revert tokenizer.get_vocab_size

Signed-off-by: Wallas Santos <[email protected]>
vllm-project · Dec 10, 2024 · 710fcc9 · 710fcc9
1 parent a9a1e3c
commit 710fcc9
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 3 deletions.
diff --git a/vllm/engine/async_llm_engine.py b/vllm/engine/async_llm_engine.py
@@ -551,7 +551,9 @@ async def build_guided_decoding_logits_processor_async(
     guided_decoding.backend = guided_decoding.backend or default_guided_backend
 
     processor = await get_guided_decoding_logits_processor(
-        guided_params=guided_decoding, tokenizer=tokenizer)
+        guided_params=guided_decoding, 
+        tokenizer=tokenizer,
+        model_config=model_config)
 
     if processor:
         if sampling_params.logits_processors is None:

diff --git a/vllm/engine/llm_engine.py b/vllm/engine/llm_engine.py
@@ -2010,7 +2010,9 @@ def _build_logits_processors(
                 self.decoding_config.guided_decoding_backend
 
             processor = get_local_guided_decoding_logits_processor(
-                guided_params=guided_decoding, tokenizer=tokenizer)
+                guided_params=guided_decoding, 
+                tokenizer=tokenizer,
+                model_config=self.model_config)
             if processor:
                 logits_processors.append(processor)
 

diff --git a/vllm/model_executor/guided_decoding/__init__.py b/vllm/model_executor/guided_decoding/__init__.py
@@ -9,6 +9,7 @@
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer
 
+    from vllm.config import ModelConfig
     from vllm.logits_process import LogitsProcessor
     from vllm.sampling_params import GuidedDecodingParams
 
@@ -87,7 +88,8 @@ def maybe_backend_fallback(
 
 async def get_guided_decoding_logits_processor(
         guided_params: GuidedDecodingParams,
-        tokenizer: PreTrainedTokenizer) -> LogitsProcessor | None:
+        tokenizer: PreTrainedTokenizer,
+        model_config: ModelConfig) -> LogitsProcessor | None:
     guided_params = maybe_backend_fallback(guided_params)
     # CFG grammar not supported by LMFE, so we use outlines instead
     if guided_params.backend == 'outlines':

diff --git a/vllm/model_executor/guided_decoding/xgrammar_decoding.py b/vllm/model_executor/guided_decoding/xgrammar_decoding.py
@@ -21,15 +21,18 @@
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer
 
+    from vllm.config import ModelConfig
     from vllm.sampling_params import GuidedDecodingParams
 
 
 # TODO: passing batch size to max threads here
 def get_local_xgrammar_guided_decoding_logits_processor(
         guided_params: GuidedDecodingParams,
         tokenizer: PreTrainedTokenizer,
+        model_config: ModelConfig,
         max_threads: int = 8):
     config = GrammarConfig.from_guided_params(guided_params=guided_params,
+                                              model_config=model_config,
                                               tokenizer=tokenizer,
                                               max_threads=max_threads)
     return XGrammarLogitsProcessor(config)
@@ -142,6 +145,7 @@ class GrammarConfig:
     @classmethod
     def from_guided_params(cls,
                            guided_params: GuidedDecodingParams,
+                           model_config: ModelConfig,
                            tokenizer: PreTrainedTokenizer,
                            max_threads: int = 8) -> GrammarConfig: