News 230720 Now we support training the LLaMA-2 models in this repo. SqueezeLLM achieves higher accuracy for both Vicuna-7B and 13B as compared to the AWQ method and also preserve the accuracy of the FP16 baseline model with 4-bit quantization. Step 3 Load and Start the Model. GPTQ Paper. Links to other models can be found in the index at the bottom. Models in the catalog are organized by collections. As the intensity of electromagnetic energy increases or decreases, it steps up or down from one quantized level to another, rather tha. According to the chart in the llama. In a nutshell Quantized models are basically compressed or "shrunken" versions, easier to run if you don&39;t have strong hardware (and is also easier on storage). Google ColabLlama 2 LangChain RetrievalQA 1. sh file is as follows, and the chat template and some default parameters are nested inside, which can. Debugger preparation. Discover Llama 2 models in AzureMLs model catalog. by Aryanne - opened 18 days ago. However, Llama. Our latest version of Llama - Llama 2 - is now accessible to individuals, creators, researchers, and businesses so they can experiment, innovate, and scale their ideas responsibly. Also supports ExLlama for inference for the best speed. text-generation-webui, the most widely used web UI. 10 ago 2023. However, Llama. 0; Get started with Logic Pro. I've tested on 2x24GB VRAM GPUs, and it works. (389) 5 months ago generate. Aug 20. These models surpass the performance of most open-source chat models on the benchmarks they were tested on. 0; use method Q41 but set offset mu - f sigma and scalingfactor 2 f sigma. First, we want to load a llama-2-7b-chat-hf model (chat model) and train it on the mlabonneguanaco-llama2-1k (1,000 samples), which will produce our fine-tuned model llama-2-7b-miniguanaco. Premium Powerups Explore. It works by reducing the precision of the weights and activations used by a model without affecting (significantly) the overall accuracy. Recently, the world of AI has been abuzz with the release of Llama 2, a large language model developed by Meta. A ftype is produced by llamamodelquantizeinternal and is passed through as-is to llamafilesaver, which writes it to disk without encoding it using GGMLQNTVERSION. dev0, 4bit quantization working with GPTQ for LLaMA. We then disable caching and set a pretraining temperature parameter. Variations Llama 2 comes in a range of parameter sizes 7B, 13B, and 70B as well as pretrained and fine-tuned variations. QLoRA was developed by members of the University of Washington&39;s UW NLP group. GPTQ is SOTA one-shot weight quantization method. &183; Loading the pre-trained Llama 2 7B model and tokenizer. okay, so at first thanks for your reply It&39;s verry appreciate -D. The advent of Llama 2 is a promising development in the field of AI. We will see that the resulting models are very fast for inference. py outputpath c4 --wbits 4 --groupsize 128 --save alpaca7b-4bit. Now laptops and Raspberry Pis can run LLMs. Llama-2-7b-Chat-GPTQ" tokenizer AutoTokenizer. GPTQ is SOTA one-shot weight quantization method. My first model with every type is. 8-bit) post-training quantization leads to significant performance drop mainly due to the high dynamic ranges of activation and a naive activation quantization strategy fails to maintain the capacity. Vicuna v1. cpp also has support for LinuxWindows. Especially for 2-bit quantization, Q21 and Q23 give basically useless results. For challenges with high quantization levels, consider the q4 variant or terminate memory-intensive applications. nf4 without double quantization significantly uses more memory than GPTQ. 3k Code Issues 94 Pull requests 12 Actions Security Insights main 24 branches 0 tags Code 227 commits. With the advent of Llama 2, running strong LLMs locally has become more and more a reality. BitsAndBytesConfig(loadin4bitTrue, 4-bit quantization bnb4bitquanttype&39;nf4&39;, Normalized float 4 bnb4bitusedoublequantTrue, Second quantization after the first bnb4bitcomputedtypebfloat16 Computation. Hugging Face repo. The successor to LLaMA (henceforce "Llama 1"), Llama 2 was trained on 40 more data, has double the context length, and was tuned on a large dataset of human preferences (over 1 million such annotations) to ensure helpfulness and safety. The launch of Llama 2 by Meta has ignited excitement within the community, marking the dawn of an era for well performed