# with a caching mechanism for keys and values to avoid a O(N^3) cost
# for auto-regression.
+# This implementation is equipped with RNN layers to replace the MHA
+
import math, warnings
import torch, einops
import ffutils
+# from blanket import blanket
+
# import memload
######################################################################
# 1 for the successive tokens.
#
# Modules able to process brackets may implement a cache that is
-# resetted when the input bracket starts at t=0
+# resetted when init_cache is True
class BracketedSequence:
import pscan
-
# X is /.../xTxD A is /.../xT Y_init is /.../xD
return Y
+def pscan_rgrad(grad_Y, A, X, Y_init, dim=-2, eps=1e-2):
+ with torch.no_grad():
+ s_A, s_X = 0, 0
+ for t in range(X.size(dim) - 1, 0, -1):
+ delta = (grad_Y[t] - s_A) / A[t].grad
+ s_A += A[t].grad * delta
+ A[t].grad = delta
+ delta = (grad_Y[t] - s_X) / X[t].grad
+ s_X += X[t].grad * delta
+ X[t].grad = delta
+
+
def pscan_shape(A, X, Y_init):
s = X.size()
A = A.reshape(-1, s[-2])
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
nb_lines,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
caterpillar_height,
attention_dropout=0.0,
len_max=1e5,
+ logger=print,
+ args=None,
):
super().__init__()
warnings.warn("Caterpillar", RuntimeWarning)
- def randw(*d):
- return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
+ def randw(*d, factor=1):
+ return nn.Parameter(torch.randn(*d) * factor / math.sqrt(d[-1]))
self.caterpillar_length = caterpillar_length
self.caterpillar_height = caterpillar_height
self.attention_dropout = attention_dropout
- self.w_G = randw(nb_heads, caterpillar_height, dim_model)
- self.b_G = nn.Parameter(
- torch.full(
- (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
- )
- )
+ self.gate_dropout_proba = args.gate_dropout_proba
+ self.gate_dropout_sync = args.gate_dropout_sync
+ self.gate_dropout_replace = args.gate_dropout_replace
+
+ ######################################################################
+
+ self.w_G = randw(nb_heads, caterpillar_height, dim_model, factor=1e-3)
+ self.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), 0.0))
self.w_K = randw(nb_heads, dim_qk, dim_model)
self.w_V = randw(nb_heads, dim_v, dim_model)
self.w_Q = randw(nb_heads, dim_qk, dim_model)
self.w_O = randw(dim_v * nb_heads, dim_model)
- self.init_K_rec = randw(caterpillar_height, caterpillar_length, dim_qk)
- self.init_V_rec = randw(caterpillar_height, caterpillar_length, dim_v)
+ self.init_K_rec = randw(
+ caterpillar_height,
+ caterpillar_length,
+ dim_qk,
+ )
+ self.init_V_rec = randw(
+ caterpillar_height,
+ caterpillar_length,
+ dim_v,
+ )
- def reset_inner_loss(self):
- self.acc_attention = 0
- self.acc_nb = 0
+ # def reset_inner_loss(self):
+ # self.acc_attention = 0
+ # self.acc_nb = 0
- def get_inner_loss(self):
- # warnings.warn("l2 regularization", RuntimeWarning)
- # return (self.acc_attention / self.acc_nb).pow(2).sum()
- return torch.tensor([0], device=self.w_Q.device)
+ # def get_inner_loss(self):
+ # warnings.warn("l2 regularization", RuntimeWarning)
+ # return (self.acc_attention / self.acc_nb).pow(2).sum()
+ # return torch.tensor([0], device=self.w_Q.device)
def forward(self, bs):
# Dimensions to make the source a bit clearer, that's needed
N = bs.x.size(0)
T = bs.x.size(1)
+ H = self.w_V.size(0)
DV = self.w_V.size(1)
DK = self.w_K.size(1)
- Dout = self.w_O.size(1)
- CH = self.caterpillar_height
- CL = self.caterpillar_length
+ DM = self.w_O.size(1)
+ R = self.caterpillar_height
+ L = self.caterpillar_length
assert (
- t0 >= CL and (t1 - t0) % CL == 0
+ t0 >= L and (t1 - t0) % L == 0
), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
+ # We cache values to deal efficiently with auto-regression
+
if bs.init_cache:
- self.rec_V = X.new_zeros(N, CH, T, DV)
- self.rec_K = X.new_zeros(N, CH, T, DK)
+ self.rec_V = X.new_zeros(N, R, T, DV)
+ self.rec_K = X.new_zeros(N, R, T, DK)
