
llmperf
LLMPerf is a library for validating and benchmarking LLMs
Stars: 439

LLMPerf is a tool designed for evaluating the performance of Language Model APIs. It provides functionalities for conducting load tests to measure inter-token latency and generation throughput, as well as correctness tests to verify the responses. The tool supports various LLM APIs including OpenAI, Anthropic, TogetherAI, Hugging Face, LiteLLM, Vertex AI, and SageMaker. Users can set different parameters for the tests and analyze the results to assess the performance of the LLM APIs. LLMPerf aims to standardize prompts across different APIs and provide consistent evaluation metrics for comparison.
README:
A Tool for evaulation the performance of LLM APIs.
git clone https://github.com/ray-project/llmperf.git
cd llmperf
pip install -e .
We implement 2 tests for evaluating LLMs: a load test to check for performance and a correctness test to check for correctness.
The load test spawns a number of concurrent requests to the LLM API and measures the inter-token latency and generation throughput per request and across concurrent requests. The prompt that is sent with each request is of the format:
Randomly stream lines from the following text. Don't generate eos tokens:
LINE 1,
LINE 2,
LINE 3,
...
Where the lines are randomly sampled from a collection of lines from Shakespeare sonnets. Tokens are counted using the LlamaTokenizer
regardless of which LLM API is being tested. This is to ensure that the prompts are consistent across different LLM APIs.
To run the most basic load test you can the token_benchmark_ray script.
- The endpoints provider backend might vary widely, so this is not a reflection on how the software runs on a particular hardware.
- The results may vary with time of day.
- The results may vary with the load.
- The results may not correlate with users’ workloads.
export OPENAI_API_KEY=secret_abcdefg
export OPENAI_API_BASE="https://api.endpoints.anyscale.com/v1"
python token_benchmark_ray.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api openai \
--additional-sampling-params '{}'
export ANTHROPIC_API_KEY=secret_abcdefg
python token_benchmark_ray.py \
--model "claude-2" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api anthropic \
--additional-sampling-params '{}'
export TOGETHERAI_API_KEY="YOUR_TOGETHER_KEY"
python token_benchmark_ray.py \
--model "together_ai/togethercomputer/CodeLlama-7b-Instruct" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api "litellm" \
--additional-sampling-params '{}'
export HUGGINGFACE_API_KEY="YOUR_HUGGINGFACE_API_KEY"
export HUGGINGFACE_API_BASE="YOUR_HUGGINGFACE_API_ENDPOINT"
python token_benchmark_ray.py \
--model "huggingface/meta-llama/Llama-2-7b-chat-hf" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api "litellm" \
--additional-sampling-params '{}'
LLMPerf can use LiteLLM to send prompts to LLM APIs. To see the environment variables to set for the provider and arguments that one should set for model and additional-sampling-params.
see the LiteLLM Provider Documentation.
python token_benchmark_ray.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api "litellm" \
--additional-sampling-params '{}'
Here, --model is used for logging, not for selecting the model. The model is specified in the Vertex AI Endpoint ID.
The GCLOUD_ACCESS_TOKEN needs to be somewhat regularly set, as the token generated by gcloud auth print-access-token
expires after 15 minutes or so.
Vertex AI doesn't return the total number of tokens that are generated by their endpoint, so tokens are counted using the LLama tokenizer.
gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID
export GCLOUD_ACCESS_TOKEN=$(gcloud auth print-access-token)
export GCLOUD_PROJECT_ID=YOUR_PROJECT_ID
export GCLOUD_REGION=YOUR_REGION
export VERTEXAI_ENDPOINT_ID=YOUR_ENDPOINT_ID
python token_benchmark_ray.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--mean-input-tokens 550 \
--stddev-input-tokens 150 \
--mean-output-tokens 150 \
--stddev-output-tokens 10 \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
--llm-api "vertexai" \
--additional-sampling-params '{}'
SageMaker doesn't return the total number of tokens that are generated by their endpoint, so tokens are counted using the LLama tokenizer.
export AWS_ACCESS_KEY_ID="YOUR_ACCESS_KEY_ID"
export AWS_SECRET_ACCESS_KEY="YOUR_SECRET_ACCESS_KEY"s
export AWS_SESSION_TOKEN="YOUR_SESSION_TOKEN"
export AWS_REGION_NAME="YOUR_ENDPOINTS_REGION_NAME"
python llm_correctness.py \
--model "llama-2-7b" \
--llm-api "sagemaker" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
see python token_benchmark_ray.py --help
for more details on the arguments.
The correctness test spawns a number of concurrent requests to the LLM API with the following format:
Convert the following sequence of words into a number: {random_number_in_word_format}. Output just your final answer.
where random_number_in_word_format could be for example "one hundred and twenty three". The test then checks that the response contains that number in digit format which in this case would be 123.
