vllm-cpu-amxbf16 0.11.2.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/entrypoints/openai/serving_score.py

@@ -0,0 +1,503 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import asyncio
+import time
+from collections.abc import AsyncGenerator, Mapping
+from typing import Any
+
+from fastapi import Request
+
+from vllm.engine.protocol import EngineClient
+from vllm.entrypoints.logger import RequestLogger
+from vllm.entrypoints.openai.protocol import (
+    ErrorResponse,
+    RerankDocument,
+    RerankRequest,
+    RerankResponse,
+    RerankResult,
+    RerankUsage,
+    ScoreRequest,
+    ScoreResponse,
+    ScoreResponseData,
+    UsageInfo,
+)
+from vllm.entrypoints.openai.serving_engine import OpenAIServing
+from vllm.entrypoints.openai.serving_models import OpenAIServingModels
+from vllm.entrypoints.score_utils import (
+    ScoreContentPartParam,
+    ScoreMultiModalParam,
+    _cosine_similarity,
+    _validate_score_input_lens,
+    compress_token_type_ids,
+    get_score_prompt,
+)
+from vllm.entrypoints.utils import _validate_truncation_size
+from vllm.inputs.data import TokensPrompt
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput
+from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
+from vllm.utils.async_utils import make_async, merge_async_iterators
+
+logger = init_logger(__name__)
+
+
+class ServingScores(OpenAIServing):
+    def __init__(
+        self,
+        engine_client: EngineClient,
+        models: OpenAIServingModels,
+        *,
+        request_logger: RequestLogger | None,
+        log_error_stack: bool = False,
+    ) -> None:
+        super().__init__(
+            engine_client=engine_client,
+            models=models,
+            request_logger=request_logger,
+            log_error_stack=log_error_stack,
+        )
+
+    async def _embedding_score(
+        self,
+        tokenizer: AnyTokenizer,
+        texts_1: list[str],
+        texts_2: list[str],
+        request: RerankRequest | ScoreRequest,
+        request_id: str,
+        tokenization_kwargs: dict[str, Any] | None = None,
+        lora_request: LoRARequest | None | None = None,
+        trace_headers: Mapping[str, str] | None = None,
+    ) -> list[PoolingRequestOutput] | ErrorResponse:
+        input_texts = texts_1 + texts_2
+
+        engine_prompts: list[TokensPrompt] = []
+        tokenize_async = make_async(
+            tokenizer.__call__, executor=self._tokenizer_executor
+        )
+
+        tokenization_kwargs = tokenization_kwargs or {}
+        tokenized_prompts = await asyncio.gather(
+            *(tokenize_async(t, **tokenization_kwargs) for t in input_texts)
+        )
+
+        for tok_result, input_text in zip(tokenized_prompts, input_texts):
+            text_token_prompt = self._validate_input(
+                request, tok_result["input_ids"], input_text
+            )
+
+            engine_prompts.append(
+                TokensPrompt(prompt_token_ids=text_token_prompt["prompt_token_ids"])
+            )
+
+        # Schedule the request and get the result generator.
+        generators: list[AsyncGenerator[PoolingRequestOutput, None]] = []
+        pooling_params = request.to_pooling_params()
+
+        try:
+            pooling_params.verify("embed", self.model_config)
+        except ValueError as e:
+            return self.create_error_response(str(e))
+
+        for i, engine_prompt in enumerate(engine_prompts):
+            request_id_item = f"{request_id}-{i}"
+
+            self._log_inputs(
+                request_id_item,
+                input_texts[i],
+                params=pooling_params,
+                lora_request=lora_request,
+            )
+
+            generators.append(
+                self.engine_client.encode(
+                    engine_prompt,
+                    pooling_params,
+                    request_id_item,
+                    lora_request=lora_request,
+                    trace_headers=trace_headers,
+                    priority=request.priority,
+                )
+            )
+
+        result_generator = merge_async_iterators(*generators)
+
+        # Non-streaming response
+        final_res_batch: list[PoolingRequestOutput] = []
+
+        embeddings: list[PoolingRequestOutput | None] = [None] * len(engine_prompts)
+
+        async for i, res in result_generator:
+            embeddings[i] = res
+
+        emb_texts_1: list[PoolingRequestOutput] = []
+        emb_texts_2: list[PoolingRequestOutput] = []
+
+        for i in range(0, len(texts_1)):
+            assert (emb := embeddings[i]) is not None
