vllm-cpu 0.11.0.post2__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/engine/protocol.py

@@ -0,0 +1,333 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+from abc import ABC, abstractmethod
+from typing import Any, AsyncGenerator, Iterable, Mapping, Optional, Union
+
+from vllm.beam_search import BeamSearchSequence, create_sort_beams_key_function
+from vllm.config import ModelConfig, VllmConfig
+from vllm.inputs.data import PromptType, TokensPrompt
+from vllm.inputs.parse import is_explicit_encoder_decoder_prompt
+from vllm.inputs.preprocess import InputPreprocessor
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.outputs import CompletionOutput, PoolingRequestOutput, RequestOutput
+from vllm.plugins.io_processors.interface import IOProcessor
+from vllm.pooling_params import PoolingParams
+from vllm.sampling_params import BeamSearchParams, SamplingParams
+from vllm.tasks import SupportedTask
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.utils import Device, collect_from_async_generator, random_uuid
+
+logger = init_logger(__name__)
+
+
+class EngineClient(ABC):
+    """Protocol class for Clients to Engine"""
+
+    @property
+    @abstractmethod
+    def is_running(self) -> bool:
+        ...
+
+    @property
+    @abstractmethod
+    def is_stopped(self) -> bool:
+        ...
+
+    @property
+    @abstractmethod
+    def errored(self) -> bool:
+        ...
+
+    @property
+    @abstractmethod
+    def dead_error(self) -> BaseException:
+        ...
+
+    @abstractmethod
+    def generate(
+        self,
+        prompt: PromptType,
+        sampling_params: SamplingParams,
+        request_id: str,
+        lora_request: Optional[LoRARequest] = None,
+        trace_headers: Optional[Mapping[str, str]] = None,
+        priority: int = 0,
+    ) -> AsyncGenerator[RequestOutput, None]:
+        """Generate outputs for a request."""
+        ...
+
+    async def beam_search(
+        self,
+        prompt: PromptType,
+        request_id: str,
+        params: BeamSearchParams,
+        lora_request: Optional[LoRARequest] = None,
+    ) -> AsyncGenerator[RequestOutput, None]:
+
+        beam_width = params.beam_width
+        max_tokens = params.max_tokens
+        ignore_eos = params.ignore_eos
+        temperature = params.temperature
+        length_penalty = params.length_penalty
+        include_stop_str_in_output = params.include_stop_str_in_output
+
+        preprocessor = await self.get_input_preprocessor()
+        tokenizer = preprocessor.get_tokenizer()
+        eos_token_id = tokenizer.eos_token_id
+
+        if is_explicit_encoder_decoder_prompt(prompt):
+            raise NotImplementedError
+        else:
+            processed_inputs = preprocessor._prompt_to_llm_inputs(prompt)
+
+        if processed_inputs["type"] == "embeds":
+            raise NotImplementedError
+
+        # This is a workaround to fix multimodal beam search; this is a
+        # bandaid fix for 2 small problems:
+        # 1. Multi_modal_data on the processed_inputs currently resolves to
+        #    `None`.
+        # 2. preprocessing above expands the multimodal placeholders. However,
+        #    this happens again in generation, so the double expansion causes
+        #    a mismatch.
+        # TODO - would be ideal to handle this more gracefully.
