Support for guided decoding for offline LLM (#6878)

Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

vllm/model_executor/guided_decoding/__init__.py

@@ -3,9 +3,10 @@ from typing import Optional, Union
 from vllm.entrypoints.openai.protocol import (
     ChatCompletionNamedToolChoiceParam, ChatCompletionRequest,
     CompletionRequest)
-from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (
-    get_lm_format_enforcer_guided_decoding_logits_processor)
+from vllm.model_executor.guided_decoding.guided_fields import (
+    GuidedDecodingRequest)
 from vllm.model_executor.guided_decoding.outlines_decoding import (
+    get_local_outlines_guided_decoding_logits_processor,
     get_outlines_guided_decoding_logits_processor)
 from vllm.sampling_params import LogitsProcessor
 
@@ -20,6 +21,8 @@ async def get_guided_decoding_logits_processor(
         return await get_outlines_guided_decoding_logits_processor(
             request, tokenizer)
     if guided_decoding_backend == 'lm-format-enforcer':
+        from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (  # noqa
+            get_lm_format_enforcer_guided_decoding_logits_processor)
         return await get_lm_format_enforcer_guided_decoding_logits_processor(
             request, tokenizer)
 
@@ -28,6 +31,25 @@ async def get_guided_decoding_logits_processor(
         "Must be one of 'outlines, 'lm-format-enforcer'")
 
 
+def get_local_guided_decoding_logits_processor(
+        guided_decoding_backend: str, guided_options: GuidedDecodingRequest,
+        tokenizer) -> Optional[LogitsProcessor]:
+    # request = _adapt_request_for_tool_use(request)
+
+    if guided_decoding_backend == 'outlines':
+        return get_local_outlines_guided_decoding_logits_processor(
+            guided_options, tokenizer)
+    if guided_decoding_backend == 'lm-format-enforcer':
+        from vllm.model_executor.guided_decoding.lm_format_enforcer_decoding import (  # noqa
+            get_local_lm_format_enforcer_guided_decoding_logits_processor)
+        return get_local_lm_format_enforcer_guided_decoding_logits_processor(
+            guided_options, tokenizer)
+
+    raise ValueError(
+        f"Unknown guided decoding backend '{guided_decoding_backend}'. "
+        "Must be one of 'outlines, 'lm-format-enforcer'")
+
+
 def _adapt_request_for_tool_use(request: Union[CompletionRequest,
                                                ChatCompletionRequest]):
     # the legacy completion API does not support tool use