+            emb_texts_1.append(emb)
+
+        for i in range(len(texts_1), len(embeddings)):
+            assert (emb := embeddings[i]) is not None
+            emb_texts_2.append(emb)
+
+        if len(emb_texts_1) == 1:
+            emb_texts_1 = emb_texts_1 * len(emb_texts_2)
+
+        final_res_batch = _cosine_similarity(
+            tokenizer=tokenizer, embed_1=emb_texts_1, embed_2=emb_texts_2
+        )
+
+        return final_res_batch
+
+    def _preprocess_score(
+        self,
+        request: RerankRequest | ScoreRequest,
+        tokenizer: AnyTokenizer,
+        tokenization_kwargs: dict[str, Any],
+        data_1: str | ScoreContentPartParam,
+        data_2: str | ScoreContentPartParam,
+    ) -> tuple[str, TokensPrompt]:
+        model_config = self.model_config
+
+        full_prompt, engine_prompt = get_score_prompt(
+            model_config=model_config,
+            data_1=data_1,
+            data_2=data_2,
+            tokenizer=tokenizer,
+            tokenization_kwargs=tokenization_kwargs,
+        )
+        self._validate_input(request, engine_prompt["prompt_token_ids"], full_prompt)
+        if request.mm_processor_kwargs is not None:
+            engine_prompt["mm_processor_kwargs"] = request.mm_processor_kwargs
+
+        return full_prompt, engine_prompt
+
+    async def _cross_encoding_score(
+        self,
+        tokenizer: AnyTokenizer,
+        data_1: list[str] | list[ScoreContentPartParam],
+        data_2: list[str] | list[ScoreContentPartParam],
+        request: RerankRequest | ScoreRequest,
+        request_id: str,
+        tokenization_kwargs: dict[str, Any] | None = None,
+        lora_request: LoRARequest | None | None = None,
+        trace_headers: Mapping[str, str] | None = None,
+    ) -> list[PoolingRequestOutput] | ErrorResponse:
+        request_prompts: list[str] = []
+        engine_prompts: list[TokensPrompt] = []
+
+        if len(data_1) == 1:
+            data_1 = data_1 * len(data_2)
+
+        if isinstance(tokenizer, MistralTokenizer):
+            raise ValueError("MistralTokenizer not supported for cross-encoding")
+
+        tokenization_kwargs = tokenization_kwargs or {}
+
+        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]
+
+        preprocess_async = make_async(
+            self._preprocess_score, executor=self._tokenizer_executor
+        )
+
+        preprocessed_prompts = await asyncio.gather(
+            *(
+                preprocess_async(
+                    request=request,
+                    tokenizer=tokenizer,
+                    tokenization_kwargs=tokenization_kwargs,
+                    data_1=t1,
+                    data_2=t2,
+                )
+                for t1, t2 in input_pairs
+            )
+        )
+
+        for full_prompt, engine_prompt in preprocessed_prompts:
+            request_prompts.append(full_prompt)
+            engine_prompts.append(engine_prompt)
+
+        # Schedule the request and get the result generator.
+        generators: list[AsyncGenerator[PoolingRequestOutput, None]] = []
+
+        default_pooling_params = request.to_pooling_params()
+
+        try:
+            default_pooling_params.verify("score", self.model_config)
+        except ValueError as e:
+            return self.create_error_response(str(e))
+
+        for i, engine_prompt in enumerate(engine_prompts):
+            request_id_item = f"{request_id}-{i}"
+
+            self._log_inputs(
+                request_id_item,
+                request_prompts[i],
+                params=default_pooling_params,
+                lora_request=lora_request,
+            )
+
+            if token_type_ids := engine_prompt.pop("token_type_ids", None):
+                pooling_params = default_pooling_params.clone()
+                compressed = compress_token_type_ids(token_type_ids)
+                pooling_params.extra_kwargs = {"compressed_token_type_ids": compressed}
+            else:
+                pooling_params = default_pooling_params
+
+            generator = self.engine_client.encode(
+                engine_prompt,
+                pooling_params,
+                request_id_item,
+                lora_request=lora_request,
+                trace_headers=trace_headers,
+                priority=request.priority,
+            )
+
+            generators.append(generator)
+
+        result_generator = merge_async_iterators(*generators)
+
+        # Non-streaming response
+        final_res_batch: list[PoolingRequestOutput | None] = [None] * len(
