273 lines
9.8 KiB
Python
273 lines
9.8 KiB
Python
from typing import Optional, Tuple, Union, Callable, List, Any, Generator
|
||
from einops import rearrange
|
||
|
||
import torch
|
||
import torch.nn.functional as F
|
||
import torch.utils.checkpoint
|
||
from torch.nn import CrossEntropyLoss
|
||
from torch import nn
|
||
import torch.nn.init as init
|
||
|
||
# for 0.1B
|
||
n_layer = 3
|
||
n_embd = 256
|
||
D_DECAY_LORA = 64
|
||
D_AAA_LORA = 64
|
||
D_MV_LORA = 32
|
||
D_GATE_LORA = 128
|
||
|
||
dim_att = n_embd
|
||
dim_ffn = n_embd * 4
|
||
|
||
vocab_size = 32
|
||
|
||
# DTYPE = torch.bfloat16
|
||
DTYPE = torch.float32 # better
|
||
|
||
head_size_a = 64 # don't change
|
||
HS = head_size_a
|
||
|
||
|
||
class RWKV_Tmix_x070(nn.Module):
|
||
def __init__(self, layer_id):
|
||
super().__init__()
|
||
self.layer_id = layer_id
|
||
|
||
self.head_size = head_size_a
|
||
self.n_head = dim_att // self.head_size
|
||
assert dim_att % self.n_head == 0
|
||
|
||
H = self.n_head
|
||
HS = self.head_size
|
||
C = n_embd
|
||
|
||
self.x_r = nn.Parameter(torch.empty(1, 1, C))
|
||
self.x_w = nn.Parameter(torch.empty(1, 1, C))
|
||
self.x_k = nn.Parameter(torch.empty(1, 1, C))
|
||
self.x_v = nn.Parameter(torch.empty(1, 1, C))
|
||
self.x_a = nn.Parameter(torch.empty(1, 1, C))
|
||
self.x_g = nn.Parameter(torch.empty(1, 1, C))
|
||
|
||
self.w0 = nn.Parameter(torch.empty(1, 1, C))
|
||
self.w1 = nn.Parameter(torch.empty(C, D_DECAY_LORA))
|
||
self.w2 = nn.Parameter(torch.empty(D_DECAY_LORA, C))
|
||
|
||
self.a0 = nn.Parameter(torch.empty(1, 1, C))
|
||
self.a1 = nn.Parameter(torch.empty(C, D_AAA_LORA))
|
||
self.a2 = nn.Parameter(torch.empty(D_AAA_LORA, C))
|
||
|
||
self.v0 = nn.Parameter(torch.empty(1, 1, C))
|
||
self.v1 = nn.Parameter(torch.empty(C, D_MV_LORA))
|
||
self.v2 = nn.Parameter(torch.empty(D_MV_LORA, C))
|
||
|
||
self.g1 = nn.Parameter(torch.empty(C, D_GATE_LORA))
|
||
self.g2 = nn.Parameter(torch.empty(D_GATE_LORA, C))
|
||
|
||
self.k_k = nn.Parameter(torch.empty(1, 1, C))
|
||
self.k_a = nn.Parameter(torch.empty(1, 1, C))
|
||
self.r_k = nn.Parameter(torch.empty(H, HS))
|
||
|
||
self.receptance = nn.Linear(C, C, bias=False)
|
||
self.key = nn.Linear(C, C, bias=False)
|
||
self.value = nn.Linear(C, C, bias=False)
|
||
self.output = nn.Linear(C, C, bias=False)
|
||
self.ln_x = nn.GroupNorm(H, C, eps=64e-5) # !!! notice eps value !!!