# We start the recurrent sequences with optimizable
# initial values. No idea if it helps.
- self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
- self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
+ self.rec_V[:, :, t0 - L : t0, :] = self.init_V_rec[None, :, :, :]
+ self.rec_K[:, :, t0 - L : t0, :] = self.init_K_rec[None, :, :, :]
+
+ self.cache_Y = X.new_zeros(N, T, DM)
+
+ V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
+ K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
- self.cache_Y = X.new_zeros(N, T, Dout)
+ # V, K = blanket(V), blanket(K)
######################################################################
# Compute the recurrent state
# This is the Gating sequence that modulates the storing of
- # the new key and value in the CH pairs of the current
- # stack. The CH gating values are independent, which means
- # that the current K/V could be stored in all the pairs of the
+ # the new key and value in the R pairs of the current
+ # stack. There are R independent gating values, which means
+ # that the current K/V may be stored in multiple pairs of the
# recurrent state, or not at all.
G = (
- torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
+ torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
).sigmoid()
- G = F.dropout(G, self.attention_dropout, self.training)
+ # Clip the gating to avoid values greater than 1 when several
+ # heads hit the same row
- V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
- K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
+ G = G / G.sum(1, keepdim=True).clamp(min=1)
+
+ # G_star = (1 - G).log().sum(1, keepdim=True).exp()
+
+ ######################################################################
+
+ def recurrence(G, V, K):
+ # We prepare the arguments for the parallel scan
+
+ A = 1 - G.sum(dim=1)
+
+ gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
+ gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
+
+ # We start from cached values, which matters in inference
+
+ init_rec_V = self.rec_V[:, :, t0 - L : t0]
+ init_rec_K = self.rec_K[:, :, t0 - L : t0]
- # We prepare the arguments for the parallel scan
+ # Here there is a trick: Since the stack at position t is
+ # computed by updating that at position t-L, the parallel
+ # scan operates with a period of L. To do so we split the
+ # sequence indexing in two axes, the second of size L, and
+ # run the parallel scan using the first as the sequence index.
- A = 1 - G.sum(1)
- gated_V = torch.einsum("nhet,nhtd->netd", G, V)
- gated_K = torch.einsum("nhet,nhtd->netd", G, K)
+ A = A.unflatten(2, (-1, L))
+ gated_V = gated_V.unflatten(2, (-1, L))
+ gated_K = gated_K.unflatten(2, (-1, L))
- init_rec_V = self.rec_V[:, :, t0 - CL : t0]
- init_rec_K = self.rec_K[:, :, t0 - CL : t0]
+ next_V = pscan_dim(A, gated_V, init_rec_V, dim=2).flatten(2, 3)
+ next_K = pscan_dim(A, gated_K, init_rec_K, dim=2).flatten(2, 3)
- # Here there is a trick: Since the stack at time t is computed
- # by updating that at time t-L, the parallel scan operates
- # with a period of L. To do so we split the time indexing in
- # two axes, the second of size CL, and run the parallel scan
- # using the other alone as the sequence index.
+ return next_V, next_K
- A = A.unflatten(2, (-1, CL))
- gated_V = gated_V.unflatten(2, (-1, CL))
- gated_K = gated_K.unflatten(2, (-1, CL))
+ #################################################################
- next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
- next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
+ next_V, next_K = recurrence(G, V, K)
- # Put back the sequence index
+ if self.training and self.gate_dropout_proba > 0.0:
+ # G is NxHxRxT where r is the caterpillar's row.
