ardfork / exllama

A more memory-efficient rewrite of the HF transformers implementation of Llama for use with quantized weights.

ExLlama

A rewrite of the HF transformers implementation of Llama with the following goals, among others:

• Designed for use with quantized weights
• Fast and memory-efficient inference (not just attention)
• Mapping across multiple devices
• Built-in (multi) LoRA support
• Companion library of funky sampling functions

Disclaimer: This is currently a preview of a work in progress. Or maybe a proof of concept. Either way any part of it is subject to change.

Hardware/software requirements

I am developing on an RTX 4090 and an RTX 3070-Ti. Both cards support the CUDA kernel, but there might be incompatibilities with older cards. I have no way of testing that right now.

I have no idea if this works on Windows/WSL, but feel free to try and contribute/give feedback.

Dependencies

This list might be incomplete:

• torch tested on 2.1.0 (nightly) with cu118, might work with older CUDA versions also
• safetensors 0.3.1
• sentencepiece
• ninja

Limitations

As of currently (working on it):

• No support for v1 models without groupsize
• I've encountered models with nonstandard layouts and datatypes (e.g. float32 embedding table). It'll take a while to make sure all the possible permutations are supported.

How to

There is no installer or package at the moment, but try this:

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

pip install safetensors sentencepiece ninja

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

python test_benchmark_inference.py -t <path_to_tokenizer.model> -c <path_to_config.json> \ 
  -m <path_to_model.safetensors> -p -ppl

Alternatively, just specify a directory containing tokenizer.model, config.json and a single .safetensors file:

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 examples:

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

python test_chatbot.py -d <path_to_model_files> -un "Maxine" -p prompt_assistant.txt -nnl \
  -temp 1.00 -topp 0.95 -beams 5 -beamlen 20

Results so far

New implementation

Model	Size	groupsize	act	Seq. len.	VRAM	Long seq.	Ind.	Ppl
Llama	7B	128	no	2,048 t	5,093 MB	2,497 t/s	103 t/s	6.45
Llama	13B	128	no	2,048 t	8,975 MB	2,291 t/s	66 t/s	5.62
Llama	30B	128	no	2,048 t	20,544 MB	1,508 t/s	32 t/s	4.60
Llama	30B	128	yes	2,048 t	20,558 MB	1,359 t/s	31 t/s	4.55
Llama	30B	32	yes	1,550 t 1	21,254 MB	1,095 t/s	31 t/s	4.52
Koala	13B	128	yes	2,048 t	8,981 MB	1,777 t/s	63 t/s	6.73
WizardLM	30B	-	no 2	2,048 t	19,950 MB	1,422 t/s	33 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, running with a desktop environment, with a few other apps also using VRAM.

All results are inference over a longer sequence, with the last 128 tokens generated individually, up to the sequence length specified. 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 the accuracy of the output. The dataset used is a small sample from WikiText, and scores are not necessarily comparable to other Llama benchmarks.

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?

	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

Comparisons

For reference, here are the best results I was able to achieve with GPTQ-for-LLaMa for the same task, using Sterlind's repo here which appears to be the fastest. The new Triton branch is, as far as I can tell, slower, and the CUDA versions have gotten slower as well over time.

	Seq. len.	VRAM	Long seq.	Ind.	Ppl
7B 4bit 128g, HF	2,048 t	8,940 MB	2,218 t/s	45 t/s	6.44
13B 4bit 128g, HF	2,048 t	14,902 MB	1,413 t/s	36 t/s	5.61
30B 4bit 128g, HF 1	256 t	21,063 MB	168 t/s	26 t/s	4.61
7B 4bit 128g, HF, 16-bit	2,048 t	6,058 MB	2,887 t/s	45 t/s	6.45
13B 4bit 128g, HF, 16-bit	2,048 t	10,563 MB	2,212 t/s	36 t/s	5.62
30B 4bit 128g, HF, 16-bit 1	1,024 t	20,715 MB	850 t/s	23 t/s	4.60

¹ Max sequence length reduced to avoid OoM

Todo

Moved the todo list here.

Recent updates

2023-05-16: Reworked the way act-order models are handled. Rows are now shuffled at load time so zeros and scales can be scanned sequentially. Left-hand side of the matmul is shuffled column-wise accordingly. Performance cost for act-order is quite small now.

2023-05-16: Removed the need to specify groupsize.

2023-05-17: Tested 32g models (30B weights take up a bit too much space still to work on 24 GB of VRAM with full context.) Added error handling to C++/CUDA parts. Cleaned up and simplified the CUDA code a lot, preparing for fused layers.

2023-05-18: Added basic layer streaming. Experimental for now. Modern GPUs should be designed for concurrent data transfer and execution, and there should be enough bandwidth to stream in every nth layer while the preceding n-1 layers are processing. It's still too slow to be useful for generation, though. Also doesn't work with multiple GPUs.

2023-05-19: Wrote a CUDA implementation of the layer norm. Turns out it was a bit of a bottleneck for the smaller models. Noticeably faster now.

2023-05-21: Added beam search implementation. It doesn't process beams in parallel which saves a lot of VRAM but does slow it down a bit. There should be ways to mitigate the slowdown. It's not clear how much better beam search performs in practice, but it's at least theoretically superior and there are other features coming which will build on it, like multi-token repetition penalties and (de-)censoring.

2023-05-22: Added option to auto-split layers across multiple GPUs based on VRAM allocation.

2023-05-22: Added option to dequantize layers at load-time which should speed up inference, but it turns out Torch's fp16 matmul is actually slower than the quantized matmul. Maybe bandwidth is the only bottleneck right now? Need to experiment some more.