Skip to content

vllm.model_executor.layers.mamba.gdn.base

Classes:

GatedDeltaNetAttention

Bases: PluggableLayer, MambaBase

Base class for GatedDeltaNet attention layer.

Source code in vllm/model_executor/layers/mamba/gdn/base.py 
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
class GatedDeltaNetAttention(PluggableLayer, MambaBase):
    """Base class for GatedDeltaNet attention layer."""

    def __init__(
        self,
        config: PreTrainedConfig,
        vllm_config: VllmConfig,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.prefix = prefix
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.layer_idx = extract_layer_index(prefix)
        self.hidden_size = config.hidden_size
        self.activation = config.hidden_act
        self.layer_norm_epsilon = config.rms_norm_eps
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        self.quant_config = vllm_config.quant_config
        self.speculative_config = vllm_config.speculative_config
        self.num_spec = (
            self.speculative_config.num_speculative_tokens
            if self.speculative_config
            else 0
        )

    @property
    def mamba_type(self) -> MambaAttentionBackendEnum:
        return MambaAttentionBackendEnum.GDN_ATTN

    def get_state_dtype(self) -> tuple[torch.dtype, ...]:
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
            self.model_config.dtype,
            self.cache_config.mamba_cache_dtype,
            self.cache_config.mamba_ssm_cache_dtype,
        )

    def get_state_splits(self) -> tuple[MambaStateSplit, ...] | None:
        if hasattr(self, "num_k_heads"):
            return MambaStateShapeCalculator.gated_delta_net_state_splits(
                self.tp_size,
                self.num_k_heads,
                self.num_v_heads,
                self.head_k_dim,
                self.head_v_dim,
            )
        # KDA layers, with as many key as value heads.
        return MambaStateShapeCalculator.kda_state_splits(
            self.tp_size, self.num_heads, self.head_dim
        )