ExLlama

A standalone Python/C++/CUDA implementation of Llama for use with 4-bit GPTQ weights, designed to be fast and memory-efficient on modern GPUs.

Disclaimer: The project is coming along, but it's still a work in progress!

Hardware requirements

I am developing on an RTX 4090 and an RTX 3090-Ti. Both cards support the CUDA kernels, but there might be incompatibilities with older cards.

Dependencies

torch tested on 2.0.1 and 2.1.0 (nightly) with cu118
safetensors 0.3.1
sentencepiece
ninja
flask (only for the web UI)

Linux/WSL prerequisites

pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu118

Windows prerequisites

To run on Windows (without WSL):

Install MSVC 2022. You can choose to install the whole Visual Studio 2022 IDE, or alternatively just the Build Tools for Visual Studio 2022 package (make sure Desktop development with C++ is ticked in the installer), it doesn't really matter which.
Install the appropriate version of PyTorch, choosing one of the CUDA versions. I am developing on the nightly build, but the stable version should also work.
Install CUDA Toolkit, (11.7 and 11.8 both seem to work, just make sure to match PyTorch's Compute Platform version).
For best performance, enable Hardware Accelerated GPU Scheduling.

How to

Install dependencies, clone repo and run benchmark:

pip install safetensors sentencepiece ninja

git clone https://github.com/turboderp/exllama
cd exllama

python test_benchmark_inference.py -d <path_to_model_files> -p -ppl

The CUDA extension is loaded at runtime so there's no need to install it separately. It will be compiled on the first run and cached to ~/.cache/torch_extensions/ which could take a little while. If nothing happens at first, give it a minute to compile.

Chatbot example:

python example_chatbot.py -d <path_to_model_files> -un "Jeff" -p prompt_chatbort.txt

Web UI

I made a simple web UI for it. Like the rest of the project, it's a work in progress. Don't look at the JavaScript, it was mostly written by ChatGPT and it will haunt your dreams. But it sort of works, and it's kinda fun, especially multibot mode:

To run it:

pip install flask

python webui/app.py -d <path_to_model_files>

Note that sessions are stored in ~/exllama_sessions/.

Docker

For security benefits and easier deployment, it is also possible to run the web UI in an isolated docker container. Note: the docker image currently only supports NVIDIA GPUs.

Requirements

It is recommended to run docker in rootless mode.

Build

The easiest way to build the docker image is using docker compose. First, set the MODEL_PATH and SESSIONS_PATH variables in the .env file to the actual directories on the host. Then run:

docker compose build

It is also possible to manually build the image:

docker build -t exllama-web

Run

Using docker compose:

docker compose up

The web UI can now be accessed on the host at http://localhost:5000.

The configuration can be viewed in docker-compose.yml and changed by creating a docker-compose.override.yml file.

Run manually:

docker run --gpus all -p 5000:5000 -v <path_to_model_files>:/app/model/ --rm -it exllama-web --host 0.0.0.0:5000

Results so far

New implementation

Model	Size	grpsz	act	Seq. len.	VRAM	Prompt	Best	Worst	Ppl
Llama	7B	128	no	2,048 t	5,194 MB	13,918 t/s	173 t/s	140 t/s	6.45
Llama	13B	128	no	2,048 t	9,127 MB	7,507 t/s	102 t/s	86 t/s	5.60
Llama	30B	128	no	2,048 t	20,795 MB	2,959 t/s	47 t/s	40 t/s	4.60
Llama	30B	128	yes	2,048 t	20,795 MB	2,784 t/s	45 t/s	37 t/s	4.55
Llama	30B	32	yes	1,550 t ¹	21,486 MB	2,636 t/s	41 t/s	37 t/s	4.52
Koala	13B	128	yes	2,048 t	9,127 MB	5,529 t/s	93 t/s	79 t/s	6.73
WizardLM	30B	-	no ²	2,048 t	20,199 MB	2,313 t/s	47 t/s	40 t/s	5.75

¹ Can not achieve full sequence length without OoM (yet)
² Not quite sure if this is act-order or not. Weights have no group index, at least

All tests done on stock RTX 4090 / 12900K, running with a desktop environment, with a few other apps also using VRAM.

