Multi-GPU Evaluation with Hugging Face accelerate

The Language Model Evaluation Harness is the backend for 🤗 Hugging Face's popular Open LLM Leaderboard, has been used in hundreds of papers is used internally by dozens of companies including NVIDIA, Cohere, Nous Research, Booz Allen Hamilton, and Mosaic ML.

Install

To install the lm-eval package from the github repository, run:

git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .

We also provide a number of optional dependencies for . Extras can be installed via pip install -e ".[NAME]"

Name	Use
anthropic	For using Anthropic's models
dev	You probably don't want to use this
gptq	For loading models with GPTQ
testing	You probably don't want to use this
multilingual	For multilingual tokenizers
openai	For using OpenAI's models
promptsource	For using PromtSource prompts
sentencepiece	For using the sentencepiece tokenizer
vllm	For loading models with vLLM
all	Loads all extras

Basic Usage

Hugging Face transformers

To evaluate a model hosted on the HuggingFace Hub (e.g. GPT-J-6B) on hellaswag you can use the following command:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 8

Additional arguments can be provided to the model constructor using the --model_args flag. Most notably, this supports the common practice of using the revisions feature on the Hub to store partially trained checkpoints, or to specify the datatype for running a model:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/pythia-160m,revision=step100000,dtype="float" \
    --tasks lambada_openai,hellaswag \
    --device cuda:0 \
    --batch_size 8

Models that are loaded via both transformers.AutoModelForCausalLM (autoregressive, decoder-only GPT style models) and transformers.AutoModelForSeq2SeqLM (such as encoder-decoder models like T5) in Huggingface are supporteded.

Batch size selection can be automated by setting the --batch_size flag to auto. This will perform automatic detection of the largest batch size that will fit on your device. On tasks where there is a large difference between the longest and shortest example, it can be helpful to periodically recompute the largest batch size, to gain a further speedup. To do this, append :N to above flag to automatically recompute the largest batch size N times. For example, to recompute the batch size 4 times, the command would be:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/pythia-160m,revision=step100000,dtype="float" \
    --tasks lambada_openai,hellaswag \
    --device cuda:0 \
    --batch_size auto:4

Alternatively, you can use lm-eval instead of lm_eval.

Note

Just like you can provide a local path to transformers.AutoModel, you can also provide a local path to lm_eval via --model_args pretrained=/path/to/model

Multi-GPU Evaluation with Hugging Face accelerate

To parallelize evaluation of HuggingFace models across multiple GPUs, we leverage the accelerate 🚀 library as follows:

accelerate launch -m lm_eval --model hf \
    --tasks lambada_openai,arc_easy \
    --batch_size 16

This will perform data-parallel evaluation: that is, placing a single full copy of your model onto each available GPU and splitting batches across GPUs to evaluate on K GPUs K times faster than on one.

If your model is is too large to be run on a single one of your GPUs then you can use accelerate with Fully Sharded Data Parallel (FSDP) that splits the weights of the model across your data parallel ranks. To enable this, ensure you select YES when asked Do you want to use FullyShardedDataParallel? when running accelerate config. To enable memory-efficient loading, select YES when asked Do you want each individually wrapped FSDP unit to broadcast module parameters from rank 0 at the start?. This will ensure only the rank 0 process loads the model and then broadcasts the parameters to the other ranks instead of having each rank load all parameters which can lead to large RAM usage spikes around the start of the script that may cause errors.

To pass even more advanced keyword arguments to accelerate, we allow for the following arguments as well:

• device_map_option: How to split model weights across available GPUs. defaults to "auto".
• max_memory_per_gpu: the max GPU memory to use per GPU in loading the model.
• max_cpu_memory: the max amount of CPU memory to use when offloading the model weights to RAM.
• offload_folder: a folder where model weights will be offloaded to disk if needed.

To use accelerate with the lm-eval command, use

accelerate launch --no_python lm-eval --model ...

Tensor Parallel + Optimized Inference with vLLM

We also support vLLM for faster inference on supported model types.

lm_eval --model vllm \
    --model_args pretrained={model_name},tensor_parallel_size={number of GPUs to use},dtype=auto,gpu_memory_utilization=0.8 \
    --tasks lambada_openai \
    --batch_size auto

For a full list of supported vLLM configurations, please reference our vLLM integration and the vLLM documentation.

Model APIs and Inference Servers

Our library also supports the evaluation of models served via several commercial APIs, and we hope to implement support for the most commonly used performant local/self-hosted inference servers.

To call a hosted model, use:

export OPENAI_API_SECRET_KEY=YOUR_KEY_HERE
lm_eval --model openai-completions \
    --model_args engine=davinci \
    --tasks lambada_openai,hellaswag

Note that for externally hosted models, configs such as --device and --batch_size should not be used and do not function. Just like you can use --model_args to pass arbitrary arguments to the model constructor for local models, you can use it to pass arbitrary arguments to the model API for hosted models. See the documentation of the hosting service for information on what arguments they support.

