MCPcopy Create free account
hub / github.com/evilsocket/cake / DecoderBlock

Class DecoderBlock

cake-core/src/models/vibevoice/vae_decoder.rs:66–80  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

64/// A single decoder block with depthwise conv mixer + FFN.
65#[derive(Debug, Clone)]
66struct DecoderBlock {
67 norm_weight: Tensor,
68 ffn_norm_weight: Tensor,
69 eps: f32,
70 gamma: Tensor,
71 /// Depthwise conv weight: (channels, 1, 7) — stored as (channels, 7) for manual impl
72 mixer_weight: Tensor,
73 mixer_bias: Tensor,
74 ffn_gamma: Tensor,
75 ffn_linear1_weight: Tensor,
76 ffn_linear1_bias: Option<Tensor>,
77 ffn_linear2_weight: Tensor,
78 ffn_linear2_bias: Option<Tensor>,
79 backend: Arc<dyn ComputeBackend>,
80}
81
82/// Manual depthwise conv1d using broadcast_mul + sum.
83/// This avoids candle's pathologically slow grouped Conv1d CUDA kernel.

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected