return self.x[:, self.first : self.first + self.nb]
def complete(self):
- return self.first == 0 and self.nb == x.size(1)
+ return self.first == 0 and self.nb == self.x.size(1)
######################################################################
"nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
) / math.sqrt(self.w_q.size(1))
- if self.record_attention:
- self.a = a
-
if self.causal:
if bs_q.first == 0:
self.cache_attzero = (
)
a = a.softmax(dim=3)
+
+ if self.record_attention:
+ self.a = a
+
a = F.dropout(a, self.attention_dropout, self.training)
y = torch.einsum(
m.weight.fill_(1.0)
def forward(self, bs):
+ # print(f"GENERATE {bs.first} {bs.first+bs.nb}")
bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
bs = self.embedding(bs)
bs = self.trunk(bs)
# unchanged.
def masked_inplace_autoregression(
- self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
+ self,
+ input,
+ ar_mask,
+ deterministic_synthesis=False,
+ forbidden_tokens=None,
+ forced_biases=None,
):
to_generate = (ar_mask.sum(0) > 0).nonzero()
if to_generate.min() > 0:
logits = output[:, s]
if forbidden_tokens is not None:
logits = logits.masked_fill(forbidden_tokens, float("-inf"))
+ if forced_biases is not None:
+ logits = logits + forced_biases[None, :]
if deterministic_synthesis:
t_next = logits.argmax(1)
else: