill/decode (PD) disaggregation: the disaggregated decode pool (HybridMambaDecodeReqToTokenPool) is not wired for the ReplaySSM ring and the disagg cache/coordination flow is unvalidated, so enabling it would silently no-op or misbehave. The check rejects any --disaggregation-mode other than null.

Source

Thrown at python/sglang/srt/server_args.py:6956

                    f"--linear-attn-decode-backend={decode!r}."
                )
            from sglang.srt.arg_groups.overrides import (
                mamba_extra_buffer_of,
            )

            if mamba_extra_buffer_of(resolved_view(self)):
                raise ValueError(
                    "--enable-linear-replayssm requires --mamba-radix-cache-strategy "
                    "no_buffer (the default); the extra_buffer ping-pong "
                    "donation path is not yet supported (follow-up). Got "
                    f"--mamba-radix-cache-strategy={cfg.mamba_radix_cache_strategy!r}."
                )
            if cfg.disaggregation_mode != "null":
                # The disaggregated decode pool (HybridMambaDecodeReqToTokenPool)
                # is not wired for the ReplaySSM ring, so the flag would silently
                # no-op there; disagg also runs a different cache/coordination
                # flow that is not yet validated for ReplaySSM (follow-up).
                raise ValueError(
                    "--enable-linear-replayssm is not supported under PD "
                    "disaggregation yet (follow-up). Got "
                    f"--disaggregation-mode={cfg.disaggregation_mode!r}."
                )
            if cfg.linear_replayssm_cache_len < 1:
                raise ValueError(
                    "--linear-replayssm-cache-len must be >= 1, got "
                    f"{cfg.linear_replayssm_cache_len}."
                )

        # ReplaySSM spec-verify (Part B of #28511): linear-chain target verify via
        # fold-every-commit -- the verify stores each draft step's raw inputs into
        # the per-slot (rawv, rawk, g, beta) window and the commit replays the
        # accepted prefix into the fp32 checkpoint. The intra-window interaction
        # uses a strictly-lower causal mask, so it is valid ONLY for a linear
        # draft chain (speculative_eagle_topk in {None, 1}, i.e. NEXTN / MTP);
        # EAGLE tree verify (topk > 1) must fall back to the recurrent verify.
        # GDN sizes the window to the draft maximum; KDA (kda_backend) keeps a

View on GitHub (pinned to 0132848349)

Solutions

  1. Run ReplaySSM without PD disaggregation: remove --disaggregation-mode (or set null)
  2. Or remove --enable-linear-replayssm on disagg deployments until support lands (follow-up)

Example fix

# before
--enable-linear-replayssm --disaggregation-mode decode
# after
--enable-linear-replayssm
Defensive patterns

Strategy: validation

Validate before calling

def validate(disagg_mode, enable_replayssm):
    return not enable_replayssm or disagg_mode == "null"

Type guard

def is_standalone(mode: str) -> bool: return mode == "null"

Prevention

When it happens

Trigger: --enable-linear-replayssm together with --disaggregation-mode prefill or decode.

Common situations: Enabling ReplaySSM on an existing PD-disaggregated deployment; adding disagg flags to a single-node ReplaySSM benchmark config.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/562d787d51cd3f20. Report an issue: GitHub.