Merge pull request #146 from theroyallab/tokenizer_data_fix

Tokenizer data fix

backends/exllamav2/grammar.py

@@ -2,9 +2,13 @@
 from exllamav2 import ExLlamaV2, ExLlamaV2Tokenizer
 from exllamav2.generator.filters import ExLlamaV2Filter, ExLlamaV2PrefixFilter
 from lmformatenforcer import JsonSchemaParser, RegexParser
-from lmformatenforcer.integrations.exllamav2 import ExLlamaV2TokenEnforcerFilter
+from lmformatenforcer.integrations.exllamav2 import (
+    ExLlamaV2TokenEnforcerFilter,
+    build_token_enforcer_tokenizer_data,
+)
 from loguru import logger
 from typing import List
+from functools import lru_cache
 
 
 class OutlinesTokenizerWrapper:
@@ -51,6 +55,18 @@ def next(self):
         return self.fsm.allowed_token_ids(self.state), set()
 
 
+@lru_cache(10)
+def _get_lmfe_tokenizer_data(tokenizer: ExLlamaV2Tokenizer):
+    return build_token_enforcer_tokenizer_data(tokenizer)
+
+
+def clear_grammar_func_cache():
+    """Flush tokenizer_data cache to avoid holding references to
+    tokenizers after unloading a model"""
+
+    _get_lmfe_tokenizer_data.clear_cache()
+
+
 class ExLlamaV2Grammar:
     """ExLlamaV2 class for various grammar filters/parsers."""
 
@@ -82,7 +98,9 @@ def add_json_schema_filter(
         # Allow JSON objects or JSON arrays at the top level
         json_prefixes = ["[", "{"]
 
-        lmfilter = ExLlamaV2TokenEnforcerFilter(schema_parser, tokenizer)
+        lmfilter = ExLlamaV2TokenEnforcerFilter(
+            schema_parser, _get_lmfe_tokenizer_data(tokenizer)
+        )
         prefix_filter = ExLlamaV2PrefixFilter(model, tokenizer, json_prefixes)
 
         # Append the filters
@@ -107,7 +125,9 @@ def add_regex_filter(
 
             return
 
-        lmfilter = ExLlamaV2TokenEnforcerFilter(pattern_parser, tokenizer)
+        lmfilter = ExLlamaV2TokenEnforcerFilter(
+            pattern_parser, _get_lmfe_tokenizer_data(tokenizer)
+        )
 
         # Append the filters
         self.filters.append(lmfilter)

backends/exllamav2/model.py

@@ -26,7 +26,10 @@
 from loguru import logger
 from typing import List, Optional, Union
 
-from backends.exllamav2.grammar import ExLlamaV2Grammar
+from backends.exllamav2.grammar import (
+    ExLlamaV2Grammar,
+    clear_grammar_func_cache,
+)
 from backends.exllamav2.utils import (
     exllama_disabled_flash_attn,
     hardware_supports_flash_attn,
@@ -704,6 +707,10 @@ async def unload(self, loras_only: bool = False, **kwargs):
             # Wait for other jobs to finish
             await self.wait_for_jobs(kwargs.get("skip_wait"))
 
+            # Delete references held in the grammar module
+            clear_grammar_func_cache()
+
+            # Unload LoRAs
             if self.generator and self.generator.generator.current_loras:
                 for lora in self.generator.generator.current_loras:
                     lora.unload()