###################################################################### 2024 Jan 07 21:37:48 (from mygpt.py) # 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) ###################################################################### ###################################################################### 2024 Jan 07 21:38:11 (from mygpt.py) # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,) ###################################################################### 2024 Jan 09 14:24:42 (from mygpt.py) # 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] # ) ###################################################################### 2024 Jan 10 08:10:39 (from mygpt.py) # That was a bad idea # G = F.dropout(G, self.attention_dropout, self.training) ###################################################################### 2024 Jan 10 08:46:13 (from mygpt.py) ################################################################# # Flashbacks. This version sucks, about to replace it if self.training and self.proba_flashback > 0.0: warnings.warn("flash back", RuntimeWarning) # 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] )