- self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
- self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
+ warnings.warn("gate dropout", RuntimeWarning)
+
+ if self.gate_dropout_sync:
+ shape_kill = (N, 1, 1)
+ else:
+ shape_kill = (N, H, R)
+
+ # Pick a point in each of the NxHxR timeline and set this
+ # entry and the following to 1
+ kill = (
+ torch.rand(*shape_kill, t1 - t0, device=G.device).sort(dim=3).indices
+ == 0
+ ).cumsum(dim=3)
+
+ # Keep these mask for only some of the NxHxR
+ kill = kill * (
+ torch.rand(*shape_kill, 1, device=G.device) <= self.gate_dropout_proba
+ )
+
+ # The coefficient to keep are the complementary
+ mask = 1 - kill
+
+ masked_next_V, masked_next_K = recurrence(G * mask, V, K)
+
+ if self.gate_dropout_replace:
+ next_V = next_V.detach()
+ next_K = next_K.detach()
+
+ warnings.warn("the rescaling is probably a bad idea", RuntimeWarning)
+
+ next_V = next_V + (masked_next_V - masked_next_V.detach()) / (
+ 1 - self.gate_dropout_proba
+ )
+ next_K = next_K + (masked_next_K - masked_next_K.detach()) / (
+ 1 - self.gate_dropout_proba
+ )
+
+ self.rec_V[:, :, t0:t1] = next_V
+ self.rec_K[:, :, t0:t1] = next_K
######################################################################
# compute the readout
Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
- # We build tensors NxHxTxFxL where N is the sample index, H
- # the head, T the time, F the row in the caterpillar, and L
+ # Q = blanket(Q)
+
+ # We build tensors NxHxTxRxL where N is the sample index, H
+ # the head, T the time, R the row in the caterpillar, and L
# the column in the caterpillar
windowed_V = moving_window(
- self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
+ self.rec_V[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
)
windowed_K = moving_window(
- self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
+ self.rec_K[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
)
- # We have an attention score for each of the CHxCL values
+ # We have an attention score for each of the RxL values
ar = torch.einsum(
- "nhtd,nftld->nhtfl",
+ "nhtd,nrtld->nhtrl",
Q,
windowed_K,
) / math.sqrt(DK)
# Compute the final output
+ # Y = blanket(Y)
+
self.cache_Y[:, t0:t1] = Y @ self.w_O
return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
nb_heads=1,
causal=False,
attention_dropout=0.0,
+ logger=print,
+ args=None,
):
super().__init__()
nb_blocks,
nb_lines=None,
caterpillar_height=None,
- dim_rec_v=-1,
causal=False,
dropout=0.0,
len_max=1e5,
- attention_layer="kvrec",
+ attention_layer="caterpillar",
+ logger=print,
+ args=None,
):
super().__init__()
- assert attention_layer in {"mha", "dumbrec", "kvrec", "caterpillar"}
+ assert attention_layer in {
+ "mha",
+ "dumbrec",
+ "kvrec",
+ "caterpillar",
+ }, f"Unknown attention operator {attention_layer}."
if attention_layer == "caterpillar":
assert nb_lines % caterpillar_height == 0
nb_heads=nb_heads,
causal=causal,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "dumbrec":
return DumbRec(
dim_model=dim_model,
dim_qk=dim_keys,
- dim_v=dim_rec_v,
+ dim_v=dim_model // nb_heads,
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "kvrec":
return KVRec(
dim_model=dim_model,
dim_qk=dim_keys,
- dim_v=dim_rec_v,
+ dim_v=dim_model // nb_heads,
nb_heads=nb_heads,
nb_lines=nb_lines,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
elif attention_layer == "caterpillar":
return Caterpillar(
dim_model=dim_model,
dim_qk=dim_keys,
- dim_v=dim_rec_v,
+ dim_v=dim_model // nb_heads,
nb_heads=nb_heads,
caterpillar_length=self.caterpillar_length,
caterpillar_height=self.caterpillar_height,
attention_dropout=dropout,
+ logger=logger,
+ args=args,
)
else:
raise ValueError(f"Unknown attention type {attention_layer}.")
