mps-bitsandbytes 0.1.1

8-bit and 4-bit quantization for PyTorch on Apple Silicon (M1/M2/M3/M4)

MPS BitsAndBytes

8-bit quantization for PyTorch on Apple Silicon (M1/M2/M3/M4).

50% memory savings for storing model weights, with no speed penalty using smart caching.

Features

• Linear8bit: Drop-in replacement for nn.Linear with int8 weights
• Smart caching: Dequantize once, run fast fp16 matmul (AMX accelerated)
• QLoRA ready: Perfect for fine-tuning large models on Mac
• Pure PyTorch: No custom kernels needed, works out of the box

Installation

pip install mps-bitsandbytes

Or from source:

git clone https://github.com/mpsops/mps-bitsandbytes
cd mps-bitsandbytes
pip install -e .

Quick Start

import torch
from mps_bitsandbytes import Linear8bit, quantize_model

# Convert existing model to 8-bit
model = YourModel().to('mps')
model = quantize_model(model, device='mps')

# Or convert individual layers
linear_8bit = Linear8bit.from_linear(some_linear_layer)

# Use normally - same API, 50% less memory for weights
output = model(input)

How It Works

1. Storage: Weights stored as int8 (1 byte per param vs 2 bytes for fp16)
2. First forward: Dequantize int8 → fp16, cache the result
3. Subsequent forwards: Use cached fp16 weights, fast AMX matmul

This gives you:

• 50% memory savings on disk and when loading weights
• Same inference speed as fp16 (once cached)
• Compatible with QLoRA training

Memory Savings

Model Size	FP16	INT8	Savings
7B params	14 GB	7 GB	7 GB
13B params	26 GB	13 GB	13 GB
70B params	140 GB	70 GB	70 GB

Configuration

# Default: cache enabled (fast, uses memory during inference)
layer = Linear8bit.from_linear(linear, use_cache=True)

# Memory-constrained: no cache (slower, minimum memory)
layer = Linear8bit.from_linear(linear, use_cache=False)

# Clear cache to free memory
layer.clear_cache()

QLoRA Training

from mps_bitsandbytes import quantize_model
from peft import get_peft_model, LoraConfig

# Load model in 8-bit
model = AutoModelForCausalLM.from_pretrained("model_name")
model = quantize_model(model.to('mps'))

# Add LoRA adapters (these stay in fp16 for gradients)
lora_config = LoraConfig(r=8, lora_alpha=16, target_modules=["q_proj", "v_proj"])
model = get_peft_model(model, lora_config)

# Train - base weights frozen in int8, LoRA in fp16
trainer.train()

Benchmarks

Tested on M1 Max, batch_size=32, hidden_dim=4096:

Method	Forward Time	Memory
FP16	1.08 ms	100 MB
INT8 (cached)	0.98 ms	50 MB + cache
INT8 (no cache)	9.65 ms	50 MB

Limitations

• First forward is slower: Need to dequantize weights once
• Cache uses memory: During inference, cached fp16 weights use extra memory
• No int8 matmul acceleration: Apple Silicon AMX only supports fp16/fp32

For maximum memory savings during inference (no cache), use use_cache=False, but expect ~10x slower inference.

Credits

• bitsandbytes - Original CUDA implementation
• LLM.int8() - Paper by Tim Dettmers et al.

License

MIT