Skip to content

vllm.models.deepseek_v4.attention

DeepseekV4 MLA Attention Layer

_select_v4_sparse_impl 

_select_v4_sparse_impl() -> type[
    DeepseekV4SparseMLAAttentionImpl
]

Pick the platform-specific V4 sparse MLA impl class. Sole platform check.

Source code in vllm/models/deepseek_v4/attention.py 
76
77
78
79
80
81
82
83
84
85
86
87
88
def _select_v4_sparse_impl() -> "type[DeepseekV4SparseMLAAttentionImpl]":
    """Pick the platform-specific V4 sparse MLA impl class. Sole platform check."""
    if current_platform.is_rocm():
        from vllm.models.deepseek_v4.amd.rocm import (
            DeepseekV4ROCMAiterMLASparseImpl,
        )

        return DeepseekV4ROCMAiterMLASparseImpl
    from vllm.models.deepseek_v4.nvidia.flashmla import (
        DeepseekV4FlashMLASparseImpl,
    )

    return DeepseekV4FlashMLASparseImpl