vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/entrypoints/openai/serving_models.py

@@ -0,0 +1,314 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+import pathlib
+from asyncio import Lock
+from collections import defaultdict
+from dataclasses import dataclass
+from http import HTTPStatus
+from typing import Optional, Union
+
+from vllm.config import ModelConfig
+from vllm.engine.protocol import EngineClient
+from vllm.entrypoints.openai.protocol import (ErrorResponse,
+                                              LoadLoRAAdapterRequest,
+                                              ModelCard, ModelList,
+                                              ModelPermission,
+                                              UnloadLoRAAdapterRequest)
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.lora.resolver import LoRAResolver, LoRAResolverRegistry
+from vllm.prompt_adapter.request import PromptAdapterRequest
+from vllm.utils import AtomicCounter
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class BaseModelPath:
+    name: str
+    model_path: str
+
+
+@dataclass
+class PromptAdapterPath:
+    name: str
+    local_path: str
+
+
+@dataclass
+class LoRAModulePath:
+    name: str
+    path: str
+    base_model_name: Optional[str] = None
+
+
+class OpenAIServingModels:
+    """Shared instance to hold data about the loaded base model(s) and adapters.
+
+    Handles the routes:
+    - /v1/models
+    - /v1/load_lora_adapter
+    - /v1/unload_lora_adapter
+    """
+
+    def __init__(
+        self,
+        engine_client: EngineClient,
+        model_config: ModelConfig,
+        base_model_paths: list[BaseModelPath],
+        *,
+        lora_modules: Optional[list[LoRAModulePath]] = None,
+        prompt_adapters: Optional[list[PromptAdapterPath]] = None,
+    ):
+        super().__init__()
+
+        self.base_model_paths = base_model_paths
+        self.max_model_len = model_config.max_model_len
+        self.engine_client = engine_client
+        self.model_config = model_config
+
+        self.static_lora_modules = lora_modules
+        self.lora_requests: list[LoRARequest] = []
+        self.lora_id_counter = AtomicCounter(0)
+
+        self.lora_resolvers: list[LoRAResolver] = []
+        for lora_resolver_name in LoRAResolverRegistry.get_supported_resolvers(
+        ):
+            self.lora_resolvers.append(
+                LoRAResolverRegistry.get_resolver(lora_resolver_name))
+        self.lora_resolver_lock: dict[str, Lock] = defaultdict(Lock)
+
+        self.prompt_adapter_requests = []
+        if prompt_adapters is not None:
+            for i, prompt_adapter in enumerate(prompt_adapters, start=1):
+                with pathlib.Path(prompt_adapter.local_path,
+                                  "adapter_config.json").open() as f:
+                    adapter_config = json.load(f)
+                    num_virtual_tokens = adapter_config["num_virtual_tokens"]
+                self.prompt_adapter_requests.append(
+                    PromptAdapterRequest(
+                        prompt_adapter_name=prompt_adapter.name,
+                        prompt_adapter_id=i,
+                        prompt_adapter_local_path=prompt_adapter.local_path,
+                        prompt_adapter_num_virtual_tokens=num_virtual_tokens))
+
+    async def init_static_loras(self):
+        """Loads all static LoRA modules.
+        Raises if any fail to load"""
+        if self.static_lora_modules is None:
+            return
+        for lora in self.static_lora_modules:
+            load_request = LoadLoRAAdapterRequest(lora_path=lora.path,
+                                                  lora_name=lora.name)
+            load_result = await self.load_lora_adapter(
+                request=load_request, base_model_name=lora.base_model_name)
+            if isinstance(load_result, ErrorResponse):
+                raise ValueError(load_result.message)
+
+    def is_base_model(self, model_name) -> bool:
+        return any(model.name == model_name for model in self.base_model_paths)
+
+    def model_name(self, lora_request: Optional[LoRARequest] = None) -> str:
+        """Returns the appropriate model name depending on the availability
+        and support of the LoRA or base model.
+        Parameters:
+        - lora: LoRARequest that contain a base_model_name.
+        Returns:
+        - str: The name of the base model or the first available model path.
+        """
+        if lora_request is not None:
+            return lora_request.lora_name
+        return self.base_model_paths[0].name
+
+    async def show_available_models(self) -> ModelList:
+        """Show available models. This includes the base model and all 
+        adapters"""
+        model_cards = [
+            ModelCard(id=base_model.name,
+                      max_model_len=self.max_model_len,
+                      root=base_model.model_path,
+                      permission=[ModelPermission()])
+            for base_model in self.base_model_paths
+        ]
+        lora_cards = [
+            ModelCard(id=lora.lora_name,
+                      root=lora.local_path,
+                      parent=lora.base_model_name if lora.base_model_name else
+                      self.base_model_paths[0].name,
+                      permission=[ModelPermission()])
+            for lora in self.lora_requests
+        ]
+        prompt_adapter_cards = [
+            ModelCard(id=prompt_adapter.prompt_adapter_name,
+                      root=self.base_model_paths[0].name,
+                      permission=[ModelPermission()])
+            for prompt_adapter in self.prompt_adapter_requests
+        ]
+        model_cards.extend(lora_cards)
+        model_cards.extend(prompt_adapter_cards)
+        return ModelList(data=model_cards)
+
+    async def load_lora_adapter(
+            self,
+            request: LoadLoRAAdapterRequest,
+            base_model_name: Optional[str] = None
+    ) -> Union[ErrorResponse, str]:
+        error_check_ret = await self._check_load_lora_adapter_request(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        lora_name, lora_path = request.lora_name, request.lora_path
+        unique_id = self.lora_id_counter.inc(1)
+        lora_request = LoRARequest(lora_name=lora_name,
+                                   lora_int_id=unique_id,
+                                   lora_path=lora_path)
+        if base_model_name is not None and self.is_base_model(base_model_name):
+            lora_request.base_model_name = base_model_name
+
+        # Validate that the adapter can be loaded into the engine
+        # This will also pre-load it for incoming requests
+        try:
+            await self.engine_client.add_lora(lora_request)
+        except BaseException as e:
+            error_type = "BadRequestError"
+            status_code = HTTPStatus.BAD_REQUEST
+            if "No adapter found" in str(e):
+                error_type = "NotFoundError"
+                status_code = HTTPStatus.NOT_FOUND
+
+            return create_error_response(message=str(e),
+                                         err_type=error_type,
+                                         status_code=status_code)
+
+        self.lora_requests.append(lora_request)
+        logger.info("Loaded new LoRA adapter: name '%s', path '%s'", lora_name,
+                    lora_path)
+        return f"Success: LoRA adapter '{lora_name}' added successfully."
+
+    async def unload_lora_adapter(
+            self,
+            request: UnloadLoRAAdapterRequest) -> Union[ErrorResponse, str]:
+        error_check_ret = await self._check_unload_lora_adapter_request(request
+                                                                        )
+        if error_check_ret is not None:
+            return error_check_ret
+
+        lora_name = request.lora_name
+        self.lora_requests = [
+            lora_request for lora_request in self.lora_requests
+            if lora_request.lora_name != lora_name
+        ]
+        logger.info("Removed LoRA adapter: name '%s'", lora_name)
+        return f"Success: LoRA adapter '{lora_name}' removed successfully."
+
+    async def _check_load_lora_adapter_request(
+            self, request: LoadLoRAAdapterRequest) -> Optional[ErrorResponse]:
+        # Check if both 'lora_name' and 'lora_path' are provided
+        if not request.lora_name or not request.lora_path:
+            return create_error_response(
+                message="Both 'lora_name' and 'lora_path' must be provided.",
+                err_type="InvalidUserInput",
+                status_code=HTTPStatus.BAD_REQUEST)
+
+        # Check if the lora adapter with the given name already exists
+        if any(lora_request.lora_name == request.lora_name
+               for lora_request in self.lora_requests):
+            return create_error_response(
+                message=
+                f"The lora adapter '{request.lora_name}' has already been "
+                "loaded.",
+                err_type="InvalidUserInput",
+                status_code=HTTPStatus.BAD_REQUEST)
+
+        return None
+
+    async def _check_unload_lora_adapter_request(
+            self,
+            request: UnloadLoRAAdapterRequest) -> Optional[ErrorResponse]:
+        # Check if either 'lora_name' or 'lora_int_id' is provided
+        if not request.lora_name and not request.lora_int_id:
+            return create_error_response(
+                message=
+                "either 'lora_name' and 'lora_int_id' needs to be provided.",
+                err_type="InvalidUserInput",
+                status_code=HTTPStatus.BAD_REQUEST)
+
+        # Check if the lora adapter with the given name exists
+        if not any(lora_request.lora_name == request.lora_name
+                   for lora_request in self.lora_requests):
+            return create_error_response(
+                message=
+                f"The lora adapter '{request.lora_name}' cannot be found.",
+                err_type="NotFoundError",
+                status_code=HTTPStatus.NOT_FOUND)
+
+        return None
+
+    async def resolve_lora(
+            self, lora_name: str) -> Union[LoRARequest, ErrorResponse]:
+        """Attempt to resolve a LoRA adapter using available resolvers.
