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

fast inference

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-02 19:07:41 +11:00
commit 0e1d67d4e1
2 changed files with 9 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

11
unsloth/models/llama.py
View file

@ -236,7 +236,7 @@ def LlamaAttention_fast_forward_inference(
kv_seq_len = seq_len + 1
if not hasattr(self, "paged_attention"):
self.paged_attention = torch.zeros((2048, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = "cuda")
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)
@ -244,6 +244,8 @@ def LlamaAttention_fast_forward_inference(
self.temp_QA = torch.empty((2, bsz, 1, hd), dtype = dtype, device = "cuda")
self.temp_KV = torch.empty((2, bsz, 1, n_kv_heads*head_dim), dtype = dtype, device = "cuda")
self.RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
self.attention = torch.empty((bsz, n_heads, 1, 2048), dtype = dtype, device = "cuda")
self.scalar = 1.0 / math_sqrt(self.head_dim)
pass
Qn = fast_linear_forward(self.q_proj, Xn, out = self.temp_QA[0])
@ -286,9 +288,8 @@ def LlamaAttention_fast_forward_inference(
Knn, Vnn = Kn, Vn
# Attention
A = torch.matmul(Qn, Knn.transpose(2, 3))
print(A.shape)
A *= 1.0 / math_sqrt(self.head_dim)
A = torch.matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,:kv_seq_len])
A *= self.scalar
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)
@ -358,6 +359,8 @@ def LlamaAttention_fast_forward(
del self.temp_QA
del self.temp_KV
del self.RH_Q
del self.attention
del self.scalar
pass
bsz, q_len, _ = hidden_states.size()

2
unsloth/models/mistral.py
View file

@ -62,6 +62,8 @@ def MistralAttention_fast_forward(
del self.temp_QA
del self.temp_KV
del self.RH_Q
del self.attention
del self.scalar
pass
bsz, q_len, _ = hidden_states.size()