vllm.models.deepseek_v4.common.spec_decode

Functions:

• needs_mtp_hidden_buffer –

  Whether the drafter reads the target's pre-hc_head residual buffer.

needs_mtp_hidden_buffer(spec_config)

Whether the drafter reads the target's pre-hc_head residual buffer.

DSpark drafts from the target's aux hidden states, so it skips the (max_num_batched_tokens, hc_mult * hidden_size) buffer.

Source code in vllm/models/deepseek_v4/common/spec_decode.py

def needs_mtp_hidden_buffer(spec_config: SpeculativeConfig | None) -> bool:
    """Whether the drafter reads the target's pre-hc_head residual buffer.

    DSpark drafts from the target's aux hidden states, so it skips the
    (max_num_batched_tokens, hc_mult * hidden_size) buffer.
    """
    return (
        spec_config is not None
        and (spec_config.use_eagle() or spec_config.uses_draft_model())
        and not spec_config.use_dspark()
    )