+
+        Args:
+            lora_name: Name/identifier of the LoRA adapter
+
+        Returns:
+            LoRARequest if found and loaded successfully.
+            ErrorResponse (404) if no resolver finds the adapter.
+            ErrorResponse (400) if adapter(s) are found but none load.
+        """
+        async with self.lora_resolver_lock[lora_name]:
+            # First check if this LoRA is already loaded
+            for existing in self.lora_requests:
+                if existing.lora_name == lora_name:
+                    return existing
+
+            base_model_name = self.model_config.model
+            unique_id = self.lora_id_counter.inc(1)
+            found_adapter = False
+
+            # Try to resolve using available resolvers
+            for resolver in self.lora_resolvers:
+                lora_request = await resolver.resolve_lora(
+                    base_model_name, lora_name)
+
+                if lora_request is not None:
+                    found_adapter = True
+                    lora_request.lora_int_id = unique_id
+
+                    try:
+                        await self.engine_client.add_lora(lora_request)
+                        self.lora_requests.append(lora_request)
+                        logger.info(
+                            "Resolved and loaded LoRA adapter '%s' using %s",
+                            lora_name, resolver.__class__.__name__)
+                        return lora_request
+                    except BaseException as e:
+                        logger.warning(
+                            "Failed to load LoRA '%s' resolved by %s: %s. "
+                            "Trying next resolver.", lora_name,
+                            resolver.__class__.__name__, e)
+                        continue
+
+            if found_adapter:
+                # An adapter was found, but all attempts to load it failed.
+                return create_error_response(
+                    message=(f"LoRA adapter '{lora_name}' was found "
+                             "but could not be loaded."),
+                    err_type="BadRequestError",
+                    status_code=HTTPStatus.BAD_REQUEST)
+            else:
+                # No adapter was found
+                return create_error_response(
+                    message=f"LoRA adapter {lora_name} does not exist",
+                    err_type="NotFoundError",
+                    status_code=HTTPStatus.NOT_FOUND)
+
+
+def create_error_response(
+        message: str,
+        err_type: str = "BadRequestError",
+        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST) -> ErrorResponse:
+    return ErrorResponse(message=message,
+                         type=err_type,
+                         code=status_code.value)

vllm/entrypoints/openai/serving_pooling.py

@@ -0,0 +1,237 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import asyncio
+import base64
+import time
+from collections.abc import AsyncGenerator
+from typing import Final, Literal, Optional, Union, cast
+
+import jinja2
+import numpy as np
+from fastapi import Request
+from typing_extensions import assert_never
+
+from vllm.config import ModelConfig
+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 (ErrorResponse,
+                                              PoolingChatRequest,
+                                              PoolingRequest, PoolingResponse,
+                                              PoolingResponseData, UsageInfo)
+from vllm.entrypoints.openai.serving_engine import OpenAIServing
+from vllm.entrypoints.openai.serving_models import OpenAIServingModels
+from vllm.logger import init_logger
+from vllm.outputs import PoolingOutput, PoolingRequestOutput
+from vllm.utils import merge_async_iterators
+
+logger = init_logger(__name__)
+
+
+def _get_data(
+    output: PoolingOutput,
+    encoding_format: Literal["float", "base64"],
+) -> Union[list[float], str]:
+    if encoding_format == "float":
+        return output.data.tolist()
+    elif encoding_format == "base64":
+        # Force to use float32 for base64 encoding
+        # to match the OpenAI python client behavior
+        pooling_bytes = np.array(output.data, dtype="float32").tobytes()
+        return base64.b64encode(pooling_bytes).decode("utf-8")
+
+    assert_never(encoding_format)
+
+
+class OpenAIServingPooling(OpenAIServing):
+
+    def __init__(
+        self,
+        engine_client: EngineClient,
+        model_config: ModelConfig,
+        models: OpenAIServingModels,
+        *,
+        request_logger: Optional[RequestLogger],
+        chat_template: Optional[str],
+        chat_template_content_format: ChatTemplateContentFormatOption,
+    ) -> None:
+        super().__init__(engine_client=engine_client,
+                         model_config=model_config,
+                         models=models,
+                         request_logger=request_logger)
+
+        self.chat_template = chat_template
+        self.chat_template_content_format: Final = chat_template_content_format
+
+    async def create_pooling(
+        self,
+        request: PoolingRequest,
+        raw_request: Optional[Request] = None,
+    ) -> Union[PoolingResponse, ErrorResponse]:
+        """
+        See https://platform.openai.com/docs/api-reference/embeddings/create
+        for the API specification. This API mimics the OpenAI Embedding API.
