Support suffix for fill-in-the-middle

Author: njhill

src/vllm_tgis_adapter/grpc/grpc_server.py

@@ -14,6 +14,7 @@
 from grpc_health.v1 import health, health_pb2, health_pb2_grpc
 from grpc_reflection.v1alpha import reflection
 from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.fim import get_fim_encoder_lookup
 from vllm.entrypoints.openai.serving_completion import merge_async_iterators
 from vllm.inputs import LLMInputs
 from vllm.sampling_params import RequestOutputKind, SamplingParams
@@ -205,6 +206,8 @@ def __init__(
         )
         self.health_servicer = health_servicer
 
+        self.get_fim_encoder = get_fim_encoder_lookup(args.fim)
+
     async def post_init(self) -> None:
         self.config = await self.engine.get_model_config()
 
@@ -250,7 +253,12 @@ async def Generate(
 
         for i, req in enumerate(request.requests):
             input_ids, max_is_token_limit[i] = await self._validate_prompt_and_tokenize(
-                sampling_params, truncate_input_tokens, req.text, tokenizer, context
+                sampling_params,
+                truncate_input_tokens,
+                req.text,
+                req.suffix,
+                tokenizer,
+                context,
             )
 
             inputs = LLMInputs(
@@ -348,6 +356,7 @@ async def GenerateStream(  # noqa: PLR0915, C901
             sampling_params,
             truncate_input_tokens,
             request.request.text,
+            request.request.suffix,
             tokenizer,
             context,
         )
@@ -778,30 +787,47 @@ def _convert_tokens(  # noqa: PLR0913
                 )
             token_infos.append(token_info)
 
-    async def _validate_prompt_and_tokenize(
+    async def _validate_prompt_and_tokenize(  # noqa: PLR0913
         self,
         sampling_params: SamplingParams,
         truncate_input_tokens: int | None,
         prompt: str,
+        suffix: str | None,
         tokenizer: AnyTokenizer,
         context: ServicerContext,
     ) -> tuple[list[int], bool]:
         assert self.config is not None
 
-        max_model_len = self.config.max_model_len
-
-        tokenizer_kwargs: dict[str, Any] = {"add_special_tokens": ADD_SPECIAL_TOKENS}
-        if truncate_input_tokens is not None:
-            tokenizer_kwargs.update(
-                {
-                    "truncation": True,
-                    "max_length": truncate_input_tokens,
-                }
-            )
+        if suffix:
+            if not (get_fim_encoder := self.get_fim_encoder):
+                await context.abort(
+                    StatusCode.INVALID_ARGUMENT,
+                    "fim support must be enabled to use suffix",
+                )
+            if truncate_input_tokens is not None:
+                await context.abort(
+                    StatusCode.INVALID_ARGUMENT,
+                    "truncate_input_tokens cannot be used with suffix",
+                )
+            fim_encoder = get_fim_encoder(tokenizer)
+            input_ids = fim_encoder.encode_with_suffix(prefix=prompt, suffix=suffix)
+        else:
+            tokenizer_kwargs: dict[str, Any] = {
+                "add_special_tokens": ADD_SPECIAL_TOKENS
+            }
+            if truncate_input_tokens is not None:
+                tokenizer_kwargs.update(
+                    {
+                        "truncation": True,
+                        "max_length": truncate_input_tokens,
+                    }
+                )
 
-        input_ids = tokenizer(prompt, **tokenizer_kwargs).input_ids
+            input_ids = tokenizer(prompt, **tokenizer_kwargs).input_ids
         token_num = len(input_ids)
 
+        max_model_len = self.config.max_model_len
+
         try:
             validate_input(sampling_params, token_num, max_model_len)
         except ValueError as tgis_validation_error:

src/vllm_tgis_adapter/grpc/pb/generation.proto

@@ -51,6 +51,8 @@ message BatchedGenerationResponse {
 
 message GenerationRequest {
   string text = 2;
+  /// Optional, for fill-in-middle
+  string suffix = 3;
 }
 
 message GenerationResponse {