vllm.py

# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import json
import logging
from dataclasses import field
from typing import Any, AsyncGenerator, List, Optional

import numpy as np

from triton_distributed.icp.data_plane import DataPlane
from triton_distributed.icp.request_plane import RequestPlane
from triton_distributed.runtime import (
    Operator,
    RemoteInferenceRequest,
    RemoteInferenceResponse,
    RemoteOperator,
)

from .stages import AggregatedStage, GenerateStage, PrefillStage, Stage


class VllmOperator(Operator):
    def __init__(
        self,
        name: str,
        version: int,
        request_plane: RequestPlane,
        data_plane: DataPlane,
        parameters: Optional[dict[str, str | int | bool | bytes]] = field(
            default_factory=dict
        ),
        repository: Optional[str] = None,
        logger: Optional[logging.Logger] = None,
        triton_core: Optional[Any] = None,
    ):
        self.name = name
        self.version = version
        self.request_plane = request_plane
        self.data_plane = data_plane
        if logger is None:
            self.logger = logging.getLogger(__name__)
        else:
            self.logger = logger
        self._stage: Stage

        self._init_stages(parameters)

    async def execute(self, requests: List[RemoteInferenceRequest]) -> None:
        for request in requests:
            response_sender = request.response_sender()
            try:
                inputs, parameters = self._prepare_inputs(request)
                self.logger.debug("Processing request")
                async for response in self._stage(
                    {
                        "inputs": inputs,
                        "parameters": parameters,
                    }
                ):
                    self.logger.debug("Sending response")
                    await response_sender.send(**response)
                    self.logger.debug("Response send")
            except Exception as e:
                self.logger.error(f"Error processing request: {e}")
                await response_sender.send(error=e, final=True)

    def _init_stages(
        self,
        parameters: Optional[dict[str, str | int | bool | bytes]] = field(
            default_factory=dict
        ),
    ):
        args = argparse.Namespace(**parameters)  # type: ignore
        self._stage = AggregatedStage(
            model=args.model_name,
            tensor_parallel_size=args.baseline_tp_size,
            gpu_memory_utilization=args.gpu_memory_utilization,
            max_model_len=args.max_model_len,
            dtype=args.dtype,
            kv_cache_dtype=args.kv_cache_dtype,
            enable_prefix_caching=args.enable_prefix_caching,
            enable_chunked_prefill=args.enable_chunked_prefill,
            enforce_eager=args.enforce_eager,
            ignore_eos=args.ignore_eos,
            max_num_seqs=args.max_num_seqs,
            disable_async_output_proc=args.disable_async_output_proc,
            disable_log_stats=args.disable_log_stats,
        )

    @staticmethod
    def _prepare_inputs(request: RemoteInferenceRequest):
        inputs, parameters = {}, {}
        for input_name, input_data in request.inputs.items():
            inputs[input_name] = np.from_dlpack(input_data)
        for key, value in request.parameters.items():
            if isinstance(value, str) and value.startswith("JSON:"):
                parameters[key] = json.loads(value[5:])
            else:
                parameters[key] = value
        return inputs, parameters


class VllmContextOperator(VllmOperator):
    def _init_stages(
        self,
        parameters: Optional[dict[str, str | int | bool | bytes]] = field(
            default_factory=dict
        ),
    ):
        args = argparse.Namespace(**parameters)  # type: ignore
        self._prefill_stage = PrefillStage(
            model=args.model_name,
            tensor_parallel_size=args.context_tp_size,
            generate_tensor_parallel_size=args.generate_tp_size,
            gpu_memory_utilization=args.gpu_memory_utilization,
            max_model_len=args.max_model_len,
            dtype=args.dtype,
            kv_cache_dtype=args.kv_cache_dtype,
            enable_prefix_caching=args.enable_prefix_caching,
            enable_chunked_prefill=args.enable_chunked_prefill,
            enforce_eager=args.enforce_eager,
            ignore_eos=args.ignore_eos,
            max_num_seqs=args.max_num_seqs,
            disable_async_output_proc=args.disable_async_output_proc,
            disable_log_stats=args.disable_log_stats,
        )
        self._generate_operator = RemoteOperator(
            "generate", self.request_plane, self.data_plane
        )

    async def execute(self, requests: List[RemoteInferenceRequest]) -> None:
        for request in requests:
            response_sender = request.response_sender()
            try:
                self.logger.info("Processing request")
                inputs, parameters = self._prepare_inputs(request)
                responses = [
                    response
                    async for response in self._prefill_stage(
                        {
                            "inputs": inputs,
                            "parameters": parameters,
                        }
                    )
                ]
                self.logger.info("Prefill finished")
                assert len(responses) == 1
                response = responses[0]
                self.logger.info("Processing generate")
                generate_responses: AsyncGenerator[
                    RemoteInferenceResponse, None
                ] = await self._generate_operator.async_infer(
                    inputs=response["outputs"],
                    parameters={**request.parameters, **response["parameters"]},
                )
                async for generate_response in generate_responses:
                    self.logger.info("Sending response")
                    parameters = {"text": generate_response.parameters["text"]}
                    await response_sender.send(
                        outputs=generate_response.outputs,
                        parameters=parameters,
                        final=generate_response.final,
                        error=generate_response.error,
                    )
                    self.logger.info("Response send")
            except Exception as e:
                self.logger.error(f"Error processing request: {e}")
                await response_sender.send(error=e, final=True)


class VllmGenerateOperator(VllmOperator):
    def _init_stages(
        self,
        parameters: Optional[dict[str, str | int | bool | bytes]] = field(
            default_factory=dict
        ),
    ):
        args = argparse.Namespace(**parameters)  # type: ignore
        args.worker_name = "generate"
        self._stage = GenerateStage(
            model=args.model_name,
            tensor_parallel_size=args.generate_tp_size,
            gpu_memory_utilization=args.gpu_memory_utilization,
            max_model_len=args.max_model_len,
            dtype=args.dtype,
            kv_cache_dtype=args.kv_cache_dtype,
            enable_prefix_caching=args.enable_prefix_caching,
            enable_chunked_prefill=args.enable_chunked_prefill,
            enforce_eager=args.enforce_eager,
            ignore_eos=args.ignore_eos,
            max_num_seqs=args.max_num_seqs,
            disable_async_output_proc=args.disable_async_output_proc,
            disable_log_stats=args.disable_log_stats,
        )