Fix the marian tokenizer importer. (huggingface#2426)

* Fix the marian tokenizer importer.

* Ignore the python caches.

.gitignore

@@ -40,3 +40,4 @@ candle-wasm-examples/*/package-lock.json
 candle-wasm-examples/**/config*.json
 .DS_Store
 .idea/*
+__pycache__

candle-examples/examples/marian-mt/convert_slow_tokenizer.py

@@ -43,6 +43,14 @@ def import_protobuf(error_message=""):
     else:
         raise ImportError(PROTOBUF_IMPORT_ERROR.format(error_message))
 
+def _get_prepend_scheme(add_prefix_space: bool, original_tokenizer) -> str:
+    if add_prefix_space:
+        prepend_scheme = "always"
+        if hasattr(original_tokenizer, "legacy") and not original_tokenizer.legacy:
+            prepend_scheme = "first"
+    else:
+        prepend_scheme = "never"
+    return prepend_scheme
 
 class SentencePieceExtractor:
     """
@@ -519,13 +527,15 @@ def normalizer(self, proto):
             )
 
     def pre_tokenizer(self, replacement, add_prefix_space):
-        return pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space)
+        prepend_scheme = _get_prepend_scheme(add_prefix_space, self.original_tokenizer)
+        return pre_tokenizers.Metaspace(replacement=replacement, prepend_scheme=prepend_scheme)
 
     def post_processor(self):
         return None
 
     def decoder(self, replacement, add_prefix_space):
-        return decoders.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space)
+        prepend_scheme = _get_prepend_scheme(add_prefix_space, self.original_tokenizer)
+        return decoders.Metaspace(replacement=replacement, prepend_scheme=prepend_scheme)
 
     def converted(self) -> Tokenizer:
         tokenizer = self.tokenizer(self.proto)
@@ -636,7 +646,8 @@ def pre_tokenizer(self, replacement, add_prefix_space):
         list_pretokenizers = []
         if self.original_tokenizer.split_by_punct:
             list_pretokenizers.append(pre_tokenizers.Punctuation(behavior="isolated"))
-        list_pretokenizers.append(pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space))
+        prepend_scheme = _get_prepend_scheme(add_prefix_space, self.original_tokenizer)
+        list_pretokenizers.append(pre_tokenizers.Metaspace(replacement=replacement, prepend_scheme=prepend_scheme))
         return pre_tokenizers.Sequence(list_pretokenizers)
 
     def normalizer(self, proto):
@@ -929,10 +940,11 @@ def unk_id(self, proto):
         return proto.trainer_spec.unk_id + self.original_tokenizer.offset
 
     def pre_tokenizer(self, replacement, add_prefix_space):
+        prepend_scheme = _get_prepend_scheme(add_prefix_space, self.original_tokenizer)
         return pre_tokenizers.Sequence(
             [
                 pre_tokenizers.WhitespaceSplit(),
-                pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space),
+                pre_tokenizers.Metaspace(replacement=replacement, prepend_scheme=prepend_scheme),
             ]
         )