large language models that were previously only accessible through company-specific APIs. Nevertheless, I encountered problems. With pretrained models ranging from 7 billion to 70 billion parameters, Llama 2 offers substantial improvements over its predecessor, Llama 1. Models in the catalog are organized by collections. But my bigger concern for production is hosting costs (needing high uptime, low-ish latency). Jump to The day began with a national frenzy over two llamas that escaped f. Quantization is costly. Multi-GPU support2 means LLaMA-65B, rivaling PaLM-540B, runs on 2x3090. You might be able to get that price down, but still it will cost more than the average API user. Quick Start Model Inference and Chat Session. War llamas feel the sting of automation. However, finetuning very large models is prohibitively expensive; regular 16-bit finetuning of a LLaMA 65B parameter model 57 requires more than 780 GB of GPU memory. This method enables 33B model finetuning on a single 24GB GPU and 65B model finetuning on a single 46GB GPU. Transformers has integrated optimum API to perform GPTQ quantization on language models. As a first step, I simply changed llama. Note We are currently working on releasing an new LLM container to support GQA for the 70B model. bin q2K 2 2. Quantization reduces neural network bit-precision, leading to smaller models and faster inference. The Kaitchup. arxiv 2305. This 4-bit post-training quantization technique is different the one that Tim Dettmers working on. Note that any Llama 2 variants can be deployed with OpenLLM if you dont have access to the official Llama 2 model. LLaMa v1 found success in fine-tuning application, with models such as Alpaca able to place well on LLM evaluation leaderboards. 90GHz Base speed 2. - GitHub - ranchlaiquantizations A collection of quantization recipes for various large models including Llama-2-70B, QWen-14B, Baichuan-2-13B, and more. Quantization methods compatible with latest llama. It extends the original Meta AI code by adding training, an optimized fine-tuning script, and the ability to run inference on a consumer GPU (using up to 8GB of memory with quantization). cpp is a big . text-generation-webui, the most widely used web UI. py invokes the quantize binary; quantize. 4-bit GPTQ models reduce VRAM usage by about 75. Similar to 79, but for Llama 2. This is a fork that adds support for ROCm's HIP to use in AMD GPUs, only supported on linux. py TheBlokeLlama-2-7B-GPTQ. To learn more about the quantization model, check out the GPTQ paper. cpp from June 6, commit 2d43387. llama2 quantized model vs. In my own experiments with Llama 2 7B and using 3 different GPUs, I also observed that GPTQ and nf4-doublequant consume a very similar amount of VRAM. When the western South A. You can see it as a way to compress LLMs. However, with finetuning, it depends on the type of data you are using to finetune on. There&39;s still the question of how much the quantization affects the quality of the output. From the README The main goal is to run the model using 4-bit quantization on a MacBook. Locally available model using GPTQ 4bit quantization. Almost indistinguishable from. This is being looked at by the llama. 1 Preliminaries In this work, we study linear quantization i. comfacebookresear Next thing thing we need to do is initialize a text-generation pipeline with Hugging Face transformers. LlaMA 2 Chat GPTQ. This wiki will walk you through the detailed steps of model quantization and local deployment using the llama. 1; TheBlokewizardLM-7B-GGML. It works by reducing the precision of the weights and activations used by a model without affecting (significantly) the overall accuracy. But my bigger concern for production is hosting costs (needing high uptime, low-ish latency). Next page Page 2 In Operation and Summary. This ends up effectively using 2. 2,3,4,5 and 6-bit quantization methods are now available in llama. 14 GB 10. (IST-DASLabgptq1) According to GPTQ paper, As the size of the model increases, the. 2022 and Feb. GGMLTYPEQ2K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Models in the catalog are organized by collections. We're now ready to quantize the model. As we have shown throughout this article, Metas Llama 2 model offers new possibilities for the open-source community. Great news if youre an Israeli war llama Your tour of duty is over. Nevertheless, I encountered problems. The framework will then do the following during the prepare module swaps, it will convert every module of type specified in (1) to the type specified. If you are interested to learn more about LLaVA Lightning, please. 5, which serves well for many use cases. Meta states that Llama 2 was trained on 2 trillion tokens of data from publicly-available sources40 percent more than its first iterationand has a context length of 4096 tokens, twice the context length of Llama 1. exe and the consecutive steps will run without issues. The release of Llama 2 now combines the best elements. The documentation for CTranslate2 contains specific instructions for llama models. This data format is also required by. LLaMa model inference. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety. Jul 25, 2023 With the release of LLaMA v1, we saw a Cambrian explosion of fine-tuned models, including Alpaca, Vicuna, and WizardLM, among others. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. Aug 20. We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3. This tutorial shows how the LLaMA 2 model has improved upon the previous version,. Integer quantization options include 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit (check out the bonus at the end of the article). Also, the paper did not experiment with applying FP8 to. Quantization is done by essentially rounding from one data type to another. LLaMA-2 & Alpaca-2 (Chinese LLaMA-2 & Alpaca-2 LLMs) - ymcuiChinese-LLaMA-Alpaca-2. cpp from June 6, commit 2d43387. Lightning-AI lit-llama Public 5. 26 GB. Link copiado A Meta e a Microsoft lan&231;aram, nesta ter&231;a-feira (18), o Llama 2, a nova gera&231;&227;o do modelo de linguagem de c&243;digo aberto, com disponibilidade gratuita para empresas, startups. We demonstrate how to use the following features. Skip to content Toggle navigation. For example, a model that was trained in 16 bit precision could be quantized to 8, 4, or even 3 bits, all. If that's so, the steps are roughly. bin q41 4 8. 56 GB Original quant. For example, if a note is ahead of the beat by 60 ticks, 50 strength would move the. Llama 2 is the latest Large Language Model (LLM) from Meta AI. compression "initializer" "range" "numinitsamples" 300. 31 ago 2023. The computation speedup and memory reduction of OmniQuant are also demonstrated on real devices. We uses bitsandbytes for quantization and is integrated with Huggingface&39;s PEFT and transformers libraries. 0 license that allows commercial use. You can view models linked from the Introducing Llama 2 tile or filter on the Meta collection, to get started with the Llama 2 models. We'll explore the mathematics behind quantization, immersion fea. Llama 2. I just tried the latest version of this, and on my 409012900K I get 77 tokens per second for Llama 7B-128g. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. I just tried the latest version of this, and on my 409012900K I get 77 tokens per second for Llama 7B-128g. Quantization, in mathematics and digital signal processing, is the process of mapping input values from a large set (often a continuous set) to output values in a (countable) smaller set, often with a finite number of elements. LLaMA-2 & Alpaca-2 (Chinese LLaMA-2 & Alpaca-2 LLMs) - ymcuiChinese-LLaMA-Alpaca-2. with different quantization bit widths based on the LLaMA model family. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. 2GB GPU memory. Created by Adobe Firefly Introduction. One person can start a fork of Llama 2 where they focus on quantization, another person can. Jul 23, 2023 I have tested the Llama2 models locally at various quantization levels, including the 70B model, on my Mac Studio. 1 QUANTIZATION METHODS. Also, the paper did not experiment with applying FP8 to. 5, for instance is one possible and very reasonable mapping. How has the llama gone from near extinction to global sensation Llamas recently have become a relatively common sight around the world. Provided files Name Quant method Bits Size Max RAM required, no GPU offloadingd Use case; llama-2-7b. In my own experiments with Llama 2 7B and using 3 different GPUs, I also observed that GPTQ and nf4-doublequant consume a very similar amount of VRAM. When raised on farms or kept as pets, llamas eat a diet similar to that of sheep a. Skip to content Toggle navigation. In order to load our 13 billion parameter model, we will need to perform some optimization