# 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
self.caterpillar_height = caterpillar_height
self.attention_dropout = attention_dropout
- warnings.warn("flash back", RuntimeWarning)
- self.proba_flashback = 1e-2
+ self.proba_gate_dropout = 0.0
self.w_G = randw(nb_heads, caterpillar_height, dim_model)
self.b_G = nn.Parameter(
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)
+
######################################################################
# 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 multiple pairs of the
+ # stack. There are CH 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]
).sigmoid()
- # That bas a bad idea
- # 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)
# We prepare the arguments for the parallel scan
gated_V = torch.einsum("nhet,nhtd->netd", G, V)
gated_K = torch.einsum("nhet,nhtd->netd", G, K)
+ # We start from cached values, which matters in inference
+
init_rec_V = self.rec_V[:, :, t0 - CL : t0]
init_rec_K = self.rec_K[:, :, t0 - CL : t0]
- # 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 as the sequence index.
+ ######################################################################
+
+ if self.training and self.proba_gate_dropout > 0.0:
+ warnings.warn("gate dropout", RuntimeWarning)
+ epsilon = 0.5
+
+ #################################################################
+ # Associative scan
+
+ # Here there is a trick: Since the stack at position t is
+ # computed by updating that at position t-CL, the parallel
+ # scan operates with a period of CL. To do so we split the
+ # sequence indexing in two axes, the second of size CL, and
+ # run the parallel scan using the first as the sequence index.
A = A.unflatten(2, (-1, CL))
gated_V = gated_V.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)
- # Put back the sequence index
-
self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
- if self.training and self.proba_flashback > 0.0:
- # This piece of code makes the assumption that there is
- # nothing informative before t0, otherwise we'd have to
- # implement a cache for V and K too. This should not be
- # too much of a problem since this is used only during
- # train, where full sequence are available
-
- n = torch.arange(N, device=X.device)[:, None, None, None]
- t = torch.arange(t0, t1, device=X.device)[None, None, :, None]
- dv = torch.arange(DV, device=X.device)[None, None, None, :]
- dk = torch.arange(DK, device=X.device)[None, None, None, :]
-
- u = (
- torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
- ) * CL
-
- src_time = t - u - t0
- src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
-
- mask = (
- torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
- ).long()
-
- self.rec_V[:, :, t0:t1] = (
- mask * V[n, src_head, src_time, dv]
- + (1 - mask) * self.rec_V[:, :, t0:t1]
- )
-
- self.rec_K[:, :, t0:t1] = (
- mask * K[n, src_head, src_time, dk]
- + (1 - mask) * self.rec_K[:, :, t0:t1]
- )
-
######################################################################
# compute the readout
nb_blocks,
nb_lines=None,
caterpillar_height=None,
- dim_rec_v=-1,
causal=False,
dropout=0.0,
len_max=1e5,
):
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
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,
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,
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,