The test does this for a number of randomly generated numbers and reports the number of responses that contain a mismatch.
To run the most basic correctness test you can run the the llm_correctness.py script.
export OPENAI_API_KEY=secret_abcdefg
export OPENAI_API_BASE=https://console.endpoints.anyscale.com/m/v1
python llm_correctness.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--max-num-completed-requests 150 \
--timeout 600 \
--num-concurrent-requests 10 \
--results-dir "result_outputs"
export ANTHROPIC_API_KEY=secret_abcdefg
python llm_correctness.py \
--model "claude-2" \
--llm-api "anthropic" \
--max-num-completed-requests 5 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs"
export TOGETHERAI_API_KEY="YOUR_TOGETHER_KEY"
python llm_correctness.py \
--model "together_ai/togethercomputer/CodeLlama-7b-Instruct" \
--llm-api "litellm" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
export HUGGINGFACE_API_KEY="YOUR_HUGGINGFACE_API_KEY"
export HUGGINGFACE_API_BASE="YOUR_HUGGINGFACE_API_ENDPOINT"
python llm_correctness.py \
--model "huggingface/meta-llama/Llama-2-7b-chat-hf" \
--llm-api "litellm" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
LLMPerf can use LiteLLM to send prompts to LLM APIs. To see the environment variables to set for the provider and arguments that one should set for model and additional-sampling-params.
see the LiteLLM Provider Documentation.
python llm_correctness.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--llm-api "litellm" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
see python llm_correctness.py --help
for more details on the arguments.
Here, --model is used for logging, not for selecting the model. The model is specified in the Vertex AI Endpoint ID.
The GCLOUD_ACCESS_TOKEN needs to be somewhat regularly set, as the token generated by gcloud auth print-access-token
expires after 15 minutes or so.
Vertex AI doesn't return the total number of tokens that are generated by their endpoint, so tokens are counted using the LLama tokenizer.
gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID
export GCLOUD_ACCESS_TOKEN=$(gcloud auth print-access-token)
export GCLOUD_PROJECT_ID=YOUR_PROJECT_ID
export GCLOUD_REGION=YOUR_REGION
export VERTEXAI_ENDPOINT_ID=YOUR_ENDPOINT_ID
python llm_correctness.py \
--model "meta-llama/Llama-2-7b-chat-hf" \
--llm-api "vertexai" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
SageMaker doesn't return the total number of tokens that are generated by their endpoint, so tokens are counted using the LLama tokenizer.
export AWS_ACCESS_KEY_ID="YOUR_ACCESS_KEY_ID"
export AWS_SECRET_ACCESS_KEY="YOUR_SECRET_ACCESS_KEY"s
export AWS_SESSION_TOKEN="YOUR_SESSION_TOKEN"
export AWS_REGION_NAME="YOUR_ENDPOINTS_REGION_NAME"
python llm_correctness.py \
--model "llama-2-7b" \
--llm-api "sagemaker" \
--max-num-completed-requests 2 \
--timeout 600 \
--num-concurrent-requests 1 \
--results-dir "result_outputs" \
The results of the load test and correctness test are saved in the results directory specified by the --results-dir
argument. The results are saved in 2 files, one with the summary metrics of the test, and one with metrics from each individual request that is returned.
The correctness tests were implemented with the following workflow in mind:
import ray
from transformers import LlamaTokenizerFast
from llmperf.ray_clients.openai_chat_completions_client import (
OpenAIChatCompletionsClient,
)
from llmperf.models import RequestConfig
from llmperf.requests_launcher import RequestsLauncher
# Copying the environment variables and passing them to ray.init() is necessary
# For making any clients work.
ray.init(runtime_env={"env_vars": {"OPENAI_API_BASE" : "https://api.endpoints.anyscale.com/v1",
"OPENAI_API_KEY" : "YOUR_API_KEY"}})
base_prompt = "hello_world"
tokenizer = LlamaTokenizerFast.from_pretrained(
"hf-internal-testing/llama-tokenizer"
)
base_prompt_len = len(tokenizer.encode(base_prompt))
prompt = (base_prompt, base_prompt_len)
# Create a client for spawning requests
clients = [OpenAIChatCompletionsClient.remote()]
req_launcher = RequestsLauncher(clients)
req_config = RequestConfig(
model="meta-llama/Llama-2-7b-chat-hf",
prompt=prompt
)
req_launcher.launch_requests(req_config)
result = req_launcher.get_next_ready(block=True)
print(result)
To implement a new LLM client, you need to implement the base class llmperf.ray_llm_client.LLMClient
and decorate it as a ray actor.
from llmperf.ray_llm_client import LLMClient
import ray
@ray.remote
class CustomLLMClient(LLMClient):
def llm_request(self, request_config: RequestConfig) -> Tuple[Metrics, str, RequestConfig]:
"""Make a single completion request to a LLM API
Returns:
Metrics about the performance charateristics of the request.
The text generated by the request to the LLM API.
The request_config used to make the request. This is mainly for logging purposes.
"""
...
The old LLMPerf code base can be found in the llmperf-legacy repo.
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for llmperf
Similar Open Source Tools