tricks. Jul 19, 2023 Here are just a few of the easiest ways to access and begin experimenting with LLaMA 2 right now 1. Thanks, it worked great, didn&39;t know it was for each sub file. The download links might change, but a single-node, bare metal setup is similar to below Ensure you can use the model via python3 and this example. It works by reducing the precision of the weights and activations used by a model without affecting (significantly) the overall accuracy. Dequantization is the inverse dequant(cFP32,XInt8) XInt8 cFP32 XFP32 (2) The problem with this approach is that if a large magnitude value (i. 4 bits quantization of LLaMA using GPTQ. dev0, 4bit quantization working with GPTQ for. cpp and whisper. Premium Powerups Explore. So let's say quantization increases perplexity by 0. 8-bit Inference Topologies. Specifically, due to their massive. Watch this video on YouTube. 40, 62, 43, 61, 59, 37 and to add desirable or remove undesirable behaviors 43, 2, 4. Open in app. High resource use and slow. 1x as compared to the state-of-the-art methods with the same memory requirement. Jul 24, 2023 Last week, Meta introduced Llama 2, a new large language model with up to 70 billion parameters. Easiest way to use the quantization debugger is to provide tf. Mar 13, 2023 Benj Edwards - 3132023, 416 PM Enlarge Ars Technica 145 Things are moving at lightning speed in AI Land. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. The fine-tuned model, Llama-2-chat, leverages publicly available instruction datasets and over 1 million human annotations. Advertisement Coins. I&39;m using ooba python server. cpp to not quantize one of the tensor types in addition to output. Apache 2. - GitHub - Lightning-AIlit-gpt Hackable implementation of state-of-the-art open-source LLMs based on nanoGPT. First, we want to load a llama-2-7b-chat-hf model (chat model) and train it on the mlabonneguanaco-llama2-1k (1,000 samples), which will produce our fine-tuned model llama-2-7b-miniguanaco. It surpasses prior techniques for 4-bit, 3-bit, and 2-bit. Alternatively, if you want to save time and. 8-bit) post-training quantization leads to significant performance drop mainly due to the high dynamic ranges of activation and a naive activation quantization strategy fails to maintain the capacity. The advent of Llama 2 is a promising development in the field of AI. These chat models can be used as chatbots. q42 and q43 are like new generations of q40 and q41. 4 bits quantization of LLaMa using GPTQ. The fine-tuned model, Llama-2-chat, leverages publicly available instruction datasets and over 1 million human annotations. We are releasing a 7B and 3B model trained on 1T tokens, as well as the preview of a 13B model trained on 600B tokens. 0 coins. QLORA backpropagates gradients through a frozen, 4. That's what the percentages are based on. In addition, Llama 2. RTN is simple Round To Nearest. We then disable caching and set a pretraining temperature parameter. Hey there fellow LLaMA enthusiasts I've been playing around with the GPTQ-for-LLaMa GitHub repo by qwopqwop200 and decided to give quantizing LLaMA models a shot. Contribute to curtisshipleyGPTQ-for-LLaMa-Lib development by creating an account on GitHub. Llama 2. sh file is as follows, and the chat template and some default parameters are nested inside, which can. Many studies on Transformer model quantization have the same observation A simple low-precision (e. Llama 2 but 75 smaller. Metas dedication to refining Llama 2 has led to its fine-tuned versions. Overall, LlaMA 2 was pre-trained using 2 trillion tokens of data taken from publicly available instruction datasets. The general theme is that without a separate bias value, then the model loses a bit if the run of digits contains only positive or only negative values because only other half of the. The bitsandbytes is a lightweight wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM. However, Llama. These figures are on the Llama base models. In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion. Besides the naive approach covered in this article, there are three main quantization techniques NF4, GPTQ, and GGML. CUDA GPU acceleration is now available for Llama 2 70B GGML files. weight (which is already known to have a significant impact on generation quality) and calculated perplexity for Q24 quantization (see issue 1240). int8 () work of Tim Dettmers. fyiinstall-llama-cpp" bash Heres what that one-liner does. Llama 2 is being released with a very permissive community license and is available for commercial use. 4-bit, 5-bit, and 8-bit quantization), each of which offers different trade-offs between efficiency and performance. Creating 3000 new POKEmon with AI (the wrong way) featuring GPT-2; Scaling LLaMA and GPTNeoX to >8k input context ; MEGABYTE Predicting Million-byte Sequences with Multiscale Transformers; Llama 2 access. Quantization Mode. 56 GB Original quant. This page describes how to compile a model with MLC LLM. okaloosa island live cam, nachovidalcom
