circumvent broken llama.cpp pre-tokenizer

outlines/integrations/llamacpp.py

@@ -49,11 +49,19 @@ def __init__(self, model: "Llama"):
         self.special_tokens: Set[int] = set()
 
         self.vocabulary: Dict[str, int] = dict()
-        for t in range(model.n_vocab()):
-            token_piece = model.tokenizer().decode([t])
-            self.vocabulary[token_piece] = t
 
-        self.decode = model.tokenizer().decode
+        tokenizer = model.tokenizer()
+
+        self.decode = tokenizer.decode
+
+        # TODO: Remove when https://github.com/ggerganov/llama.cpp/pull/5613 is resolved
+        try:
+            self.vocabulary = model.tokenizer_.hf_tokenizer.get_vocab()
+        except AttributeError:
+            # ###
+            for t in range(model.n_vocab()):
+                token_piece = model.tokenizer().decode([t])
+                self.vocabulary[token_piece] = t
 
     def convert_token_to_string(self, token: str) -> str:
         return token

outlines/models/llamacpp.py

@@ -1,4 +1,5 @@
 import dataclasses
+import warnings
 from typing import TYPE_CHECKING, Iterator, List, Optional, TypedDict, Union
 
 from typing_extensions import Unpack
@@ -288,6 +289,16 @@ def llamacpp(
     if "verbose" not in llamacpp_model_params:
         llamacpp_model_params["verbose"] = False
 
+    # TODO: Remove when https://github.com/ggerganov/llama.cpp/pull/5613 is resolved
+    if "tokenizer" not in llamacpp_model_params:
+        warnings.warn(
+            "The pre-tokenizer in `llama.cpp` handles unicode improperly "
+            + "(https://github.com/ggerganov/llama.cpp/pull/5613)\n"
+            + "Outlines may raise a `RuntimeError` when building the regex index.\n"
+            + "To circumvent this error when using `models.llamacpp()` you may pass the argument"
+            + "`tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained(<hf_repo_id>)`\n"
+        )
+
     model = Llama.from_pretrained(repo_id, filename, **llamacpp_model_params)
 
     return LlamaCpp(model)

tests/generate/test_integration_llamacpp.py

@@ -247,3 +247,35 @@ def test_llamacpp_cfg(model):
     prompt = "<|im_start|>user\nOutput a short and valid JSON object with two keys.<|im_end|>\n><|im_start|>assistant\n"
     result = generate.cfg(model, grammars.arithmetic)(prompt, seed=11)
     assert isinstance(result, str)
+
+
+@pytest.mark.parametrize(
+    "repo,model_path,hf_tokenizer_uri",
+    [
+        ("Qwen/Qwen1.5-0.5B-Chat-GGUF", "*q2*.gguf", "Qwen/Qwen1.5-0.5B-Chat"),
+        ("TheBloke/phi-2-GGUF", "*Q2*.gguf", "microsoft/phi-2"),
+    ],
+)
+def test_byte_tokenizer_regression(repo, model_path, hf_tokenizer_uri):
+    """Reproduce https://github.com/outlines-dev/outlines/issues/820"""
+    import llama_cpp
+
+    model = models.llamacpp(
+        repo,
+        model_path,
+        tokenizer=llama_cpp.llama_tokenizer.LlamaHFTokenizer.from_pretrained(
+            hf_tokenizer_uri
+        ),
+    )
+    generator = generate.choice(model, ["skirt", "dress", "pen", "jacket"])
+    generator("Pick the odd word out: skirt, dress, pen, jacket")
+
+
+def test_llama_cpp_pre_tokenizer_remains_broken():
+    """If fails, llama.cpp pre-tokenizer is fixed -> revert #892, remove `with pytest.raises`"""
+    repo = "Qwen/Qwen1.5-0.5B-Chat-GGUF"
+    model_path = "*q2*.gguf"
+
+    model = models.llamacpp(repo, model_path)
+    with pytest.raises(RuntimeError):
+        generate.choice(model, ["skirt", "dress", "pen", "jacket"])