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 02:49:26 +11:00
commit d6ab9c92d7
1 changed files with 13 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

17
unsloth/models/llama.py
View file

@ -74,6 +74,15 @@ pass
from math import sqrt as math_sqrt
KV_CACHE_INCREMENT = 128 # KV Cache update size
@torch.compile(options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"fallback_random" : False,
"shape_padding" : True,
"triton.cudagraphs" : False,
"trace.enabled" : True,
"trace.graph_diagram" : True,
}, dynamic = True,)
def LlamaAttention_fast_forward_inference(
self,
hidden_states: torch.Tensor,
@ -142,9 +151,9 @@ def LlamaAttention_fast_forward_inference(
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])
Vn = fast_linear_forward(self.v_proj, Xn, out = self.temp_KV[1])
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])
Vn = fast_linear_forward(self.v_proj, Xn)#, out = self.temp_KV[1])
Qn = Qn.view(bsz, 1, n_heads, head_dim).transpose(1, 2)
Kn = Kn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
Vn = Vn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
@ -202,7 +211,7 @@ def LlamaAttention_fast_forward_inference(
A = torch.matmul(A, Vnn)#, out = Qn)
A = A.transpose(1, 2)
A = A.reshape(bsz, 1, self.hidden_size)
A = fast_linear_forward(self.o_proj, A, out = self.temp_QA[1])
A = fast_linear_forward(self.o_proj, A)#, out = self.temp_QA[1])
return A, (Kn, Vn)
pass