+        prompt_token_ids = prompt.get("prompt_token_ids")
+        multi_modal_data = prompt.get("multi_modal_data")
+
+        prompt_text = processed_inputs.get("prompt")
+        mm_processor_kwargs = processed_inputs.get("mm_processor_kwargs")
+
+        tokenized_length = len(prompt_token_ids)
+
+        sort_beams_key = create_sort_beams_key_function(
+            eos_token_id, length_penalty)
+
+        beam_search_params = SamplingParams(
+            logprobs=2 * beam_width,
+            max_tokens=1,
+            temperature=temperature,
+        )
+        all_beams = [
+            BeamSearchSequence(tokens=prompt_token_ids,
+                               cum_logprob=0,
+                               logprobs=[],
+                               multi_modal_data=multi_modal_data,
+                               mm_processor_kwargs=mm_processor_kwargs,
+                               lora_request=lora_request)
+        ]
+        completed = []
+
+        for _ in range(max_tokens):
+            prompts_batch, lora_req_batch = zip(*[(
+                TokensPrompt(prompt_token_ids=beam.tokens,
+                             multi_modal_data=beam.multi_modal_data,
+                             mm_processor_kwargs=beam.mm_processor_kwargs),
+                beam.lora_request,
+            ) for beam in all_beams])
+
+            tasks = []
+
+            request_id = f"beam_search-{random_uuid()}"
+            for i, (individual_prompt,
+                    lora_req) in enumerate(zip(prompts_batch, lora_req_batch)):
+                request_id_item = f"{request_id}-{i}"
+                task = asyncio.create_task(
+                    collect_from_async_generator(
+                        self.generate(individual_prompt,
+                                      beam_search_params,
+                                      request_id_item,
+                                      lora_request=lora_req)))
+                tasks.append(task)
+
+            output = await asyncio.gather(*tasks)
+
+            output = [x[0] for x in output]
+
+            new_beams = []
+            for i, current_beam in enumerate(all_beams):
+                result = output[i]
+
+                if result.outputs[0].logprobs is not None:
+                    logprobs = result.outputs[0].logprobs[0]
+                    for token_id, logprob_obj in logprobs.items():
+                        if token_id == eos_token_id and \
+                            not ignore_eos:
+                            completed.append(
+                                BeamSearchSequence(
+                                    tokens=current_beam.tokens +
+                                    [token_id] if include_stop_str_in_output
+                                    else current_beam.tokens,
+                                    logprobs=current_beam.logprobs +
+                                    [logprobs],
+                                    cum_logprob=current_beam.cum_logprob +
+                                    logprob_obj.logprob,
+                                    finish_reason="stop",
+                                    stop_reason=eos_token_id))
+                        else:
+                            new_beams.append(
+                                BeamSearchSequence(
+                                    tokens=current_beam.tokens + [token_id],
+                                    logprobs=current_beam.logprobs +
+                                    [logprobs],
+                                    lora_request=current_beam.lora_request,
+                                    cum_logprob=current_beam.cum_logprob +
+                                    logprob_obj.logprob,
+                                    multi_modal_data=current_beam.
+                                    multi_modal_data,
+                                    mm_processor_kwargs=current_beam.
+                                    mm_processor_kwargs))
+
+            sorted_beams = sorted(new_beams, key=sort_beams_key, reverse=True)
+            all_beams = sorted_beams[:beam_width]
+
+        completed.extend(all_beams)
+        sorted_completed = sorted(completed, key=sort_beams_key, reverse=True)
+        best_beams = sorted_completed[:beam_width]
+
+        for beam in best_beams:
+            if (beam.tokens[-1] == eos_token_id and not ignore_eos):
+                # Skip the eos token in the text.
+                tokens = beam.tokens[tokenized_length:-1]
+            else:
+                tokens = beam.tokens[tokenized_length:]
+            beam.text = tokenizer.decode(tokens)
+
+        yield RequestOutput(
+            request_id=request_id,
+            prompt=prompt_text,
+            outputs=[
+                CompletionOutput(text=beam.text,
+                                 cumulative_logprob=beam.cum_logprob,
+                                 token_ids=beam.tokens[tokenized_length:],
+                                 index=i,
+                                 logprobs=beam.logprobs,
+                                 finish_reason=beam.finish_reason if
+                                 beam.finish_reason is not None else "length",
+                                 stop_reason=beam.stop_reason)
+                for (i, beam) in enumerate(best_beams)
+            ],
+            finished=True,
+            prompt_token_ids=prompt_token_ids,
+            prompt_logprobs=None)
+
+    @abstractmethod
+    def encode(
+        self,
+        prompt: PromptType,
+        pooling_params: PoolingParams,
+        request_id: str,
+        lora_request: Optional[LoRARequest] = None,
+        trace_headers: Optional[Mapping[str, str]] = None,
+        priority: int = 0,
+        tokenization_kwargs: Optional[dict[str, Any]] = None,
+    ) -> AsyncGenerator[PoolingRequestOutput, None]:
+        """Generate outputs for a request from a pooling model."""
+        ...
+
+    @abstractmethod
+    async def abort(self, request_id: Union[str, Iterable[str]]) -> None:
+        """Abort a request.
