Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

complex rotary memory #185

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

complex rotary memory #185

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
65 changes: 65 additions & 0 deletions RWKV-v4neo/src/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,12 +450,66 @@ def forward(self, x):
x = RUN_CUDA(B, T, C, H, r, k, v, w=self.time_decay, u=self.time_faaaa)

return self.jit_func_2(x, g)

########################################################################################################
# RWKV: RWKV Wavenet-mem + rotary memory
########################################################################################################
class Short_Mem(nn.Module):
def __init__(self, args, shiftAmount=1):
super().__init__()
# self.time_shift1 = TimeShift(args.n_embd, shiftAmount=shiftAmount, batch=args.micro_bsz)
self.time_shift1 = nn.ZeroPad2d((0, 0, shiftAmount, -shiftAmount))
self.activation = nn.Sequential(
nn.Linear(args.n_embd*2, args.n_embd, bias=False),
nn.Sigmoid(),
)

def forward(self, x):
xv = self.activation(torch.cat([self.time_shift1(x),x], dim=-1))
return xv

class WaveNet_Mem(Short_Mem):
def __init__(self, args, layer_id, modulo=12, undialated=False):
if undialated:
super().__init__(args, shiftAmount=1)
else:
super().__init__(args, shiftAmount=2**(layer_id%modulo))

class Rotary_Memory(nn.Module):
def __init__(self, args, layer_id):
nn.Module.__init__(self)
self.args = args
self.layer_id = layer_id

self.complexsize = args.n_embd
self.short = WaveNet_Mem(args, layer_id, undialated=True)
self.key = nn.Linear(args.n_embd,self.complexsize*2, bias=False, dtype=torch.bfloat16)
# self.cumprod = CumProd(torch.complex(torch.ones(args.micro_bsz, 1, self.complexsize), torch.zeros(args.micro_bsz, 1, self.complexsize)))
# self.cummax = CumMax()
self.activation = nn.Linear(self.complexsize*2, args.n_embd, bias=False, dtype=torch.bfloat16)

def forward(self, x):
B, T, C = x.size()
k = self.key(x).float()


complexval = torch.view_as_complex(k.reshape(B, T, self.complexsize,2))
# scale = self.cummax(torch.abs(complexval))
scale = torch.cummax(torch.abs(complexval), dim=-2)[0]
complexval2 = complexval / scale
# kv = self.cumprod(complexval2)
kv = torch.cumprod(complexval2, dim=-2)
out = self.activation(torch.view_as_real(kv).reshape(B, T, self.complexsize*2)) * self.short(x* scale)

return out

########################################################################################################
# RWKV: RWKV Time-mix + RWKV Channel-mix
########################################################################################################




class RWKV_TimeMix(MyModule):
def __init__(self, args, layer_id):
super().__init__()
Expand Down Expand Up @@ -632,6 +686,11 @@ def __init__(self, args, layer_id):
self.ln1 = nn.LayerNorm(args.n_embd)
self.ln2 = nn.LayerNorm(args.n_embd)

self.use_mem = 'c' in os.environ["RWKV_MY_TESTING"]
if self.use_mem:
self.ln3 = nn.LayerNorm(args.n_embd)
self.mem = Rotary_Memory(args, layer_id)

if self.layer_id == 0:
self.ln0 = nn.LayerNorm(args.n_embd)
if args.my_pos_emb > 0:
Expand Down Expand Up @@ -663,6 +722,8 @@ def __init__(self, args, layer_id):
if args.dropout > 0:
self.drop0 = nn.Dropout(p = args.dropout)
self.drop1 = nn.Dropout(p = args.dropout)
if self.use_mem:
self.drop2 = nn.Dropout(p = args.dropout)

def forward(self, x, x_emb=None):
args = self.args
Expand All @@ -678,12 +739,16 @@ def forward(self, x, x_emb=None):
x = x + self.ffnPre(self.ln1(x))
else:
x = x + self.att(self.ln1(x))
if self.use_mem:
x = x + self.mem(self.ln3(x))
x = x + self.ffn(self.ln2(x))
else:
if self.layer_id == 0 and args.pre_ffn > 0:
x = self.drop0(x + self.ffnPre(self.ln1(x)))
else:
x = self.drop0(x + self.att(self.ln1(x)))
if self.use_mem:
x = self.drop2(x + self.mem(self.ln3(x)))
x = self.drop1(x + self.ffn(self.ln2(x)))

if args.tiny_att_dim > 0 and self.layer_id == args.tiny_att_layer:
Expand Down