# 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
# 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:
##############################
-# This is one order of magnitude more complicated than I expected, not
-# elegant, slow, hopefully not buggy
-
-
-def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
- # starting flash backs
- fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long()
- fb_start[:, :, -CL:] = 0
- fb_start[:, :, :CL] = 0
-
- # Remove series longer than CL
- fb_body = fb_start.clone()
- fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)]
- fb_body = fb_body.cumsum(dim=2)
- fb_start = fb_start * (fb_body == 1)
-
- # Set a origin source time (starting time of the chunck to copy
- # here) We set it as the current time minus a multiple of CL to be
- # consistent with the "rolling" caterpillar
- t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :]
- src_time = fb_start * (
- t
- - CL
- * (
- 1
- + (
- torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1)
- ).long()
- )
- )
- src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL]
- src_time = src_time.cumsum(dim=2)
-
- src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device)
- src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL]
- src_head = src_head.cumsum(dim=2)
-
- # combine
- src_delta = fb_start.clone()
- src_delta[:, :, CL:] -= fb_start[:, :, :-CL]
- src_delta = src_delta.cumsum(dim=2)
- src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL]
- src_time += src_delta.cumsum(dim=2) - 1
-
- return src_time, src_head
-
-
-def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
- N, H, CH = V.size(0), V.size(1), rec_V.size(1)
-
- fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
-
- fbt_V = fbt[:, :, :, None]
- fbh_V = fbh[:, :, :, None]
- t = fbt_V.clamp(min=0)
- n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
- d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
- q = V[:, :, t0:t1][n, fbh_V, t, d]
- rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
-
- fbt_K = fbt[:, :, :, None]
- fbh_K = fbh[:, :, :, None]
- t = fbt_K.clamp(min=0)
- n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
- d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
- q = K[:, :, t0:t1][n, fbh_K, t, d]
- rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
-
-
-######################################################################
-
class Caterpillar(nn.Module):
def __init__(
self.caterpillar_height = caterpillar_height
self.attention_dropout = attention_dropout
- warnings.warn("flash back", RuntimeWarning)
- self.proba_flashback = 0.1
+ 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:
- insert_flash_back(
- self.rec_V,
- V,
- self.rec_K,
- K,
- t0,
- t1,
- CL,
- proba=self.proba_flashback / CL,
- )
-
- # 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)[None, None, None, :]
- # dk = torch.arange(DK)[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)
-
- # mk = (
- # torch.rand(self.rec_V[:, :, t0:t1].size()) <= self.proba_flashback
- # ).long()
- # self.rec_V[:, :, t0:t1] = V[n, src_head, src_time, dv]
- # self.rec_K[:, :, t0:t1] = K[n, src_head, src_time, dk]
-
- exit(0)
-
######################################################################
# 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,