|
||
|
||
def forward(self, x, v_first):
|
||
B, T, C = x.size() # seq_len
|
||
H = self.n_head # 12
|
||
|
||
xx = torch.zeros_like(x) # time_shift [1, seq_len, 768] -> [1, seq_len, 768]
|
||
xx[:, 0, :] = -x[:, 0, :]
|
||
xx[:, 1:, :] = x[:, :-1, :] - x[:, 1:, :]
|
||
|
||
xr = x + xx * self.x_r # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
xw = x + xx * self.x_w # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
xk = x + xx * self.x_k # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
xv = x + xx * self.x_v # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
xa = x + xx * self.x_a # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
xg = x + xx * self.x_g # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
|
||
r = self.receptance(xr) # Linear [1, seq_len, 768] -> [1, seq_len, 768]
|
||
|
||
xw = torch.tanh(xw @ self.w1) # -> [1, seq_len, 64]
|
||
xw = xw @ self.w2 + self.w0 # -> [1, seq_len, 768]
|
||
xw = -F.softplus(-xw) # 函数的输出范围为 [0,∞)
|
||
w = xw - 0.5 # -> [1, seq_len, 768]
|
||
|
||
k = self.key(xk) # Linear [1, seq_len, 768] -> [1, seq_len, 768]
|
||
|
||
v = self.value(xv) # Linear [1, seq_len, 768] -> [1, seq_len, 768]
|
||
|
||
if self.layer_id == 0:
|
||
v_first = v # store the v of the first layer
|
||
else:
|
||
xv = (xv @ self.v1) @ self.v2
|
||
xv = xv + self.v0
|
||
xv = torch.sigmoid(xv)
|
||
v = v + (v_first - v) * xv # add value residual # -> [1, seq_len, 768]
|
||
|
||
xa = (xa @ self.a1) @ self.a2
|
||
xa = xa + self.a0
|
||
a = torch.sigmoid(xa) # -> [1, seq_len, 768]
|
||
|
||
xg = xg @ self.g1
|
||
xg = torch.sigmoid(xg)
|
||
g = xg @ self.g2 # -> [1, seq_len, 768]
|
||
|
||
kk = k * self.k_k # [1, seq_len, 768] * [1, 1, 768] -> [1, seq_len, 768]
|
||
kk = F.normalize(kk.view(B, T, H, -1), dim=-1, p=2.0).view(B, T, C) # -> [1, seq_len, 768]
|
||
|
||
k = k * (1 + (a - 1) * self.k_a) # -> [1, seq_len, 768]
|
||
|
||
# start op
|
||
a_op = -kk
|
||
b_op = kk * a
|
||
|
||
B, T, C = r.size() # 768
|
||
H = C // HS # 12
|
||
|
||
r_op = r.view(B, T, H, HS, 1).float() # -> [1, seq_len, 12, 64, 1]
|
||
k_op = k.view(B, T, H, 1, HS).float() # -> [1, seq_len, 12, 1, 64]
|
||
v_op = v.view(B, T, H, HS, 1).float() # -> [1, seq_len, 12, 64, 1]
|
||
a_op = a_op.view(B, T, H, HS, 1).float() # -> [1, seq_len, 12, 64, 1]
|
||
b_op = b_op.view(B, T, H, 1, HS).float() # -> [1, seq_len, 12, 1, 64]
|
||
|
||
w_op = w.view(B, T, H, HS).float()
|
||
w_op = torch.exp(-torch.exp(w_op)) # -> [1, seq_len, 12, 64]
|
||
w_op = w_op.view(B, T, H, 1, HS) # -> [1, seq_len, 12, 1, 64]
|
||
|
||
out = torch.zeros((B, T, H, HS), device=r_op.device, dtype=torch.float) # -> [1, seq_len, 12, 64]
|
||
state = torch.zeros((B, H, HS, HS), device=r_op.device, dtype=torch.float) # -> [1, seq_len, 12, 64]
|
||
|
||
for t in range(T):
|
||
rr_op = r_op[:, t, :] # [1, seq_len, 12, 64, 1] -> [1, 12, 64, 1]
|
||
kk_op = k_op[:, t, :] # [1, seq_len, 12, 1, 64] -> [1, 12, 1, 64]
|
||
vv_op = v_op[:, t, :] # [1, seq_len, 12, 64, 1] -> [1, 12, 64, 1]
|
||
aa_op = a_op[:, t, :] # [1, seq_len, 12, 64, 1] -> [1, 12, 64, 1]
|
||
bb_op = b_op[:, t, :] # [1, seq_len, 12, 1, 64] -> [1, 12, 1, 64]
|
||
ww_op = w_op[:, t, :] # [1, seq_len, 12, 64] -> [1, 12, 1, 64]
|
||
|
||
state = state * ww_op + state @ aa_op @ bb_op + vv_op @ kk_op # -> [1, 12, 64, 64]
|
||
out[:, t, :] = (state @ rr_op).view(B, H, HS) # -> [1, seq_len, 12, 64]
|
||
|
||
x = out.view(B, T, C).to(dtype=DTYPE) # -> [1, seq_len, 768]
|
||
# end op
|
||
|
||
x = self.ln_x(x.view(B * T, C)).view(B, T, C) # -> [1, seq_len, 768]
|
||
|
||
xx = r.view(B, T, H, -1) * k.view(B, T, H, -1) # -> [1, seq_len, 12, 64]
|
||
xx = xx * self.r_k # -> [1, seq_len, 12, 64]