API or Inference Server	Implemented?	--model <xxx> name	Models supported:	Request Types:
OpenAI Completions	✔️	openai, openai-completions, gooseai	up to code-davinci-002	generate_until, loglikelihood, loglikelihood_rolling
OpenAI ChatCompletions	❌ Not yet - needs testing!	N/A	All ChatCompletions API models	generate_until (no logprobs)
Anthropic	✔️	anthropic	Supported Anthropic Engines	generate_until (no logprobs)
GooseAI	✔️ (not separately maintained)	openai, openai-completions, gooseai (same interface as OpenAI Completions)		generate_until, loglikelihood, loglikelihood_rolling
Textsynth	✔️	textsynth	All supported engines	generate_until, loglikelihood, loglikelihood_rolling
Cohere	⌛ - blocked on Cohere API bug	N/A	All cohere.generate() engines	generate_until, loglikelihood, loglikelihood_rolling
Llama.cpp (via llama-cpp-python)	✔️	gguf, ggml	All models supported by llama.cpp	generate_until, loglikelihood, loglikelihood_rolling
vLLM	✔️	vllm	Most HF Causal Language Models	generate_until, loglikelihood, loglikelihood_rolling
Your inference server here!	...	...	...	...

It is on our roadmap to create task variants designed to enable models which do not serve logprobs/loglikelihoods to be compared with generation performance of open-source models.

Other Frameworks

A number of other libraries contain scripts for calling the eval harness through their library. These include GPT-NeoX, Megatron-DeepSpeed, and mesh-transformer-jax.

Additional Features

If you have a Metal compatible Mac, you can run the eval harness using the MPS back-end by replacing --device cuda:0 with --device mps (requires PyTorch version 2.1 or higher).

Note

You can inspect what the LM inputs look like by running the following command:

python write_out.py \
    --tasks all_tasks \
    --num_fewshot 5 \
    --num_examples 10 \
    --output_base_path /path/to/output/folder

This will write out one text file for each task.

To verify the data integrity of the tasks you're performing in addition to running the tasks themselves, you can use the --check_integrity flag:

lm_eval --model openai \
    --model_args engine=davinci \
    --tasks lambada_openai,hellaswag \
    --check_integrity

Advanced Usage Tips

For models loaded with the HuggingFace transformers library, any arguments provided via --model_args get passed to the relevant constructor directly. This means that anything you can do with AutoModel can be done with our library. For example, you can pass a local path via pretrained= or use models finetuned with PEFT by taking the call you would run to evaluate the base model and add ,peft=PATH to the model_args argument:

lm_eval --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6b,parallelize=True,load_in_4bit=True,peft=nomic-ai/gpt4all-j-lora \
    --tasks openbookqa,arc_easy,winogrande,hellaswag,arc_challenge,piqa,boolq \
    --device cuda:0

GPTQ quantized models can be loaded by specifying their file names in ,gptq=NAME (or ,gptq=True for default names) in the model_args argument:

lm_eval --model hf \
    --model_args pretrained=model-name-or-path,gptq=model.safetensors,gptq_use_triton=True \
    --tasks hellaswag

We support wildcards in task names, for example you can run all of the machine-translated lambada tasks via --task lambada_openai_mt_*.

To save evaluation results provide an --output_path. We also support logging model responses with the --log_samples flag for post-hoc analysis.

Additionally, one can provide a directory with --use_cache to cache the results of prior runs. This allows you to avoid repeated execution of the same (model, task) pairs for re-scoring.

For a full list of supported arguments, check out the interface guide in our documentation!

How to Contribute or Learn More?

For more information on the library and how everything fits together, check out all of our documentation pages! We plan to post a larger roadmap of desired + planned library improvements soon, with more information on how contributors can help.

You can also ask for help, or discuss new features with the maintainers in the #lm-thunderdome channel of the EleutherAI discord! If you've used the library and have had a positive (or negative) experience, we'd love to hear from you!

Implementing new tasks

To implement a new task in the eval harness, see this guide.

In general, we following the following priority list for addressing concerns about prompting and other eval details:

1. If there is widespread agreement among people who train LLMs, use the agreed upon procedure.
2. If there is a clear and unambiguous official implementation, use that procedure.
3. If there is widespread agreement among people who evaluate LLMs, use the agreed upon procedure.
4. If there are multiple common implementations but not universal or widespread agreement, use our preferred option among the common implementations. As before, prioritize choosing from among the implementations found in LLM training papers.

These are guidelines and not rules, and can be overruled in special circumstances.

We try to prioritize agreement with the procedures used by other groups to decrease the harm when people inevitably compare runs across different papers despite our discouragement of the practice. Historically, we also prioritized the implementation from "Language Models are Few Shot Learners" as our original goal was specifically to compare results with that paper.