+            engine_prompts
+        )
+
+        async for i, res in result_generator:
+            final_res_batch[i] = res
+
+        return [out for out in final_res_batch if out is not None]
+
+    async def _run_scoring(
+        self,
+        data_1: list[str] | str | ScoreMultiModalParam,
+        data_2: list[str] | str | ScoreMultiModalParam,
+        request: ScoreRequest | RerankRequest,
+        request_id: str,
+        raw_request: Request | None = None,
+    ) -> list[PoolingRequestOutput] | ErrorResponse:
+        lora_request = self._maybe_get_adapters(request)
+
+        tokenizer = await self.engine_client.get_tokenizer()
+
+        truncate_prompt_tokens = getattr(request, "truncate_prompt_tokens", None)
+
+        tokenization_kwargs: dict[str, Any] = {}
+        _validate_truncation_size(
+            self.max_model_len, truncate_prompt_tokens, tokenization_kwargs
+        )
+
+        trace_headers = (
+            None
+            if raw_request is None
+            else await self._get_trace_headers(raw_request.headers)
+        )
+
+        if not self.model_config.is_multimodal_model and (
+            isinstance(data_1, dict) or isinstance(data_2, dict)
+        ):
+            raise ValueError(
+                f"MultiModalParam is not supported for {self.model_config.architecture}"  # noqa: E501
+            )
+
+        if isinstance(data_1, str):
+            data_1 = [data_1]
+        elif isinstance(data_1, dict):
+            data_1 = data_1.get("content")  # type: ignore[assignment]
+
+        if isinstance(data_2, str):
+            data_2 = [data_2]
+        elif isinstance(data_2, dict):
+            data_2 = data_2.get("content")  # type: ignore[assignment]
+
+        _validate_score_input_lens(data_1, data_2)  # type: ignore[arg-type]
+
+        if self.model_config.is_cross_encoder:
+            return await self._cross_encoding_score(
+                tokenizer=tokenizer,
+                data_1=data_1,  # type: ignore[arg-type]
+                data_2=data_2,  # type: ignore[arg-type]
+                request=request,
+                request_id=request_id,
+                tokenization_kwargs=tokenization_kwargs,
+                lora_request=lora_request,
+                trace_headers=trace_headers,
+            )
+
+        else:
+            return await self._embedding_score(
+                tokenizer=tokenizer,
+                texts_1=data_1,  # type: ignore[arg-type]
+                texts_2=data_2,  # type: ignore[arg-type]
+                request=request,
+                request_id=request_id,
+                tokenization_kwargs=tokenization_kwargs,
+                lora_request=lora_request,
+                trace_headers=trace_headers,
+            )
+
+    async def create_score(
+        self,
+        request: ScoreRequest,
+        raw_request: Request | None = None,
+    ) -> ScoreResponse | ErrorResponse:
+        """
+        Score API similar to Sentence Transformers cross encoder
+
+        See https://sbert.net/docs/package_reference/cross_encoder
+        """
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        request_id = f"score-{self._base_request_id(raw_request)}"
+        created_time = int(time.time())
+
+        try:
+            final_res_batch = await self._run_scoring(
+                request.text_1,
+                request.text_2,
+                request,
+                request_id,
+                raw_request,
+            )
+            if isinstance(final_res_batch, ErrorResponse):
+                return final_res_batch
+
+            return self.request_output_to_score_response(
+                final_res_batch,
+                request_id,
+                created_time,
+                self.models.model_name(),
+            )
+        except asyncio.CancelledError:
+            return self.create_error_response("Client disconnected")
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            return self.create_error_response(str(e))
+
+    async def do_rerank(
+        self, request: RerankRequest, raw_request: Request | None = None
+    ) -> RerankResponse | ErrorResponse:
+        """
+        Rerank API based on JinaAI's rerank API; implements the same
+        API interface. Designed for compatibility with off-the-shelf
+        tooling, since this is a common standard for reranking APIs
+
+        See example client implementations at
+        https://github.com/infiniflow/ragflow/blob/main/rag/llm/rerank_model.py
+        numerous clients use this standard.
+        """
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        request_id = f"rerank-{self._base_request_id(raw_request)}"
+        documents = request.documents
+        top_n = (
+            request.top_n
+            if request.top_n > 0
+            else (
+                len(documents)
+                if isinstance(documents, list)
+                else len(documents["content"])
+            )
+        )
+
+        try:
+            final_res_batch = await self._run_scoring(
+                request.query,
+                documents,
+                request,
+                request_id,
+                raw_request,
+            )
+            if isinstance(final_res_batch, ErrorResponse):
+                return final_res_batch
+
+            return self.request_output_to_rerank_response(
+                final_res_batch,
+                request_id,
+                self.models.model_name(),
+                documents,
+                top_n,
+            )
+        except asyncio.CancelledError:
+            return self.create_error_response("Client disconnected")
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            return self.create_error_response(str(e))
+
+    def request_output_to_score_response(
+        self,
+        final_res_batch: list[PoolingRequestOutput],
+        request_id: str,
+        created_time: int,
+        model_name: str,
+    ) -> ScoreResponse:
+        items: list[ScoreResponseData] = []
+        num_prompt_tokens = 0
+
+        for idx, final_res in enumerate(final_res_batch):
+            classify_res = ScoringRequestOutput.from_base(final_res)
+
+            item = ScoreResponseData(
+                index=idx,
+                score=classify_res.outputs.score,
+            )
+            prompt_token_ids = final_res.prompt_token_ids
+
+            items.append(item)
+            num_prompt_tokens += len(prompt_token_ids)
+
+        usage = UsageInfo(
+            prompt_tokens=num_prompt_tokens,
+            total_tokens=num_prompt_tokens,
+        )
+
+        return ScoreResponse(
+            id=request_id,
+            created=created_time,
+            model=model_name,
+            data=items,
+            usage=usage,
+        )
+
+    def request_output_to_rerank_response(
+        self,
+        final_res_batch: list[PoolingRequestOutput],
+        request_id: str,
+        model_name: str,
+        documents: list[str] | ScoreMultiModalParam,
+        top_n: int,
+    ) -> RerankResponse:
+        """
+        Convert the output of do_rank to a RerankResponse
+        """
+        results: list[RerankResult] = []
+        num_prompt_tokens = 0
+        for idx, final_res in enumerate(final_res_batch):
+            classify_res = ScoringRequestOutput.from_base(final_res)
+
+            result = RerankResult(
+                index=idx,
+                document=RerankDocument(text=documents[idx])
+                if isinstance(documents, list)
+                else RerankDocument(multi_modal=documents["content"][idx]),
+                relevance_score=classify_res.outputs.score,
+            )
+            results.append(result)
+            prompt_token_ids = final_res.prompt_token_ids
+            num_prompt_tokens += len(prompt_token_ids)
+
+        # sort by relevance, then return the top n if set
+        results.sort(key=lambda x: x.relevance_score, reverse=True)
+        if top_n < len(documents):
+            results = results[:top_n]
+
+        return RerankResponse(
+            id=request_id,
+            model=model_name,
+            results=results,
+            usage=RerankUsage(total_tokens=num_prompt_tokens),
+        )

vllm/entrypoints/openai/serving_tokenization.py

@@ -0,0 +1,203 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from typing import Any, Final
+
+import jinja2
+from fastapi import Request
+
+from vllm.engine.protocol import EngineClient
+from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption
+from vllm.entrypoints.logger import RequestLogger
+from vllm.entrypoints.openai.protocol import (
+    DetokenizeRequest,
+    DetokenizeResponse,
+    ErrorResponse,
+    TokenizeChatRequest,
+    TokenizeRequest,
+    TokenizeResponse,
+    TokenizerInfoResponse,
+)
+from vllm.entrypoints.openai.serving_engine import OpenAIServing
+from vllm.entrypoints.openai.serving_models import OpenAIServingModels
+from vllm.entrypoints.renderer import RenderConfig
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+
+logger = init_logger(__name__)
+
+
+class OpenAIServingTokenization(OpenAIServing):
+    def __init__(
+        self,
+        engine_client: EngineClient,
+        models: OpenAIServingModels,
+        *,
+        request_logger: RequestLogger | None,
+        chat_template: str | None,
+        chat_template_content_format: ChatTemplateContentFormatOption,
+        trust_request_chat_template: bool = False,
+        log_error_stack: bool = False,
+    ) -> None:
+        super().__init__(
+            engine_client=engine_client,
+            models=models,
+            request_logger=request_logger,
+            log_error_stack=log_error_stack,
+        )
+
+        self.chat_template = chat_template
+        self.chat_template_content_format: Final = chat_template_content_format
+        self.trust_request_chat_template = trust_request_chat_template
+
+    async def create_tokenize(
+        self,
+        request: TokenizeRequest,
+        raw_request: Request,
+    ) -> TokenizeResponse | ErrorResponse:
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        request_id = f"tokn-{self._base_request_id(raw_request)}"
+
+        try:
+            lora_request = self._maybe_get_adapters(request)
+
+            tokenizer = await self.engine_client.get_tokenizer()
+            renderer = self._get_renderer(tokenizer)
+
+            if isinstance(request, TokenizeChatRequest):
+                tool_dicts = (
+                    None
+                    if request.tools is None
+                    else [tool.model_dump() for tool in request.tools]
+                )
+                error_check_ret = self._validate_chat_template(
+                    request_chat_template=request.chat_template,
+                    chat_template_kwargs=request.chat_template_kwargs,
+                    trust_request_chat_template=self.trust_request_chat_template,
+                )
+                if error_check_ret is not None:
+                    return error_check_ret
+                (
+                    _,
+                    _,
+                    engine_prompts,
+                ) = await self._preprocess_chat(
+                    request,
+                    tokenizer,
+                    request.messages,
+                    tool_dicts=tool_dicts,
+                    chat_template=request.chat_template or self.chat_template,
+                    chat_template_content_format=self.chat_template_content_format,
+                    add_generation_prompt=request.add_generation_prompt,
+                    continue_final_message=request.continue_final_message,
+                    chat_template_kwargs=request.chat_template_kwargs,
+                    add_special_tokens=request.add_special_tokens,
+                )
+            else:
+                engine_prompts = await renderer.render_prompt(
+                    prompt_or_prompts=request.prompt,
+                    config=self._build_render_config(request),
+                )
+        except (ValueError, TypeError, jinja2.TemplateError) as e:
+            logger.exception("Error in preprocessing prompt inputs")
+            return self.create_error_response(f"{e} {e.__cause__}")
+
+        input_ids: list[int] = []
+        for engine_prompt in engine_prompts:
+            self._log_inputs(
+                request_id, engine_prompt, params=None, lora_request=lora_request
+            )
+
+            if isinstance(engine_prompt, dict) and "prompt_token_ids" in engine_prompt:
+                input_ids.extend(engine_prompt["prompt_token_ids"])
+
+        token_strs = None
+        if request.return_token_strs:
+            token_strs = tokenizer.convert_ids_to_tokens(input_ids)
+
+        return TokenizeResponse(
+            tokens=input_ids,
+            token_strs=token_strs,
+            count=len(input_ids),
+            max_model_len=self.max_model_len,
+        )
+
+    async def create_detokenize(
+        self,
+        request: DetokenizeRequest,
+        raw_request: Request,
+    ) -> DetokenizeResponse | ErrorResponse:
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        request_id = f"tokn-{self._base_request_id(raw_request)}"
+
+        lora_request = self._maybe_get_adapters(request)
+
+        tokenizer = await self.engine_client.get_tokenizer()
+
+        self._log_inputs(
+            request_id, request.tokens, params=None, lora_request=lora_request
+        )
+
+        prompt_input = await self._tokenize_prompt_input_async(
+            request,
+            tokenizer,
+            request.tokens,
+        )
+        input_text = prompt_input["prompt"]
+
+        return DetokenizeResponse(prompt=input_text)
+
+    async def get_tokenizer_info(
+        self,
+    ) -> TokenizerInfoResponse | ErrorResponse:
+        """Get comprehensive tokenizer information."""
+        try:
+            tokenizer = await self.engine_client.get_tokenizer()
+            info = TokenizerInfo(tokenizer, self.chat_template).to_dict()
+            return TokenizerInfoResponse(**info)
+        except Exception as e:
+            return self.create_error_response(f"Failed to get tokenizer info: {str(e)}")
+
+    def _build_render_config(self, request: TokenizeRequest) -> RenderConfig:
+        return RenderConfig(add_special_tokens=request.add_special_tokens)
+
+
+@dataclass
+class TokenizerInfo:
+    tokenizer: AnyTokenizer
+    chat_template: str | None
+
+    def to_dict(self) -> dict[str, Any]:
+        """Return the tokenizer configuration."""
+        return self._get_tokenizer_config()
+
+    def _get_tokenizer_config(self) -> dict[str, Any]:
+        """Get tokenizer configuration directly from the tokenizer object."""
+        config = dict(getattr(self.tokenizer, "init_kwargs", None) or {})
+
+        # Remove file path fields
+        config.pop("vocab_file", None)
+        config.pop("merges_file", None)
+
+        config = self._make_json_serializable(config)
+        config["tokenizer_class"] = type(self.tokenizer).__name__
+        if self.chat_template:
+            config["chat_template"] = self.chat_template
+        return config
+
+    def _make_json_serializable(self, obj):
+        """Convert any non-JSON-serializable objects to serializable format."""
+        if hasattr(obj, "content"):
+            return obj.content
+        elif isinstance(obj, dict):
+            return {k: self._make_json_serializable(v) for k, v in obj.items()}
+        elif isinstance(obj, list):
+            return [self._make_json_serializable(item) for item in obj]
+        else:
+            return obj