+        """
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        encoding_format = request.encoding_format
+        if request.dimensions is not None:
+            return self.create_error_response(
+                "dimensions is currently not supported")
+
+        model_name = self._get_model_name(request.model)
+        request_id = f"pool-{self._base_request_id(raw_request)}"
+        created_time = int(time.time())
+
+        truncate_prompt_tokens = None
+
+        if request.truncate_prompt_tokens is not None:
+            if request.truncate_prompt_tokens <= self.max_model_len:
+                truncate_prompt_tokens = request.truncate_prompt_tokens
+            else:
+                return self.create_error_response(
+                    "truncate_prompt_tokens value is "
+                    "greater than max_model_len."
+                    " Please, select a smaller truncation size.")
+
+        try:
+            (
+                lora_request,
+                prompt_adapter_request,
+            ) = self._maybe_get_adapters(request)
+
+            tokenizer = await self.engine_client.get_tokenizer(lora_request)
+
+            if prompt_adapter_request is not None:
+                raise NotImplementedError("Prompt adapter is not supported "
+                                          "for pooling models")
+
+            if isinstance(request, PoolingChatRequest):
+                (
+                    _,
+                    request_prompts,
+                    engine_prompts,
+                ) = await self._preprocess_chat(
+                    request,
+                    tokenizer,
+                    request.messages,
+                    chat_template=request.chat_template or self.chat_template,
+                    chat_template_content_format=self.
+                    chat_template_content_format,
+                    # In pooling requests, we are not generating tokens,
+                    # so there is no need to append extra tokens to the input
+                    add_generation_prompt=False,
+                    continue_final_message=False,
+                    truncate_prompt_tokens=truncate_prompt_tokens,
+                    add_special_tokens=request.add_special_tokens,
+                )
+            else:
+                (request_prompts,
+                 engine_prompts) = await self._preprocess_completion(
+                     request,
+                     tokenizer,
+                     request.input,
+                     truncate_prompt_tokens=truncate_prompt_tokens,
+                     add_special_tokens=request.add_special_tokens,
+                 )
+        except (ValueError, TypeError, jinja2.TemplateError) as e:
+            logger.exception("Error in preprocessing prompt inputs")
+            return self.create_error_response(str(e))
+
+        # Schedule the request and get the result generator.
+        generators: list[AsyncGenerator[PoolingRequestOutput, None]] = []
+        try:
+            pooling_params = request.to_pooling_params()
+
+            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=pooling_params,
+                                 lora_request=lora_request,
+                                 prompt_adapter_request=prompt_adapter_request)
+
+                trace_headers = (None if raw_request is None else await
+                                 self._get_trace_headers(raw_request.headers))
+
+                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)
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            return self.create_error_response(str(e))
+
+        result_generator = merge_async_iterators(*generators)
+
+        num_prompts = len(engine_prompts)
+
+        # Non-streaming response
+        final_res_batch: list[Optional[PoolingRequestOutput]]
+        final_res_batch = [None] * num_prompts
+        try:
+            async for i, res in result_generator:
+                final_res_batch[i] = res
+
+            assert all(final_res is not None for final_res in final_res_batch)
+
+            final_res_batch_checked = cast(list[PoolingRequestOutput],
+                                           final_res_batch)
+
+            response = self.request_output_to_pooling_response(
+                final_res_batch_checked,
+                request_id,
+                created_time,
+                model_name,
+                encoding_format,
+            )
+        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))
+
+        return response
+
+    def request_output_to_pooling_response(
+        self,
+        final_res_batch: list[PoolingRequestOutput],
+        request_id: str,
+        created_time: int,
+        model_name: str,
+        encoding_format: Literal["float", "base64"],
+    ) -> PoolingResponse:
+        items: list[PoolingResponseData] = []
+        num_prompt_tokens = 0
+
+        for idx, final_res in enumerate(final_res_batch):
+            item = PoolingResponseData(
+                index=idx,
+                data=_get_data(final_res.outputs, encoding_format),
+            )
+            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 PoolingResponse(
+            id=request_id,
+            created=created_time,
+            model=model_name,
+            data=items,
+            usage=usage,
+        )