+
+        Args:
+            request_id: The unique id of the request,
+                        or an iterable of such ids.
+        """
+        ...
+
+    @abstractmethod
+    async def get_vllm_config(self) -> VllmConfig:
+        """Get the vllm configuration of the vLLM engine."""
+        ...
+
+    @abstractmethod
+    async def get_model_config(self) -> ModelConfig:
+        """Get the model configuration of the vLLM engine."""
+        ...
+
+    @abstractmethod
+    async def get_input_preprocessor(self) -> InputPreprocessor:
+        """Get the input processor of the vLLM engine."""
+        ...
+
+    @abstractmethod
+    async def get_tokenizer(self) -> AnyTokenizer:
+        """Get the tokenizer"""
+        ...
+
+    async def get_io_processor(self) -> IOProcessor:
+        raise NotImplementedError
+
+    @abstractmethod
+    async def is_tracing_enabled(self) -> bool:
+        ...
+
+    @abstractmethod
+    async def do_log_stats(self) -> None:
+        ...
+
+    @abstractmethod
+    async def check_health(self) -> None:
+        """Raise if unhealthy"""
+        ...
+
+    @abstractmethod
+    async def start_profile(self) -> None:
+        """Start profiling the engine"""
+        ...
+
+    @abstractmethod
+    async def stop_profile(self) -> None:
+        """Start profiling the engine"""
+        ...
+
+    @abstractmethod
+    async def reset_mm_cache(self) -> None:
+        """Reset the multi-modal cache"""
+        ...
+
+    @abstractmethod
+    async def reset_prefix_cache(self,
+                                 device: Optional[Device] = None) -> None:
+        """Reset the prefix cache"""
+        ...
+
+    @abstractmethod
+    async def sleep(self, level: int = 1) -> None:
+        """Sleep the engine"""
+        ...
+
+    @abstractmethod
+    async def wake_up(self, tags: Optional[list[str]] = None) -> None:
+        """Wake up the engine"""
+        ...
+
+    @abstractmethod
+    async def is_sleeping(self) -> bool:
+        """Check whether the engine is sleeping"""
+        ...
+
+    @abstractmethod
+    async def add_lora(self, lora_request: LoRARequest) -> bool:
+        """Load a new LoRA adapter into the engine for future requests."""
+        ...
+
+    async def scale_elastic_ep(self,
+                               new_data_parallel_size: int,
+                               drain_timeout: int = 300) -> None:
+        """Scale the engine"""
+        raise NotImplementedError
+
+    async def collective_rpc(self,
+                             method: str,
+                             timeout: Optional[float] = None,
+                             args: tuple = (),
+                             kwargs: Optional[dict] = None):
+        """Perform a collective RPC call to the given path."""
+        raise NotImplementedError
+
+    async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
+        """Get supported tasks"""
+        raise NotImplementedError

vllm/entrypoints/api_server.py

@@ -0,0 +1,178 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+NOTE: This API server is used only for demonstrating usage of AsyncEngine
+and simple performance benchmarks. It is not intended for production use.
+For production use, we recommend using our OpenAI compatible server.
+We are also not going to accept PRs modifying this file, please
+change `vllm/entrypoints/openai/api_server.py` instead.
+"""
+import asyncio
+import json
+import ssl
+from argparse import Namespace
+from collections.abc import AsyncGenerator
+from typing import Any, Optional
+
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse, Response, StreamingResponse
+
+import vllm.envs as envs
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.launcher import serve_http
+from vllm.entrypoints.utils import with_cancellation
+from vllm.logger import init_logger
+from vllm.sampling_params import SamplingParams
+from vllm.usage.usage_lib import UsageContext
+from vllm.utils import FlexibleArgumentParser, random_uuid, set_ulimit
+from vllm.version import __version__ as VLLM_VERSION
+
+logger = init_logger("vllm.entrypoints.api_server")
+
+app = FastAPI()
+engine = None
+
+
+@app.get("/health")
+async def health() -> Response:
+    """Health check."""
+    return Response(status_code=200)
+
+
+@app.post("/generate")
+async def generate(request: Request) -> Response:
+    """Generate completion for the request.
+
+    The request should be a JSON object with the following fields:
+    - prompt: the prompt to use for the generation.
