Quantisation
38 updates on Quantisation.
llm.npu prefills 1,106 tok/s for Qwen1.5-1.8B on a Snapdragon 8 Gen 3
Peking University and BUPT move prompt processing onto the Hexagon NPU and report up to 43.6x faster prefill and up to 59.5x lower energy use.
Apple Intelligence pairs a 3-billion-parameter on-device model with a server model
Apple reports 0.6 ms per prompt token and 30 tokens per second on an iPhone 15 Pro for a model compressed to an average of 3.7 bits per weight.
ExecuTorch alpha runs Llama 2 7B on iPhone 15 Pro and Galaxy phones
PyTorch's edge runtime brought 4-bit Llama 2 7B to iPhone and Galaxy handsets, added early Llama 3 8B support and leaned on Apple, Arm and Qualcomm.
Microsoft runs Phi-3-mini offline on an iPhone 14 at over 12 tokens per second
The 3.8-billion-parameter model takes about 1.8 GB at 4-bit and scores 69 percent on MMLU, which Microsoft compares to Mixtral 8x7B and GPT-3.5.
One shared on-device LLM keeps a context per app and switches in 0.27 seconds
Peking University, BUPT and Tsinghua researchers compress and swap each app conversation context in 16-token chunks, switching up to 20 times faster.
BitNet b1.58 gives every weight three values and runs 3B in 2.22 GB
Microsoft trained models whose every weight is -1, 0 or 1, which replaces multiplication with addition, and reports parity with full-precision Llama from 3B.
MobiLlama is a fully transparent 0.5B model that runs in 770 MB on a phone
MBZUAI published a 0.5B model that shares one feed-forward block across all layers and reports 7.02 tok/s in 770 MB on a Snapdragon 685 phone.
Arm runs Llama 2 7B at 9.6 tokens per second on three mobile CPU cores
Arm demonstrated Llama 2 7B generating 9.6 tokens per second on three Cortex-A700 CPU cores, using int4 kernels it says beat llama.cpp by 20 percent.
MIT HAN Lab publishes TinyChatEngine for 4-bit LLMs on laptops and Raspberry Pi
The from-scratch C and C++ engine from Song Han's group at MIT runs 4-bit LLaMA2-7B on x86, Apple silicon, Raspberry Pi and Nvidia GPUs.
LLMFarm runs llama.cpp models offline on iOS and macOS
The iOS and macOS app loads a quantised model file off the device and chats with it offline, with 3B the practical size on an iPhone.
Qualcomm argues for hybrid AI, citing 10 times the cost per generative AI search query
Qualcomm argues that cloud-only inference cannot scale, and puts models of 1B to 10B parameters on phones and laptops at INT4.
Georgi Gerganov publishes llama.cpp, LLM inference in plain C and C++
The dependency-free C and C++ inference engine supports 1.5-bit to 8-bit quantisation and treats Apple silicon as a first-class target.