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 2024-01-23 01:33:51 +11:00
commit 1ba28d8e26
1 changed files with 14 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

23
unsloth/models/llama.py
View file

@ -139,7 +139,7 @@ def LlamaAttention_fast_forward_inference(
RH_Q[:, :, :h] = Qn[:, :, h:]; RH_Q[:, :, h:] = Qn[:, :, :h]; torch.neg(RH_Q[:, :, :h], out = RH_Q[:, :, :h]);
Qn *= cos; Qn.addcmul_(RH_Q, sin);
RH_K = torch.empty((n_kv_heads, head_dim), dtype = dtype, device = "cuda")
RH_K = RH_Q[:n_kv_heads, :] # torch.empty((n_kv_heads, head_dim), dtype = dtype, device = "cuda")
RH_K[:, :, :h] = Qn[:, :, h:]; RH_K[:, :, h:] = Kn[:, :, :h]; torch.neg(RH_K[:, :, :h], out = RH_K[:, :, :h]);
Kn *= cos; Kn.addcmul_(RH_K, sin);
@ -149,11 +149,15 @@ def LlamaAttention_fast_forward_inference(
# Grouped query attention
if n_groups != 1:
_, _, cached_len, _ = Kn.shape
Knn = Kn[:, :, None, :, :].expand(1, n_kv_heads, n_groups, cached_len, head_dim)
Vnn = Vn[:, :, None, :, :].expand(1, n_kv_heads, n_groups, cached_len, head_dim)
Knn = Knn.reshape(1, n_heads, cached_len, head_dim)
Vnn = Vnn.reshape(1, n_heads, cached_len, head_dim)
# _, _, cached_len, _ = Kn.shape
# Knn = Kn[:, :, None, :, :].expand(1, n_kv_heads, n_groups, cached_len, head_dim)
# Vnn = Vn[:, :, None, :, :].expand(1, n_kv_heads, n_groups, cached_len, head_dim)
# Knn = Knn.reshape(1, n_heads, cached_len, head_dim)
# Vnn = Vnn.reshape(1, n_heads, cached_len, head_dim)
Knn = Kn[:, :, None, :, :].expand(n_kv_heads, n_groups, kv_seq_len, head_dim)
Vnn = Vn[:, :, None, :, :].expand(n_kv_heads, n_groups, kv_seq_len, head_dim)
Knn = Knn.reshape(n_heads, kv_seq_len, head_dim)
Vnn = Vnn.reshape(n_heads, kv_seq_len, head_dim)
else:
Knn, Vnn = Kn, Vn
@ -175,7 +179,8 @@ pass
torch_silu = torch.nn.functional.silu
def fast_mlp_inference(self, X):
X = X.view(self.hidden_size)
hidden_size = self.hidden_size
X = X.view(hidden_size)
# gate = self.gate_proj(X)
# up = self.up_proj(X)
@ -185,8 +190,8 @@ def fast_mlp_inference(self, X):
gate *= up
# X = self.down_proj(gate)
down = fast_linear_forward(self.down_proj, gate)
X = down.view(1, 1, self.hidden_size)
down = fast_linear_forward(self.down_proj, gate, out = up[:hidden_size])
X = down.view(1, 1, hidden_size)
return X
pass