Fix sampling bug when logits_all=False

Author: abetlen

llama_cpp/llama.py

@@ -1029,16 +1029,16 @@ def eval(self, tokens: Sequence[int]):
             )
             self._ctx.decode(self._batch)
             # Save tokens
-            self.input_ids[self.n_tokens : self.n_tokens + n_tokens] = batch
+            self.input_ids[n_past : n_past + n_tokens] = batch
             # Save logits
-            rows = n_tokens if self.context_params.logits_all else 1
+            rows = n_tokens
             cols = self._n_vocab
             offset = (
                 0 if self.context_params.logits_all else n_tokens - 1
             )  # NOTE: Only save the last token logits if logits_all is False
-            self.scores[self.n_tokens + offset : self.n_tokens + n_tokens, :].reshape(
+            self.scores[n_past + offset : n_past + n_tokens, :].reshape(
                 -1
-            )[:] = self._ctx.get_logits()[: rows * cols]
+            )[:] = self._ctx.get_logits()[offset * cols: rows * cols]
             # Update n_tokens
             self.n_tokens += n_tokens