This document provides different strategies for quantizing the various models available in LitGPT to reduce GPU memory usage, which is useful for running larger models on certain GPU hardware.
All the examples below were run on an A100 40GB GPU with CUDA 12.1.
Note
Quantization also supports finetuning via QLoRA
It's useful to start with a baseline to have a reference point for memory savings via the various quantization methods.
litgpt generate tiiuae/falcon-7b \
--precision 32-true \
--max_new_tokens 256
...
Time for inference 1: 6.93 sec total, 36.96 tokens/sec.
Memory used: 28.95 GBFirst, using a lower precision compared to 32-bit float can result in two times reduced memory consumption. You can either try setting --precision 16-true for regular 16-bit precision or --precision bf16-true if your GPU supports brain-float 16-bit precision. (This brief video explains the difference between regular 16-bit and bf16-bit precision.)
In short, when --precision bf16-true or --precision 16-true is used, the model weights will automatically be converted and consume less memory.
However, this might not be enough for large models or when using GPUs with limited memory.
litgpt generate tiiuae/falcon-7b \
--precision bf16-true \
--max_new_tokens 256
...
Time for inference 1: 5.37 sec total, 47.66 tokens/sec.
Memory used: 14.50 GBTo reduce the memory requirements further, LitGPT supports several quantization techniques, which are shown below.
Tip
Most quantization examples below also use the --precision bf16-true setting explained above. If your GPU does not support bfloat16, you can change it to --precision 16-true.
Enabled with bitsandbytes. Check out the paper to learn more about how it works.
Important
bitsandbytes only supports CUDA devices and the Linux operating system.
Windows users should use WSL2.
Uses the normalized float 4 (nf4) data type. This is recommended over "fp4" based on the paper's experimental results and theoretical analysis.
pip install bitsandbytes
litgpt generate tiiuae/falcon-7b \
--quantize bnb.nf4 \
--precision bf16-true \
--max_new_tokens 256
...
Time for inference 1: 6.80 sec total, 37.62 tokens/sec
Memory used: 5.72 GBEnabled with bitsandbytes. Check out the paper to learn more about how it works.
"dq" stands for "Double Quantization" which reduces the average memory footprint by quantizing the quantization constants. In average, this amounts to about 0.37 bits per parameter (approximately 3 GB for a 65B model).
pip install bitsandbytes
litgpt generate tiiuae/falcon-7b \
--quantize bnb.nf4-dq \
--precision bf16-true \
--max_new_tokens 256
...
Time for inference 1: 8.09 sec total, 30.87 tokens/sec
Memory used: 5.38 GBEnabled with bitsandbytes. Check out the paper to learn more about how it works.
Uses pure FP4 quantization.
pip install bitsandbytes
litgpt generate tiiuae/falcon-7b \
--quantize bnb.fp4 \
--precision bf16-true \
--max_new_tokens 256
...
Time for inference 1: 6.92 sec total, 36.98 tokens/sec
Memory used: 5.72 GBEnabled with bitsandbytes. Check out the paper to learn more about how it works.
"dq" stands for "Double Quantization" which reduces the average memory footprint by quantizing the quantization constants. In average, this amounts to about 0.37 bits per parameter (approximately 3 GB for a 65B model).
pip install bitsandbytes
litgpt generate tiiuae/falcon-7b \
--quantize bnb.fp4-dq \
--precision bf16-true \
--max_new_tokens 256
...
Time for inference 1: 10.02 sec total, 25.54 tokens/sec
Memory used: 5.38 GBEnabled with bitsandbytes. Check out the paper to learn more about how it works.
pip install bitsandbytes
litgpt generate tiiuae/falcon-7b \
--quantize bnb.int8 \
--precision 16-true \
--max_new_tokens 256
...
Time for inference 1: 20.22 sec total, 12.66 tokens/sec
Memory used: 8.70 GB