sgl-project/sglang · error · ValueError

Kimi-K3 DCP + DSPARK currently requires SGLANG_RAGGED_VERIFY

Error message

Kimi-K3 DCP + DSPARK currently requires SGLANG_RAGGED_VERIFY_MODE=static; compact/cap-accept are not validated under DCP (got {ragged_mode.value!r}).

What it means

Kimi-K3 DCP combined with DSPARK speculative decoding has only been validated with SGLANG_RAGGED_VERIFY_MODE=static. The compact and cap-accept ragged verify modes are untested/incorrect under DCP, so any other value read from the env is rejected.

Source

Thrown at python/sglang/srt/arg_groups/overrides.py:633

    cfg = resolving_view(server_args)
    if cfg.dcp_size > 1:
        overrides = {}
        if cfg.enable_symm_mem:
            logger.warning(
                "Kimi-K3 DCP disables --enable-symm-mem due to decode CUDA "
                "graph correctness issues."
            )
            overrides["enable_symm_mem"] = False

        if cfg.speculative_algorithm == "DSPARK":
            from sglang.srt.speculative.ragged_verify import (
                RaggedVerifyMode,
                read_ragged_verify_mode,
            )

            ragged_mode = read_ragged_verify_mode()
            if ragged_mode is not RaggedVerifyMode.STATIC:
                raise ValueError(
                    "Kimi-K3 DCP + DSPARK currently requires "
                    "SGLANG_RAGGED_VERIFY_MODE=static; compact/cap-accept are "
                    f"not validated under DCP (got {ragged_mode.value!r})."
                )

            # DSPARK target-verify + draft-extend must run on the decode
            # (cutedsl_mla) backend, whose _run_decode_kernel implements the DCP
            # signature (causal_seqs / cp_world / cp_rank). The default
            # "prefill" routes verify to trtllm_mla, whose base _run_decode_kernel
            # lacks that DCP path (TypeError: unexpected kwarg 'causal_seqs').
            overrides["speculative_attention_mode"] = "decode"

        prefill_backend, decode_backend = attention_backends_of(cfg)
        if decode_backend == "cutedsl_mla" or decode_backend is None:
            _require_kimi_k3_cutedsl_dcp_support()
            logger.info(
                "Kimi-K3 DCP keeps decode attention backend 'cutedsl_mla' "
                f"(prefill={prefill_backend!r} -> 'trtllm_mla')."

View on GitHub (pinned to 0132848349)

Solutions

1. Set SGLANG_RAGGED_VERIFY_MODE=static (or unset it if static is the default)
2. Remove compact/cap_accept from launch scripts/Dockerfiles for Kimi-K3 DCP runs
3. If you need those modes, disable DCP for this model

Example fix

# before
export SGLANG_RAGGED_VERIFY_MODE=compact
# after
export SGLANG_RAGGED_VERIFY_MODE=static

import os
mode = os.environ.get('SGLANG_RAGGED_VERIFY_MODE', 'static')
assert mode == 'static', f'unset/override SGLANG_RAGGED_VERIFY_MODE for Kimi-K3 DCP (got {mode})'

except ValueError as e:
    if 'RAGGED_VERIFY_MODE' in str(e):
        os.environ['SGLANG_RAGGED_VERIFY_MODE'] = 'static'; retry()
    raise

When it happens

Trigger: Env SGLANG_RAGGED_VERIFY_MODE set to compact or cap_accept while running Kimi-K3 with DCP and DSPARK spec decoding; read_ragged_verify_mode() returns a non-STATIC mode in _kimi_k3_overrides.

Common situations: Tuning ragged verify mode for throughput on non-DCP setups and leaving the env exported in shell profiles or Docker ENV; shared cluster images with the var pre-set.

Related errors

• DSpark with dp attention + moe_a2a_backend={} requires SGLAN
• DSV4 ragged verify does not support context parallel (CP); s
• DSpark speculative_num_draft_tokens must be >= 2 (= gamma +
• Invalid ltx2_two_stage_device_mode={mode!r}. Expected one of
• --enable-hisparse is not supported with the unified-KV path

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d7326ee57f8a95e6. Report an issue: GitHub.