llmperf
LLMPerf is a tool designed for evaluating the performance of Language Model APIs. It provides functionalities for conducting load tests to measure inter-token latency and generation throughput, as well as correctness tests to verify the responses. The tool supports various LLM APIs including OpenAI, Anthropic, TogetherAI, Hugging Face, LiteLLM, Vertex AI, and SageMaker. Users can set different parameters for the tests and analyze the results to assess the performance of the LLM APIs. LLMPerf aims to standardize prompts across different APIs and provide consistent evaluation metrics for comparison.

marqo
Marqo is more than a vector database, it's an end-to-end vector search engine for both text and images. Vector generation, storage and retrieval are handled out of the box through a single API. No need to bring your own embeddings.

SPAG
This repository contains the implementation of Self-Play of Adversarial Language Game (SPAG) as described in the paper 'Self-playing Adversarial Language Game Enhances LLM Reasoning'. The SPAG involves training Language Models (LLMs) in an adversarial language game called Adversarial Taboo. The repository provides tools for imitation learning, self-play episode collection, and reinforcement learning on game episodes to enhance LLM reasoning abilities. The process involves training models using GPUs, launching imitation learning, conducting self-play episodes, assigning rewards based on outcomes, and learning the SPAG model through reinforcement learning. Continuous improvements on reasoning benchmarks can be observed by repeating the episode-collection and SPAG-learning processes.

llm2vec
LLM2Vec is a simple recipe to convert decoder-only LLMs into text encoders. It consists of 3 simple steps: 1) enabling bidirectional attention, 2) training with masked next token prediction, and 3) unsupervised contrastive learning. The model can be further fine-tuned to achieve state-of-the-art performance.

elia
Elia is a powerful terminal user interface designed for interacting with large language models. It allows users to chat with models like Claude 3, ChatGPT, Llama 3, Phi 3, Mistral, and Gemma. Conversations are stored locally in a SQLite database, ensuring privacy. Users can run local models through 'ollama' without data leaving their machine. Elia offers easy installation with pipx and supports various environment variables for different models. It provides a quick start to launch chats and manage local models. Configuration options are available to customize default models, system prompts, and add new models. Users can import conversations from ChatGPT and wipe the database when needed. Elia aims to enhance user experience in interacting with language models through a user-friendly interface.

