Simple and efficient pytorch-native transformer text generation.

Featuring:

1. Very low latency
2. <1000 lines of python
3. No dependencies other than PyTorch and sentencepiece
4. int8/int4 quantization
5. Speculative decoding
6. Tensor parallelism
7. Supports Nvidia and AMD GPUs

This is NOT intended to be a "framework" or "library" - it is intended to show off what kind of performance you can get with native PyTorch :) Please copy-paste and fork as you desire.

For an in-depth walkthrough of what's in this codebase, see this blog post.

Please check the rest of this page about benchmark of LLaMA family models.

We also supported Mixtral 8x7B which is a high-quality sparse mixture of experts (MoE) model, the average token generation rates are:

Note that the benchmarks run on an 8xA100-80GB, power limited to 330W with a hybrid cube mesh topology. Note that all benchmarks are run at batch size=1, making the reported tokens/s numbers equivalent to "tokens/s/user". In addition, they are run with a very small prompt length (just 5 tokens).

For more details about Mixtral 8x7B, please check this page or this note.

In the spirit of keeping the repo minimal, here are various examples of extensions you can make to gpt-fast as PRs.

• Google Gemma
• xAI Grok-1
• Databricks DBRX

Projects inspired by gpt-fast in the community:

• gpt-blazing: applies the same performance optimization strategy to more models (e.g., baichuan2).
• gptfast: applies a subset of the performance optimizations to all Huggingface models
• gpt-accelera: extends gpt-fast to SFT/RM/PPO training and batched inference to optimize the throughput

Download PyTorch nightly Install sentencepiece and huggingface_hub

pip install sentencepiece huggingface_hub

To download llama models, go to https://huggingface.co/meta-llama/Llama-2-7b and go through steps to obtain access. Then login with huggingface-cli login

Models tested/supported

tinyllamas/stories{15,42,100}
openlm-research/open_llama_7b
meta-llama/Llama-2-7b-chat-hf
meta-llama/Llama-2-13b-chat-hf
meta-llama/Llama-2-70b-chat-hf
codellama/CodeLlama-7b-Python-hf
codellama/CodeLlama-34b-Python-hf
mistralai/Mistral-7B-v0.1
mistralai/Mistral-7B-Instruct-v0.1
mistralai/Mistral-7B-Instruct-v0.2
meta-llama/Meta-Llama-3-8B

For example, to convert Llama-2-7b-chat-hf

export MODEL_REPO=meta-llama/Llama-2-7b-chat-hf
./scripts/prepare.sh $MODEL_REPO

Benchmarks run on an 8xA100-80GB, power limited to 330W with a hybrid cube mesh topology. Note that all benchmarks are run at batch size=1, making the reported tokens/s numbers equivalent to "tokens/s/user". In addition, they are run with a very small prompt length (just 5 tokens).

Verifier: Llama-70B (int4), Draft: Llama-7B (int4): 48.4 tok/s

Benchmarks run on one GCD of a MI-250x.

Model definition in model.py, generation code in generate.py.

python generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/model.pth --prompt "Hello, my name is"

To squeeze out a little bit more performance, you can also compile the prefill with --compile_prefill. This will increase compilation times though.

Choose device to use by

# The current support devices: cuda, cpu
export DEVICE=cuda

To generate this version of the model

# Spits out model at checkpoints/$MODEL_REPO/model_int8.pth
python quantize.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --mode int8

To run with int8, just pass the int8 checkpoint to generate.py.

python generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/model_int8.pth --device $DEVICE

To generate int4 version of model

# Spits out model at checkpoints/$MODEL_REPO/model_int4.g32.$DEVICE.pth
python quantize.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --mode int4 --groupsize 32

To run with int4, just pass the int4 checkpoint to generate.py.

python generate.py --checkpoint_path checkpoints/$MODEL_REPO/model_int4.g32.pth --compile

To generate with speculative sampling (DRAFT_MODEL_REPO should point to a smaller model compared with MODEL_REPO).

In this example, the "smaller" model is just the int8 quantized version of the model.

export DRAFT_MODEL_REPO=meta-llama/Llama-2-7b-chat-hf
python generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/model.pth --draft_checkpoint_path checkpoints/$DRAFT_MODEL_REPO/model_int8.pth

Note: Running on an A100 80GB, albeit power-limited to 330 watts. Empirically, seems like peak bandwidth is about 1700 GB/s.

ENABLE_INTRA_NODE_COMM=1 torchrun --standalone --nproc_per_node=2 generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/model.pth

We use the EleutherAI evaluation harness to evaluate our model accuracy. To evaluate the accuracy, make sure the evaluation harness is installed and pass your model checkpoint and desired tasks to eval.py.

python eval.py --checkpoint_path checkpoints/$MODEL_REPO/model.pth --compile --tasks hellaswag winogrande

Note: Generative tasks are currently not supported for gpt-fast

Installation Instructions for the evaluation harness: https://github.com/EleutherAI/lm-evaluation-harness/tree/master#install

We have a pure pytorch implementation of GPTQ that utilizes torch._dynamo.export to access the model structure. You can generate a GPTQ quantized version of int4 quantization by using the same command to quantize it but adding 'gptq' to the quantization mode i.e.

# Spits out model at checkpoints/$MODEL_REPO/model_int4-gptq.g32.pth
python quantize.py --mode int4-gptq --calibration_tasks wikitext --calibration_seq_length 2048

You can then eval or generate text with this model in the same way as above.

gpt-fast is released under the BSD 3 license.

Thanks to:

• Lightning AI for supporting pytorch and work in flash attention, int8 quantization, and LoRA fine-tuning.
• GGML for driving forward fast, on device inference of LLMs
• Karpathy for spearheading simple, interpretable and fast LLM implementations
• MLC-LLM for pushing 4-bit quantization performance on heterogeneous hardware