######################################################################
if __name__ == "__main__":
- print("Basic check.")
+ import argparse
+
+ import numpy as np
+ import matplotlib.pyplot as plt
+ import matplotlib.collections as mc
+
+ args = argparse.Namespace(
+ gate_dropout_proba=0.0, gate_dropout_sync=True, gate_dropout_replace=False
+ )
+
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+ dim_model, dim_keys, nb_heads = 512, 64, 1
+ dropout = 0.1
+
+ caterpillar = Caterpillar(
+ dim_model=dim_model,
+ dim_qk=dim_keys,
+ dim_v=dim_model // nb_heads,
+ nb_heads=nb_heads,
+ caterpillar_length=16,
+ caterpillar_height=32,
+ attention_dropout=dropout,
+ args=args,
+ ).to(device)
+
+ qkv = QKVAttention(
+ dim_model=dim_model,
+ dim_qk=dim_keys,
+ dim_v=dim_model // nb_heads,
+ nb_heads=nb_heads,
+ causal=True,
+ attention_dropout=dropout,
+ args=args,
+ ).to(device)
+
+ linear = CacheWrapper(nn.Linear(512, 512)).to(device)
+
+ x = torch.randn(1, 256, dim_model)
+
+ x = x.to(device)
+ x.requires_grad_()
+
+ ######################################################################
+
+ fig = plt.figure()
+ fig.set_figheight(6)
+ fig.set_figwidth(8)
+
+ ax = fig.add_subplot(1, 1, 1)
+
+ # ax.set_xlim(-1.5, 1.5)
+ # ax.set_ylim(-1.5, 1.5)
+ # ax.set(aspect=1)
+ # ax.spines.right.set_visible(False)
+ # ax.spines.top.set_visible(False)
+
+ # dt = 0.01
+ # t = np.arange(dt, 20.0, dt)
+ # ax.semilogx(t, np.exp(-t / 5.0))
+ # ax.grid()
+ ax.set_yscale("log")
+
+ ######################################################################
+
+ for label, model, thickness in [
+ ("nn.Linear", linear, 0.2),
+ ("mygpy.QKVAttention", qkv, 1),
+ ("mygpt.Caterpillar", caterpillar, 2),
+ ]:
+ y = model(BracketedSequence(x, 32, x.size(1) - 32, init_cache=True)).x
+
+ for n, p in [("input", x)] + list(model.named_parameters()):
+ print(f"Processing {model}.{n}")
+ data = []
+ for t in range(y.size(1)):
+ sg = 0
+ for d in torch.randperm(y.size(2))[:8]:
+ sg += torch.autograd.grad(y[0, t, d], p, retain_graph=True)[0]
+ assert not sg.isinf().any()
+ assert not sg.isnan().any()
+ data.append([t, sg.sum().item()])
+
+ data = torch.tensor(data)
+ # cx, cy = data[:, 0], data[:, 1]
+ cy = data[:, 1].sort().values
+ cx = torch.linspace(0, 1, cy.size(0))
+ ax.plot(
+ cx, cy, label=label + "." + n, linewidth=thickness
+ ) # , color='gray', label='Input')
+
+ # ax.legend(frameon=False, loc="top right")
+
+ # Put a legend to the right of the current axis
+ box = ax.get_position()
+ ax.set_position([box.x0, box.y0, box.width * 0.8, box.height])
+ ax.legend(loc="center left", bbox_to_anchor=(1, 0.5))
+
+ filename = "plot.pdf"
+ print(f"saving {filename}")
+ fig.savefig(filename, bbox_inches="tight")
+
+ # if args.window and hasattr(plt.get_current_fig_manager(), 'window'):
+ # plt.get_current_fig_manager().window.setGeometry(2, 2, 1024, 768)
+ # plt.show()
+
+ exit(0)
+
+ ######################################################################
m = Caterpillar(
dim_model=4,
print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
exit(0)
- device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
vocabulary_size = 128
x = torch.randint(vocabulary_size, (6, 1024))