|
||
xx = xx.sum(dim=-1, keepdim=True) # -> [1, seq_len, 12, 1]
|
||
xx = xx * v.view(B, T, H, -1) # [1, seq_len, 12, 1] x [1, seq_len, 12, 64] -> [1, seq_len, 12, 64]
|
||
xx = xx.view(B, T, C) # -> [1, seq_len, 768]
|
||
x = x + xx # -> [1, seq_len, 768]
|
||
x = self.output(x * g) # Linear -> [1, seq_len, 768]
|
||
return x, v_first
|
||
|
||
|
||
class RWKV_CMix_x070(nn.Module):
|
||
def __init__(self, layer_id):
|
||
super().__init__()
|
||
self.layer_id = layer_id
|
||
|
||
# with torch.no_grad():
|
||
self.x_k = nn.Parameter(torch.empty(1, 1, n_embd))
|
||
|
||
self.key = nn.Linear(n_embd, dim_ffn, bias=False)
|
||
self.value = nn.Linear(dim_ffn, n_embd, bias=False)
|
||
|
||
def forward(self, x):
|
||
|
||
shift = torch.zeros_like(x)
|
||
shift[:, 1:, :] = x[:, :-1, :]
|
||
|
||
xx = shift - x # time_shift -> [1, seq_len, 768]
|
||
|
||
k = x + xx * self.x_k # -> [1, seq_len, 768]
|
||
k = torch.relu(self.key(k)) ** 2 # Linear -> [1, seq_len, 768]
|
||
return self.value(k) # Linear -> [1, seq_len, 768]
|
||
|
||
|
||
class Block(nn.Module):
|
||
def __init__(self, layer_id):
|
||
super().__init__()
|
||
self.layer_id = layer_id
|
||
|
||
self.ln0 = nn.LayerNorm(n_embd) # only used in block 0, should be fused with emb
|
||
self.ln1 = nn.LayerNorm(n_embd)
|
||
self.ln2 = nn.LayerNorm(n_embd)
|
||
|
||
self.att = RWKV_Tmix_x070(layer_id)
|
||
self.ffn = RWKV_CMix_x070(layer_id)
|
||
|
||
def forward(self, x, v_first):
|
||
|
||
if self.layer_id == 0:
|
||
x = self.ln0(x) # LayerNorm -> [1, seq_len, 768] normal at dim 768 * γ + β
|
||
ln = self.ln1(x) # LayerNorm -> [1, seq_len, 768] normal at dim 768 * γ + β
|
||
|
||
xx, v_first = self.att(ln, v_first) # [1, seq_len, 768] -> [1, seq_len, 768]
|
||
x = x + xx # [1, seq_len, 768] -> [1, seq_len, 768]
|
||
x = x + self.ffn(self.ln2(x)) # [1, seq_len, 768] -> [1, seq_len, 768]
|
||
|
||
return x, v_first
|
||
|
||
|
||
class RWKV(nn.Module):
|
||
def __init__(self):
|
||
super().__init__()
|
||
self.emb = nn.Embedding(vocab_size, n_embd)
|
||
|
||
self.blocks = nn.ModuleList([Block(i) for i in range(n_layer)])
|
||
|
||
self.ln_out = nn.LayerNorm(n_embd)
|
||
self.head = nn.Linear(n_embd, vocab_size, bias=False)
|
||
|
||
def forward(self, idx):
|
||
|
||
x = self.emb(idx) # [1, seq_len] -> [1, seq_len, 768]
|
||
|
||
v_first = torch.empty_like(x) # -> [1, seq_len, 768]
|
||
for block in self.blocks:
|
||
x, v_first = block(x, v_first)
|
||
|
||
x = self.ln_out(x) # [1, seq_len, 768] -> [1, seq_len, 768]
|
||
x = self.head(x) # [1, seq_len, 768] -> [1, seq_len, 65536]
|
||
|
||
return x
|
||
|
||
|
||
class RWKVLMHeadModel(nn.Module):
|
||
def __init__(self, config):
|
||
super().__init__()
|
||
self.config = config
|
||
self.rwkv = RWKV()
|
||
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
||
self.hook_attention = None
|
||
|
||
for name, param in self.rwkv.named_parameters():
|
||
init.normal_(param.data)
|
||
|
||
def forward(
|
||
self,
|
||
input_ids: Optional[torch.LongTensor] = None,
|
||
labels: Optional[torch.LongTensor] = None,
|
||
token_type_ids: Optional[torch.LongTensor] = None,
|
||
**kwargs,
|
||
):
|
||
lm_logits = self.rwkv(input_ids)
|
||
|
||
loss = None
|
||
if labels is not None:
|
||
labels = labels.to(lm_logits.device)
|
||
shift_labels = labels[..., 1:].contiguous().view(-1)
|
||
shift_logits = lm_logits[..., :-1, :].contiguous()
|
||
shift_logits = shift_logits.view(-1, shift_logits.size(-1))
|
||
mask = shift_labels < self.config.vocab_size
|
||
shift_labels = shift_labels[mask]
|
||
shift_logits = shift_logits[mask]
|
||
loss = CrossEntropyLoss()(shift_logits, shift_labels)
|
||
|
||
return lm_logits, loss
|