Fix tests

fairseq/data.py

@@ -442,7 +442,7 @@ def numpy_seed(seed):
 
 def get_dummy_batch(ntokens, src_dict, dst_dict, src_len=128, tgt_len=128):
     bsz = int(ntokens / max(src_len, tgt_len))
-    bsz = (bsz // 8) * 8
+    bsz = math.ceil(bsz / 8) * 8
     assert src_dict.pad() == dst_dict.pad()
     pad_idx = src_dict.pad()
     src_vocab, dst_vocab = len(src_dict), len(dst_dict)

fairseq/dictionary.py

@@ -93,9 +93,10 @@ class Dictionary(object):
                 multiple of 8, which is important on some hardware (e.g., Nvidia
                 Tensor Cores).
         """
+        if nwords == -1:
+            nwords = len(self)
+
         if padding_factor > 1:
-            if nwords == -1:
-                nwords = len(self)
             i = 0
             while nwords % padding_factor != 0:
                 if nwords >= len(self):

scripts/average_checkpoints.py

@@ -44,7 +44,10 @@ def average_checkpoints(inputs):
         for k in params_keys:
             if k not in params_dict:
                 params_dict[k] = []
-            params_dict[k].append(model_params[k].float())
+            p = model_params[k]
+            if isinstance(p, torch.HalfTensor):
+                p = p.float()
+            params_dict[k].append(p)
 
     averaged_params = collections.OrderedDict()
     # v should be a list of torch Tensor.

tests/utils.py

@@ -21,7 +21,7 @@ def dummy_dictionary(vocab_size, prefix='token_'):
     for i in range(vocab_size):
         token = prefix + str(i)
         d.add_symbol(token)
-    d.finalize()
+    d.finalize(padding_factor=1)  # don't add extra padding symbols
     return d