CAG
Cache-Augmented Generation (CAG) is an alternative paradigm to Retrieval-Augmented Generation (RAG) that eliminates real-time retrieval delays and errors by preloading all relevant resources into the model's context. CAG leverages extended context windows of large language models (LLMs) to generate responses directly, providing reduced latency, improved reliability, and simplified design. While CAG has limitations in knowledge size and context length, advancements in LLMs are addressing these issues, making CAG a practical and scalable alternative for complex applications.

langserve_ollama
LangServe Ollama is a tool that allows users to fine-tune Korean language models for local hosting, including RAG. Users can load HuggingFace gguf files, create model chains, and monitor GPU usage. The tool provides a seamless workflow for customizing and deploying language models in a local environment.

mjai.app
mjai.app is a platform for mahjong AI competition. It contains an implementation of a mahjong game simulator for evaluating submission files. The simulator runs Docker internally, and there is a base class for developing bots that communicate via the mjai protocol. Submission files are deployed in a Docker container, and the Docker image is pushed to Docker Hub. The Mjai protocol used is customized based on Mortal's Mjai Engine implementation.

chatgpt-arcana.el
ChatGPT-Arcana is an Emacs package that allows users to interact with ChatGPT directly from Emacs, enabling tasks such as chatting with GPT, operating on code or text, generating eshell commands from natural language, fixing errors, writing commit messages, and creating agents for web search and code evaluation. The package requires an API key from OpenAI's GPT-3 model and offers various interactive functions for enhancing productivity within Emacs.

flux-aio
Flux All-In-One is a lightweight distribution optimized for running the GitOps Toolkit controllers as a single deployable unit on Kubernetes clusters. It is designed for bare clusters, edge clusters, clusters with restricted communication, clusters with egress via proxies, and serverless clusters. The distribution follows semver versioning and provides documentation for specifications, installation, upgrade, OCI sync configuration, Git sync configuration, and multi-tenancy configuration. Users can deploy Flux using Timoni CLI and a Timoni Bundle file, fine-tune installation options, sync from public Git repositories, bootstrap repositories, and uninstall Flux without affecting reconciled workloads.

clickclickclick
ClickClickClick is a framework designed to enable autonomous Android and computer use using various LLM models, both locally and remotely. It supports tasks such as drafting emails, opening browsers, and starting games, with current support for local models via Ollama, Gemini, and GPT 4o. The tool is highly experimental and evolving, with the best results achieved using specific model combinations. Users need prerequisites like `adb` installation and USB debugging enabled on Android phones. The tool can be installed via cloning the repository, setting up a virtual environment, and installing dependencies. It can be used as a CLI tool or script, allowing users to configure planner and finder models for different tasks. Additionally, it can be used as an API to execute tasks based on provided prompts, platform, and models.

ai-artifacts
AI Artifacts is an open source tool that replicates Anthropic's Artifacts UI in the Claude chat app. It utilizes E2B's Code Interpreter SDK and Core SDK for secure AI code execution in a cloud sandbox environment. Users can run AI-generated code in various languages such as Python, JavaScript, R, and Nextjs apps. The tool also supports running AI-generated Python in Jupyter notebook, Next.js apps, and Streamlit apps. Additionally, it offers integration with Vercel AI SDK for tool calling and streaming responses from the model.

promptwright
Promptwright is a Python library designed for generating large synthetic datasets using a local LLM and various LLM service providers. It offers flexible interfaces for generating prompt-led synthetic datasets. The library supports multiple providers, configurable instructions and prompts, YAML configuration for tasks, command line interface for running tasks, push to Hugging Face Hub for dataset upload, and system message control. Users can define generation tasks using YAML configuration or Python code. Promptwright integrates with LiteLLM to interface with LLM providers and supports automatic dataset upload to Hugging Face Hub.

fragments
Fragments is an open-source tool that leverages Anthropic's Claude Artifacts, Vercel v0, and GPT Engineer. It is powered by E2B Sandbox SDK and Code Interpreter SDK, allowing secure execution of AI-generated code. The tool is based on Next.js 14, shadcn/ui, TailwindCSS, and Vercel AI SDK. Users can stream in the UI, install packages from npm and pip, and add custom stacks and LLM providers. Fragments enables users to build web apps with Python interpreter, Next.js, Vue.js, Streamlit, and Gradio, utilizing providers like OpenAI, Anthropic, Google AI, and more.

