Add possessive quantifiers to avoid catastrophic backtracking

Cargo.toml

@@ -12,7 +12,7 @@ crate-type = ["cdylib"]
 pyo3 = { version = "0.20.0", features = ["extension-module"] }
 
 # tiktoken dependencies
-fancy-regex = "0.11.0"
-regex = "1.8.3"
+fancy-regex = "0.13.0"
+regex = "1.10.3"
 rustc-hash = "1.1.0"
 bstr = "1.5.0"

tests/test_encoding.py

@@ -11,6 +11,23 @@
 from .test_helpers import ENCODING_FACTORIES, MAX_EXAMPLES
 
 
+@pytest.mark.skip(reason="Takes a really long time to finish, but was added to reproduce a crash.")
+@pytest.mark.parametrize("make_enc", ENCODING_FACTORIES)
+def test_extremely_big_encoding(make_enc: Callable[[], tiktoken.Encoding]):
+    enc = make_enc()
+    for c in ["^", "0", "a", "'s"]: # TODO " ", "\n" are still failing
+        print(f"Validating `{c}`")
+
+        big_value = c * 1_000_000
+        assert big_value == enc.decode(enc.encode(big_value))
+
+        big_value = " " + big_value
+        assert big_value == enc.decode(enc.encode(big_value))
+
+        big_value = big_value + "\n"
+        assert big_value == enc.decode(enc.encode(big_value))
+
+
 def test_simple():
     enc = tiktoken.get_encoding("gpt2")
     assert enc.encode("hello world") == [31373, 995]

tiktoken_ext/openai_public.py

@@ -6,6 +6,11 @@
 FIM_SUFFIX = "<|fim_suffix|>"
 ENDOFPROMPT = "<|endofprompt|>"
 
+# The pattern in the original GPT-2 release is:
+# r"""'s|'t|'re|'ve|'m|'ll|'d| ?[\p{L}]+| ?[\p{N}]+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
+# This is equivalent, but executes faster:
+_legacy_splitter_regex = r"""'(?:[sdmt]|ll|ve|re)| ?\p{L}++| ?\p{N}++| ?[^\s\p{L}\p{N}]++|\s+(?!\S)|\s++"""
+
 
 def gpt2():
     mergeable_ranks = data_gym_to_mergeable_bpe_ranks(
@@ -17,10 +22,7 @@ def gpt2():
     return {
         "name": "gpt2",
         "explicit_n_vocab": 50257,
-        # The pattern in the original GPT-2 release is:
-        # r"""'s|'t|'re|'ve|'m|'ll|'d| ?[\p{L}]+| ?[\p{N}]+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
-        # This is equivalent, but executes faster:
-        "pat_str": r"""'(?:[sdmt]|ll|ve|re)| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        "pat_str": _legacy_splitter_regex,
         "mergeable_ranks": mergeable_ranks,
         "special_tokens": {ENDOFTEXT: 50256},
     }
@@ -34,7 +36,7 @@ def r50k_base():
     return {
         "name": "r50k_base",
         "explicit_n_vocab": 50257,
-        "pat_str": r"""'(?:[sdmt]|ll|ve|re)| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        "pat_str": _legacy_splitter_regex,
         "mergeable_ranks": mergeable_ranks,
         "special_tokens": {ENDOFTEXT: 50256},
     }
@@ -48,7 +50,7 @@ def p50k_base():
     return {
         "name": "p50k_base",
         "explicit_n_vocab": 50281,
-        "pat_str": r"""'(?:[sdmt]|ll|ve|re)| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        "pat_str": _legacy_splitter_regex,
         "mergeable_ranks": mergeable_ranks,
         "special_tokens": {ENDOFTEXT: 50256},
     }
@@ -62,7 +64,7 @@ def p50k_edit():
     special_tokens = {ENDOFTEXT: 50256, FIM_PREFIX: 50281, FIM_MIDDLE: 50282, FIM_SUFFIX: 50283}
     return {
         "name": "p50k_edit",
-        "pat_str": r"""'(?:[sdmt]|ll|ve|re)| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        "pat_str": _legacy_splitter_regex,
         "mergeable_ranks": mergeable_ranks,
         "special_tokens": special_tokens,
     }
@@ -82,7 +84,7 @@ def cl100k_base():
     }
     return {
         "name": "cl100k_base",
-        "pat_str": r"""'(?i:[sdmt]|ll|ve|re)|[^\r\n\p{L}\p{N}]?+\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]++[\r\n]*|\s*[\r\n]|\s+(?!\S)|\s+""",
+        "pat_str": r"""'(?i:[sdmt]|ll|ve|re)|[^\r\n\p{L}\p{N}]?+\p{L}++|\p{N}{1,3}+| ?[^\s\p{L}\p{N}]++[\r\n]*+|\s*[\r\n]|\s+(?!\S)|\s++""",
         "mergeable_ranks": mergeable_ranks,
         "special_tokens": special_tokens,
     }