This commit is contained in:
Yessen Kanapin 2023-08-29 22:22:40 +00:00
parent f4932aec89
commit 6a017f5208

View File

@ -189,6 +189,7 @@ class FlashLlamaAttention(torch.nn.Module):
self.rotary_emb = PositionRotaryEmbedding.static(
dim=self.head_size, base=config.rope_theta, device=weights.device,
config=config
)
self.softmax_scale = self.head_size**-0.5