+class Gang(nn.Module):
+ def __init__(self, models, nb_models_for_generation, mode="groupthink"):
+ super().__init__()
+ self.models = models
+ self.nb_models_for_generation = nb_models_for_generation
+ self.mode = mode
+
+ def forward(self, bs):
+ # If first = 0, we are re-starting an auto-regressive process,
+ # that's the right moment to randomize who gonna do it
+ if bs.first == 0:
+ self.models_to_use = [
+ self.models[k]
+ for k in torch.randperm(len(self.models))[
+ : self.nb_models_for_generation
+ ]
+ ]
+
+ all_the_logits = torch.cat(
+ [model(bs).x[None] for model in self.models_to_use], dim=0
+ )
+
+ if self.mode == "groupthink":
+ y = all_the_logits.mean(dim=0)
+ elif self.mode == "groupwork":
+ m = torch.rand(all_the_logits.size(), device=all_the_logits.device)
+ m = (m.sort(dim=0).indices == 0).long()
+ y = (y * m).sum(dim=0)
+ else:
+ raise ValueError(f"Invalid mode {self.mode}")
+
+ return BracketedSequence(y, bs.first, bs.nb)
+
+
+######################################################################
+
+# ar_mask is a tensor with 0s and 1s, of same shape as input, with
+# 1s where tokens should be generated. The others are kept
+# unchanged.
+
+
+def one_batch_masked_inplace_autoregression(
+ model,
+ input,
+ ar_mask,
+ seq_logproba,
+ temperature=1.0,
+ deterministic_synthesis=False,
+ forbidden_tokens=None,
+ forced_biases=None,
+):
+ to_generate = (ar_mask.sum(0) > 0).nonzero()
+
+ if to_generate.min() > 0:
+ model(
+ 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 = model(BracketedSequence(input, s, 1)).x
+
+ logits = output[:, s]
+
+ logits = (logits / temperature).log_softmax(dim=-1)
+
+ 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:
+ dist = torch.distributions.categorical.Categorical(logits=logits)
+ t_next = dist.sample()
+
+ all_n = torch.arange(t_next.size(0))
+ seq_logproba += logits[all_n, t_next].sum(dim=-1)
+
+ input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
+
+