+    - stream: whether to stream the results or not.
+    - other fields: the sampling parameters (See `SamplingParams` for details).
+    """
+    request_dict = await request.json()
+    return await _generate(request_dict, raw_request=request)
+
+
+@with_cancellation
+async def _generate(request_dict: dict, raw_request: Request) -> Response:
+    prompt = request_dict.pop("prompt")
+    stream = request_dict.pop("stream", False)
+    sampling_params = SamplingParams(**request_dict)
+    request_id = random_uuid()
+
+    assert engine is not None
+    results_generator = engine.generate(prompt, sampling_params, request_id)
+
+    # Streaming case
+    async def stream_results() -> AsyncGenerator[bytes, None]:
+        async for request_output in results_generator:
+            prompt = request_output.prompt
+            assert prompt is not None
+            text_outputs = [
+                prompt + output.text for output in request_output.outputs
+            ]
+            ret = {"text": text_outputs}
+            yield (json.dumps(ret) + "\n").encode("utf-8")
+
+    if stream:
+        return StreamingResponse(stream_results())
+
+    # Non-streaming case
+    final_output = None
+    try:
+        async for request_output in results_generator:
+            final_output = request_output
+    except asyncio.CancelledError:
+        return Response(status_code=499)
+
+    assert final_output is not None
+    prompt = final_output.prompt
+    assert prompt is not None
+    text_outputs = [prompt + output.text for output in final_output.outputs]
+    ret = {"text": text_outputs}
+    return JSONResponse(ret)
+
+
+def build_app(args: Namespace) -> FastAPI:
+    global app
+
+    app.root_path = args.root_path
+    return app
+
+
+async def init_app(
+    args: Namespace,
+    llm_engine: Optional[AsyncLLMEngine] = None,
+) -> FastAPI:
+    app = build_app(args)
+
+    global engine
+
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    engine = (llm_engine
+              if llm_engine is not None else AsyncLLMEngine.from_engine_args(
+                  engine_args, usage_context=UsageContext.API_SERVER))
+    app.state.engine_client = engine
+    return app
+
+
+async def run_server(args: Namespace,
+                     llm_engine: Optional[AsyncLLMEngine] = None,
+                     **uvicorn_kwargs: Any) -> None:
+    logger.info("vLLM API server version %s", VLLM_VERSION)
+    logger.info("args: %s", args)
+
+    set_ulimit()
+
+    app = await init_app(args, llm_engine)
+    assert engine is not None
+
+    shutdown_task = await serve_http(
+        app,
+        sock=None,
+        enable_ssl_refresh=args.enable_ssl_refresh,
+        host=args.host,
+        port=args.port,
+        log_level=args.log_level,
+        timeout_keep_alive=envs.VLLM_HTTP_TIMEOUT_KEEP_ALIVE,
+        ssl_keyfile=args.ssl_keyfile,
+        ssl_certfile=args.ssl_certfile,
+        ssl_ca_certs=args.ssl_ca_certs,
+        ssl_cert_reqs=args.ssl_cert_reqs,
+        **uvicorn_kwargs,
+    )
+
+    await shutdown_task
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument("--host", type=str, default=None)
+    parser.add_argument("--port", type=parser.check_port, default=8000)
+    parser.add_argument("--ssl-keyfile", type=str, default=None)
+    parser.add_argument("--ssl-certfile", type=str, default=None)
+    parser.add_argument("--ssl-ca-certs",
+                        type=str,
+                        default=None,
+                        help="The CA certificates file")
+    parser.add_argument(
+        "--enable-ssl-refresh",
+        action="store_true",
+        default=False,
+        help="Refresh SSL Context when SSL certificate files change")
+    parser.add_argument(
+        "--ssl-cert-reqs",
+        type=int,
+        default=int(ssl.CERT_NONE),
+        help="Whether client certificate is required (see stdlib ssl module's)"
+    )
+    parser.add_argument(
+        "--root-path",
+        type=str,
+        default=None,
+        help="FastAPI root_path when app is behind a path based routing proxy")
+    parser.add_argument("--log-level", type=str, default="debug")
+    parser = AsyncEngineArgs.add_cli_args(parser)
+    args = parser.parse_args()
+
+    asyncio.run(run_server(args))