GGUF files store all data in little-endian. On big-endian systems (AIX, z/OS), llama.cpp detects the mismatch and fails. This pre-converted model works directly.

Model Details

Field	Value
Parameters	1.54B
Architecture	Qwen2 transformer, 28 layers, GQA (12 Q / 2 KV heads)
Quantization	Q4_K_M
Context length	32,768 tokens
File	`Qwen2.5-1.5B-Instruct-Q4_K_M-be.gguf` (940 MB)
Endianness	Big-endian
Source	Qwen/Qwen2.5-1.5B-Instruct

Performance on IBM POWER9 (AIX 7.3)

Tested at 16 threads, SMT-2:

Generation: 7.55 tok/s
Prompt processing: 2.92 tok/s

Quick Start (AIX)

bash

git clone https://gitlab.com/librepower/llama-aix.git
cd llama-aix
./scripts/fetch_upstream.sh && ./scripts/build_aix_73.sh
wget https://huggingface.co/librepowerai/Qwen2.5-1.5B-Instruct-Q4_K_M-BE/resolve/main/Qwen2.5-1.5B-Instruct-Q4_K_M-be.gguf
./build/bin/llama-simple -m Qwen2.5-1.5B-Instruct-Q4_K_M-be.gguf -n 256 -t 16 "Your prompt"

LibrePower — Unlocking IBM Power Systems through open source. https://librepower.org | hello@librepower.org

librepowerai/Qwen2.5-1.5B-Instruct-Q4_K_M-BE

Qwen2.5-1.5B-Instruct Q4_K_M — Big-Endian

Why Big-Endian?

Model Details

Performance on IBM POWER9 (AIX 7.3)

Quick Start (AIX)

Related

librepowerai/
Qwen2.5-1.5B-Instruct-Q4_K_M-BE