ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update llama.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-02 04:18:34 +11:00
commit 7ad1a1fa62
1 changed files with 11 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

23
unsloth/models/llama.py
View file

@ -259,20 +259,19 @@ def LlamaAttention_fast_forward_inference(
Kn *= cos; Kn.addcmul_(RH_K, sin);
# New KV cache
print(K1.shape, V1.shape)
print(Kn.shape, Vn.shape)
Kn = torch.cat([K1, Kn], dim = 2)
Vn = torch.cat([V1, Vn], dim = 2)
print(Kn.shape, Vn.shape)
# if not hasattr(self, "paged_attention"):
# self.paged_attention = torch.empty((2048, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = "cuda")
# self.paged_attention_K = self.paged_attention[0]
# self.paged_attention_V = self.paged_attention[1]
# self.paged_attention_K[:,:,:seq_len,:] = K1
# self.paged_attention_V[:,:,:seq_len,:] = V1
# pass
# Kn = self.paged_attention_K[:,:,seq_len,:] = Kn
# Vn = self.paged_attention_V[:,:,seq_len,:] = Vn
if not hasattr(self, "paged_attention"):
self.paged_attention = torch.empty((2048, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = "cuda")
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
self.paged_attention_K[:seq_len] = K1.permute(2, 0, 1, 3)
self.paged_attention_K[:seq_len] = K1.permute(2, 0, 1, 3)
pass
self.paged_attention_K[seq_len] = Kn.permute(2, 0, 1, 3)
self.paged_attention_V[seq_len] = Vn.permute(2, 0, 1, 3)
Kn = self.paged_attention_K.permute(1, 2, 0, 3)
Vn = self.paged_attention_V.permute(1, 2, 0, 3)
# Grouped query attention
if n_groups != 1: