torch.nn.attention.bias.CausalBias

class torch.nn.attention.bias.CausalBias(variant, seq_len_q, seq_len_kv)

表示因果注意力模式的偏差。有关偏差结构的概述，请参阅CausalVariant 枚举。

此类用于定义因果（三角形）注意力偏差。为了构建偏差，存在两个工厂函数：causal_upper_left() 和causal_lower_right()。

示例

from torch.nn.attention.bias import causal_lower_right

bsz, num_heads, seqlen_q, seqlen_kv, head_dim = 32, 8, 4, 12, 8

# Create a lower-right causal bias
attn_bias = causal_lower_right(seqlen_q, seqlen_kv)

q = torch.randn(bsz, num_heads, seqlen_q, head_dim, device="cuda", dtype=torch.float16)
k = torch.randn(bsz, num_heads, seqlen_kv, head_dim, device="cuda", dtype=torch.float16)
v = torch.randn(bsz, num_heads, seqlen_kv, head_dim, device="cuda", dtype=torch.float16)

out = F.scaled_dot_product_attention(q, k, v, attn_bias)

警告

此类为原型，可能会发生更改。