text-generation-webui, the most widely used web UI. . Llama 2 quantization
Weight storage data type int3. Llama 2 family of models. Llama 2. Changed to support new features proposed by GPTQ. It is because the fine-tuned model Llama-2-Chat model leverages publicly available instruction datasets and over 1 million human annotations. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. Easy-to-use fine-tuning framework using PEFT (PTSFTRLHF with QLoRA) (LLaMA-2, BLOOM, Falcon, Baichuan) bloom transformers falcon llama quantization language-model fine-tuning peft pre-training llm rlhf qlora baichuan-7b llama2. Discover Llama 2 models in AzureMLs model catalog. cpplight-cuda This image only includes the main executable file. You can view models linked from the Introducing Llama 2 tile or filter on the Meta collection, to get started with the Llama 2 models. - Support LoRA finetuning with quantization &183; Issue 54 &183;. We observe large improvements over training-free methods, especially in the. However, the memory required can be reduced by using swap memory. Multi-GPU support2 means LLaMA-65B, rivaling PaLM-540B, runs on 2x3090. Compile Models via MLC. Besides quantization improvements (which will be covered in a future post), ONNX Runtime compresses the size of the cosine and sine caches used in each of the rotary embeddings by 50. Locally available model using GPTQ 4bit quantization. bin q41 4 8. In the SpQR paper they mention that often outliers are clustered together and for that they propose a two level quantization similar to the ones just released in llama. Fine-tuned LLMs, called Llama-2-chat, are optimized for dialogue use cases. We will see that the resulting models are very fast for inference. chat (question-answer) support similar to llamallama2. Smith 24 Jul 2023 4 min read Getty Images GPT-4 Microsoft OpenAI Qualcomm large language models meta Last week, Meta introduced Llama 2, a new large language model with up to 70 billion parameters. cpp no longer supports GGML models. fastLLaMa is an experimental high-performance framework designed to tackle the challenges associated with deploying large language models (LLMs) in production environments. Llama 2, a large language model, is a product of an uncommon alliance between Meta and Microsoft, two competing tech giants at the forefront of artificial intelligence research. Saved searches Use saved searches to filter your results more quickly. Now consider a weight element ww, if we multiply wwith s>1 and the inversely scale x, we will have Q(ws)(xs), which is Q(ws) x s Round(ws) x 1 s, (2) where is the new quantization scaler. This trend encouraged different businesses to launch their own base models with licenses suitable for commercial use, such as OpenLLaMA, Falcon, XGen, etc. bil-ash opened this issue 5 days ago &183; 0 comments. We dont need much VRAM but we need a lot of CPU RAM. GPU acceleration is now available for Llama 2 70B GGML files, with both CUDA (NVidia) and Metal (macOS). llama-hfllama-7b c4 --benchmark 2048 --check model inference with the saved model CUDAVISIBLEDEVICES0. quantize () method quantizes the model and exports it to the OpenVINO format. 4 bit is nice because you can store many more values than 23 bit. 230712 Now we support training the Baichuan-13B model in this repo. Embeddings generation. Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language. 75 on the C4 dataset (see Sec. w2 tensors,. This release includes model weights and starting code for pretrained and fine-tuned Llama language models ranging from 7B to 70B parameters. Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language modelling tasks, but also by their extremely high computational and storage costs. With quantization, you can run LLaMA with a 4GB memory GPU. This data format is also required by. In this blog post, we show all the steps involved in training a LlaMa model to. This tutorial shows how the LLaMA 2 model has improved upon the previous version,. 