"Prompt" speed is inference over the sequence length listed minus 128 tokens. "Worst" is the average speed for the last 128 tokens of the full context (worst case) and "Best" lists the speed for the first 128 tokens in an empty sequence (best case.)

VRAM usage is as reported by PyTorch and does not include PyTorch's own overhead (CUDA kernels, internal buffers etc.) This is somewhat unpredictable anyway. Best bet is to just optimize VRAM usage by the model, probably aiming for 20 GB on a 24 GB GPU to ensure there is room for a desktop environment and all of Torch's internals.

Perplexity is measured only to verify that the models are working. The dataset used is a particular, small sample from WikiText, so scores are not necessarily comparable to other Llama benchmarks.

Dual GPU results

Since many seem to be interested in running 65B models, I can confirm that this works with two 24 GB GPUs. The following benchmarks are from a 4090 + 3090-Ti with -gs 17.2,24:

Model	Size	groupsize	act	Seq. len.	VRAM	Prompt	Best	Worst	Ppl
Llama	65B	128	yes	2,048 t	39,804 MB	1,109 t/s	20 t/s	18 t/s	4.20
Llama	65B	32	yes	2,048 t	43,424 MB	1,037 t/s	17 t/s	16 t/s	4.11

Testing long sequences

The following tests were all done on 30B/65B, 4bit 128g with various settings, just to test the max sequence length and get a sense of what can be achieved with different or multiple GPUs right now. Llama goes incoherent generating past 2048 tokens anyway, but with some fine-tuning, who knows? Note that these tests were run a while ago and the speeds are no longer current.

	Size	Seq. len.	VRAM	Long seq.	Ind.
4090/24GB	30B	2,516 t	22,145 MB	1140 t/s	28 t/s
4090/24GB + 3070Ti/8GB	30B	3,932 t	22,055 MB + 7,377 MB	840 t/s	22 t/s
A6000/48GB (headless)	30B	9,032 t	46,863 MB	645 t/s	12 t/s
A100/80GB (headless)	65B	9,520 t	79,009 MB	650 t/s	9 t/s

Todo

Moved the todo list here.

Compatibility

I downloaded a whole bunch of GPTQ models to test compatibility. Here is the list of models confirmed to be working right now.

Recent updates

2023-05-24: Added fused rotary embeddings and some minor optimizations. 13% faster on 7B, 9% on 13B. Small improvement on larger models. Added best-case scores to benchmark results and some clarification. For easier comparisons to other implementations, or whatever.

2023-05-27: Better memory management in CUDA. Introduced auto switch between Torch's SDP backend and regular matmul attention with some tweaks. Finished CUDA MLP. All in all about 10% faster with these updates.

2023-05-29: Web UI is almost up and running. Having to learn JavaScript, and it turns out I hate JavaScript. But ChatGPT is an incredible resource for learning new languages, I gotta say, so it's not as painful as it could have been. Anyway, in the process of working with the UI I discovered I've been measuring prompt speed incorrectly. Either Torch or CUDA or the GPU driver does some sort of caching or self-calibration or lazy initialization during the first pass through the model, so subsequent passes are actually way faster than what I've been recording. Doesn't do much for individual tokens, but benchmarks updated anyway. Closing in on 10k tokens/second for 7B. (!)

2023-06-02: Web UI is now in a fairly working state. Expect it to be a little scuffed in places. There will be a rewrite at some point to make the client-side code less seizure-inducing. It has multibot mode, chat rewind and editing features, sessions, and more. I'm going to build it out with support for instruct prompting and such, in time.

2023-06-04: Refactored a whole bunch to move more of the work into the extension, setting up for more tuning options to come soon and eventually auto tuning. Also optimized a little, for about a 5% speedup.

2023-06-06: Some minor optimizations. Also it should now compile the extension more easily and run more seamlessly on Windows.

2023-06-09: Fused most of the self-attention step. More to come. Slight speedup already, but more importantly went from 69% actual CPU utilization to 37%. This should do a lot to address the bottleneck on CPUs with lower single-threaded performance.

2023-06-10: Docker support now! And some minor optimizations. Cleaned up the project a bit.

2023-06-11: Added some concurrency a couple of places. It's only beneficial on the 4090, on small models where the cores are somewhat underutilized and the L2 cache can keep up. For the 3090 it's detrimental to performance, so it's disabled by default. YMMV. Use -cs to try it out.

lhl / exllama