vllm-project/vllm · error · ValueError

MLA DSpark does not currently support decode context paralle

Error message

MLA DSpark does not currently support decode context parallelism; set decode_context_parallel_size=1.

What it means

Raised when the speculative method is 'dspark', the draft model architecture is K3DSparkModel (an MLA-style DSpark draft), and the target parallel config has decode_context_parallel_size > 1. The MLA DSpark drafting path does not implement decode context parallelism, so vLLM refuses the combination at config validation rather than producing incorrect results.

Source

Thrown at vllm/config/speculative.py:1054

                        getattr(hf, "dspark_target_layer_ids", None) is None
                        and getattr(hf, "target_layer_ids", None) is not None
                    ):
                        hf.dspark_target_layer_ids = hf.target_layer_ids
                    if (
                        getattr(hf, "n_predict", None) is None
                        and getattr(hf, "block_size", None) is not None
                    ):
                        hf.n_predict = hf.block_size

                if self.method in ("dflash", "dspark"):
                    self.parallel_drafting = True

                if (
                    self.method == "dspark"
                    and "K3DSparkModel" in self.draft_model_config.architectures
                    and self.target_parallel_config.decode_context_parallel_size > 1
                ):
                    raise ValueError(
                        "MLA DSpark does not currently support decode context "
                        "parallelism; set decode_context_parallel_size=1."
                    )

                if self.num_speculative_tokens is not None and hasattr(
                    self.draft_model_config.hf_config, "num_lookahead_tokens"
                ):
                    self.draft_model_config.hf_config.num_lookahead_tokens = (
                        self.num_speculative_tokens
                    )

                n_predict = getattr(
                    self.draft_model_config.hf_config, "n_predict", None
                )
                if n_predict is not None:
                    if self.num_speculative_tokens is None:
                        # Default to max value defined in draft model config.
                        self.num_speculative_tokens = n_predict

View on GitHub (pinned to c794754062)

Solutions

1. Set decode_context_parallel_size=1 for this deployment
2. Use a non-MLA DSpark draft architecture if you must keep decode context parallelism
3. Drop speculative decoding (method=dspark) if decode context parallelism is the hard requirement

Example fix

# before
vllm serve model --speculative-config '{"method": "dspark", "model": "k3-draft"}' -decode-context-parallel-size 4
# after
vllm serve model --speculative-config '{"method": "dspark", "model": "k3-draft"}' -decode-context-parallel-size 1

if spec_cfg.get("method") == "dspark" and dcp_size > 1:
    # K3DSpark (MLA) drafts reject decode context parallelism
    raise ValueError("set decode_context_parallel_size=1 when using an MLA DSpark draft")

def dspark_allows_dcp(draft_architectures: list[str], dcp_size: int) -> bool:
    return dcp_size == 1 or "K3DSparkModel" not in draft_architectures

When it happens

Trigger: Launching with --speculative-config method=dspark plus a K3DSpark draft and -decode-context-parallel-size (or decode_context_parallel_size in ParallelConfig) greater than 1.

Common situations: Reusing a decode-context-parallel launch template (tuned for long-context throughput) when switching the draft model to a K3DSpark/MLA variant; enabling DCP cluster-wide via env or defaults while adding speculative decoding.

Related errors

• target_model_config must be present for dspark
• dspark_draft_topk is only supported by DSpark
• rejection_sample_method='synthetic' requires exactly one of
• synthetic_acceptance_rates must have length {n}, got {rates}
• synthetic_acceptance_rates entries must be in [0, 1], got {r

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/10aeac18e04722fd. Report an issue: GitHub.