75 GiB is for a "pure" 3-bit quantization, the others obtained by quantizing various subsets of the tensors with 4 or 5 bits. The dataset, available. 12 min read. Step 3 Load and Start the Model. To keep a more human feel, you can use strength to simply move notes closer to the nearest grid point, but not all the way there. Jul 22, 2023 Llama. Jul 22, 2023 Llama. This is a fork of the LLaMA code that runs LLaMA-13B comfortably within 24 GiB of RAM. You can view models linked from the Introducing Llama 2 tile or filter on the Meta collection, to get started with the Llama 2 models. 15 hours ago maybefalseiStock Unreleased via Getty Images. from typing import List. Use p4d instances for deploying Llama 70B it. The pre-trained LLaMA2 weights are quantized and frozen. Use this discussion to Coordinate. Watch this video on YouTube. Model card for Alpaca-30B. About GGML. Double quantization is necessary to match GPTQ quantization performance. However, Llama. Obtaining and using the Facebook LLaMA 2 model. 3, LLaMA-2, XGen, etc. Jul 22, 2023 Rohan Paul rohanpaulai Run LlaMa-2 for Inferencing with 4 bit quantization First get access to the model following this official guide - github. Transformers has integrated optimum API to perform GPTQ quantization on language models. Besides the naive approach covered in this article, there are three main quantization techniques NF4, GPTQ, and GGML. from huggingfacehub import login. Sign up Product. Jul 23, 2023 I have tested the Llama2 models locally at various quantization levels, including the 70B model, on my Mac Studio. Jul 24, 2023 Matthew S. cpp is a big . Jul 24, 2023 Matthew S. Jul 24, 2023 Matthew S. Table 2 LLaMA TPU v4 HBM requirements (i. py --outtype f16. int8int16 quantization for reduced resource usage and model size. LetterRip 3 mo. cpp Efficient inference implementation with ARM NEON, AVX2 and CUDA - see . GPTQ is SOTA one-shot weight quantization method. Apache 2. bin q80 8 7. RTN is simple Round To Nearest. Running data type float16. Jul 24, 2023 Matthew S. Bigger models - 70B -- use Grouped-Query Attention (GQA) for improved inference scalability. The CodeLlama-7B model alone has 10 different quantization variants. Jul 24, 2023 Fig 1. safetensors Benchmarks looked great as expected,. The content of the chat. cpp is a port of Llama in CC, which makes it possible to run Llama 2 locally using 4-bit integer quantization on Macs. cpp, llama-cpp-python. I believe I've heard that the llama-2-7b and 13b compatible with v1. Jul 19, 2023 Here are just a few of the easiest ways to access and begin experimenting with LLaMA 2 right now 1. GPTQ is SOTA one-shot weight quantization method. Llama 2 (Meta) Meta release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Google ColabLlama 2 LangChain RetrievalQA 1. The bitsandbytes is a lightweight wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM. Creating 3000 new POKEmon with AI (the wrong way) featuring GPT-2; Scaling LLaMA and GPTNeoX to >8k input context ; MEGABYTE Predicting Million-byte Sequences with Multiscale Transformers; Llama 2 access. QA-LoRA expects to find a file pytorchmodel. The current stack quantize. Llama 2 is a family of state-of-the-art open-access large language models released by Meta today, and we&x27;re excited to fully support the launch with comprehensive integration in Hugging Face. Step 1. Like nelson said, 2K per month. Support 4-bit integer quantization 27; Example of Cerebras-GPT inference examplesgpt-2; Example of FLAN-T5 inference 12;. This code is based on GPTQ. The missing blue bars suggest the model size doesnt fit in the specified TPU hardware. 82 GB Original quant method, 4-bit. exe The reason being some version of VC will build the output to. Performance is quite good. Join our Discord Server for updates, support & collaboration Dataset creation, training, weight merging, and quantization instructions are in the docs. Jul 23, 2023 I have tested the Llama2 models locally at various quantization levels, including the 70B model, on my Mac Studio. 2 (at this time of writing, we tested it on bitsandbytes0. &183; Loading the pre-trained Llama 2 7B model and tokenizer. I'd recommend using TheBloke's quantization script with AutoGPTQ. . jobs hiring in lexington ky