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 2024-02-08 02:52:43 +11:00
commit 277ca9eecf
1 changed files with 8 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

16
unsloth/models/llama.py
View file

@ -72,7 +72,7 @@ pass
from math import sqrt as math_sqrt
KV_CACHE_INCREMENT = 128 # KV Cache update size
KV_CACHE_INCREMENT = 1024 # KV Cache update size
@torch.compile(options = {
"epilogue_fusion" : True,
@ -144,12 +144,12 @@ def LlamaAttention_fast_forward_inference(
self.RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
self.attention = torch.empty((bsz, n_heads, 1, KV_CACHE_INCREMENT+seq_len), dtype = dtype, device = "cuda")
self.scalar = 1.0 / math_sqrt(self.head_dim)
elif kv_seq_len >= self.paged_attention.shape[0]:
self.paged_attention.resize_((self.paged_attention.shape[0]+KV_CACHE_INCREMENT, 2, bsz, n_kv_heads, head_dim))
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
self.attention.resize_((bsz, n_heads, 1, self.attention.shape[-1]+KV_CACHE_INCREMENT))
pass
# elif kv_seq_len >= self.paged_attention.shape[0]:
# self.paged_attention.resize_((self.paged_attention.shape[0]+KV_CACHE_INCREMENT, 2, bsz, n_kv_heads, head_dim))
# self.paged_attention_K = self.paged_attention[:,0]
# self.paged_attention_V = self.paged_attention[:,1]
# self.attention.resize_((bsz, n_heads, 1, self.attention.shape[-1]+KV_CACHE_INCREMENT))
# pass
Qn = fast_linear_forward(self.q_proj, Xn)#, out = self.temp_QA[0])
Kn = fast_linear_forward(self.k_proj, Xn)#, out = self.temp_KV[0])
@ -206,7 +206,7 @@ def LlamaAttention_fast_forward_inference(
# Attention
A = torch.matmul(Qn, Knn.transpose(2, 3)) #out = self.attention[:,:,:,:attention_size])
A *= self.scalar
A *= 1.0 / (self.head_dim**0.5)
A[:] = torch.nn.functional.softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)
A = torch.matmul(A, Vnn)#, out = Qn)
A = A.transpose(1, 2)