Add LlamaSequenceGenerator

We currently store the logits processor in the `LlamaCpp` instance. This
causes issues when doing successive generations with different
generators. In this PR we create a new `LlamaSequenceGenerator` instance
every time we create a new generator, and store the logits processor in
this instance which solves the issue.

Fixes #700.

outlines/generate/cfg.py

@@ -3,7 +3,11 @@
 from outlines.fsm.fsm import CFGFSM
 from outlines.generate.api import SequenceGenerator
 from outlines.models import OpenAI
-from outlines.models.llamacpp import CFGLogitsProcessor, LlamaCpp
+from outlines.models.llamacpp import (
+    CFGLogitsProcessor,
+    LlamaCpp,
+    LlamaSequenceGenerator,
+)
 from outlines.samplers import Sampler, multinomial
 
 
@@ -45,9 +49,9 @@ def cfg_llamacpp(
         )
 
     logits_processor = CFGLogitsProcessor(cfg_str, model.tokenizer)
-    model.logits_processor = logits_processor
+    generator = LlamaSequenceGenerator(logits_processor, model)
 
-    return model
+    return generator
 
 
 @cfg.register(OpenAI)

outlines/generate/regex.py

@@ -3,7 +3,11 @@
 from outlines.fsm.fsm import RegexFSM
 from outlines.generate.api import SequenceGenerator
 from outlines.models import OpenAI
-from outlines.models.llamacpp import LlamaCpp, RegexLogitsProcessor
+from outlines.models.llamacpp import (
+    LlamaCpp,
+    LlamaSequenceGenerator,
+    RegexLogitsProcessor,
+)
 from outlines.samplers import Sampler, multinomial
 
 
@@ -49,9 +53,9 @@ def regex_llamacpp(
         )
 
     logits_processor = RegexLogitsProcessor(regex_str, model.tokenizer)
-    model.logits_processor = logits_processor
+    generator = LlamaSequenceGenerator(logits_processor, model)
 
-    return model
+    return generator
 
 
 @regex.register(OpenAI)

outlines/generate/text.py

@@ -3,6 +3,7 @@
 from outlines.fsm.fsm import StopAtEosFSM
 from outlines.generate import SequenceGenerator
 from outlines.models import LlamaCpp, OpenAI
+from outlines.models.llamacpp import LlamaSequenceGenerator
 from outlines.samplers import Sampler, multinomial
 
 
@@ -44,7 +45,9 @@ def text_llamacpp(model: LlamaCpp, sampler: Sampler = multinomial()):
             + "than the multinomial sampler."
         )
 
-    return model
+    generator = LlamaSequenceGenerator(None, model)
+
+    return generator
 
 
 @text.register(OpenAI)

outlines/models/llamacpp.py

@@ -8,17 +8,12 @@
 from outlines.fsm.fsm import CFGFSM, FSM, FSMState, RegexFSM
 
 
-class LlamaCpp:
-    """Represents a `llama_cpp` model."""
-
+class LlamaSequenceGenerator:
     def __init__(
-        self, model_path, logits_processor: Optional["LogitsProcessor"] = None, **kwargs
+        self, logits_processor: Optional["LogitsProcessor"], model: "LlamaCpp"
     ):
-        from llama_cpp import Llama
-
+        self.model = model.model
         self.logits_processor = logits_processor
-        self.model = Llama(model_path, **kwargs)
-        self.tokenizer = LlamaCppTokenizer(self)
 
     def __call__(
         self,
@@ -89,6 +84,16 @@ def stream(
         )
 
 
+class LlamaCpp:
+    """Represents a `llama_cpp` model."""
+
+    def __init__(self, model_path, **kwargs):
+        from llama_cpp import Llama
+
+        self.model = Llama(model_path, **kwargs)
+        self.tokenizer = LlamaCppTokenizer(self)
+
+
 class LlamaCppTokenizer:
     def __init__(self, model, **kwargs):
         self.eos_token_id = model.model.token_eos()