- def masked_inplace_autoregression(
- self,
- input,
- ar_mask,
- temperature=1.0,
- deterministic_synthesis=False,
- forbidden_tokens=None,
- forced_biases=None,
- ):
- sum_logits = 0
- to_generate = (ar_mask.sum(0) > 0).nonzero()
- if to_generate.min() > 0:
- self(
- BracketedSequence(input, 0, to_generate.min())
- ) # Needed to initialize the model's cache
- for s in range(to_generate.min(), to_generate.max() + 1):
- output = self(BracketedSequence(input, s, 1)).x
- logits = output[:, s]
-
- logits = logits.log_softmax(dim=-1) / temperature
-
- 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)
+ def decode(self, z_shape):
+ bs = self.decoder(z_shape)
+ bs = self.readout(bs)
+ return bs
+
+ def partial_forward(self, bs, start_layer=None, end_layer=None):
+ if start_layer is None:
+ # 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)
+ if end_layer is not None:
+ return self.trunk[:end_layer](bs)