Remove extra contiguous (#5953)
This commit is contained in:
@@ -752,7 +752,7 @@ class DeepseekV2AttentionMLA(nn.Module):
|
||||
q_nope_out = q_nope_out.transpose(0, 1)
|
||||
|
||||
k_nope = latent_cache[..., : self.kv_lora_rank]
|
||||
k_nope = self.kv_a_layernorm(k_nope.contiguous()).unsqueeze(1)
|
||||
k_nope = self.kv_a_layernorm(k_nope).unsqueeze(1)
|
||||
k_pe = latent_cache[..., self.kv_lora_rank :].unsqueeze(1)
|
||||
|
||||
q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
|
||||
|
||||
Reference in New Issue
Block a user