PhoGPT
PhoGPT is an open-source 4B-parameter generative model series for Vietnamese, including the base pre-trained monolingual model PhoGPT-4B and its chat variant, PhoGPT-4B-Chat. PhoGPT-4B is pre-trained from scratch on a Vietnamese corpus of 102B tokens, with an 8192 context length and a vocabulary of 20K token types. PhoGPT-4B-Chat is fine-tuned on instructional prompts and conversations, demonstrating superior performance. Users can run the model with inference engines like vLLM and Text Generation Inference, and fine-tune it using llm-foundry. However, PhoGPT has limitations in reasoning, coding, and mathematics tasks, and may generate harmful or biased responses.

agent-mimir
Agent Mimir is a command line and Discord chat client 'agent' manager for LLM's like Chat-GPT that provides the models with access to tooling and a framework with which accomplish multi-step tasks. It is easy to configure your own agent with a custom personality or profession as well as enabling access to all tools that are compatible with LangchainJS. Agent Mimir is based on LangchainJS, every tool or LLM that works on Langchain should also work with Mimir. The tasking system is based on Auto-GPT and BabyAGI where the agent needs to come up with a plan, iterate over its steps and review as it completes the task.
For similar tasks

llmperf
LLMPerf is a tool designed for evaluating the performance of Language Model APIs. It provides functionalities for conducting load tests to measure inter-token latency and generation throughput, as well as correctness tests to verify the responses. The tool supports various LLM APIs including OpenAI, Anthropic, TogetherAI, Hugging Face, LiteLLM, Vertex AI, and SageMaker. Users can set different parameters for the tests and analyze the results to assess the performance of the LLM APIs. LLMPerf aims to standardize prompts across different APIs and provide consistent evaluation metrics for comparison.
For similar jobs

weave
Weave is a toolkit for developing Generative AI applications, built by Weights & Biases. With Weave, you can log and debug language model inputs, outputs, and traces; build rigorous, apples-to-apples evaluations for language model use cases; and organize all the information generated across the LLM workflow, from experimentation to evaluations to production. Weave aims to bring rigor, best-practices, and composability to the inherently experimental process of developing Generative AI software, without introducing cognitive overhead.

LLMStack
LLMStack is a no-code platform for building generative AI agents, workflows, and chatbots. It allows users to connect their own data, internal tools, and GPT-powered models without any coding experience. LLMStack can be deployed to the cloud or on-premise and can be accessed via HTTP API or triggered from Slack or Discord.

VisionCraft
The VisionCraft API is a free API for using over 100 different AI models. From images to sound.

kaito
Kaito is an operator that automates the AI/ML inference model deployment in a Kubernetes cluster. It manages large model files using container images, avoids tuning deployment parameters to fit GPU hardware by providing preset configurations, auto-provisions GPU nodes based on model requirements, and hosts large model images in the public Microsoft Container Registry (MCR) if the license allows. Using Kaito, the workflow of onboarding large AI inference models in Kubernetes is largely simplified.

PyRIT
PyRIT is an open access automation framework designed to empower security professionals and ML engineers to red team foundation models and their applications. It automates AI Red Teaming tasks to allow operators to focus on more complicated and time-consuming tasks and can also identify security harms such as misuse (e.g., malware generation, jailbreaking), and privacy harms (e.g., identity theft). The goal is to allow researchers to have a baseline of how well their model and entire inference pipeline is doing against different harm categories and to be able to compare that baseline to future iterations of their model. This allows them to have empirical data on how well their model is doing today, and detect any degradation of performance based on future improvements.

tabby
Tabby is a self-hosted AI coding assistant, offering an open-source and on-premises alternative to GitHub Copilot. It boasts several key features: * Self-contained, with no need for a DBMS or cloud service. * OpenAPI interface, easy to integrate with existing infrastructure (e.g Cloud IDE). * Supports consumer-grade GPUs.

spear
SPEAR (Simulator for Photorealistic Embodied AI Research) is a powerful tool for training embodied agents. It features 300 unique virtual indoor environments with 2,566 unique rooms and 17,234 unique objects that can be manipulated individually. Each environment is designed by a professional artist and features detailed geometry, photorealistic materials, and a unique floor plan and object layout. SPEAR is implemented as Unreal Engine assets and provides an OpenAI Gym interface for interacting with the environments via Python.

Magick
Magick is a groundbreaking visual AIDE (Artificial Intelligence Development Environment) for no-code data pipelines and multimodal agents. Magick can connect to other services and comes with nodes and templates well-suited for intelligent agents, chatbots, complex reasoning systems and realistic characters.