63 lines
2.2 KiB
Markdown
63 lines
2.2 KiB
Markdown
---
|
||
language:
|
||
- id
|
||
license: apache-2.0
|
||
library_name: gguf
|
||
pipeline_tag: text-generation
|
||
base_model: AksaraLLM/AksaraLLM-Qwen-1.5B
|
||
tags:
|
||
- gguf
|
||
- llama.cpp
|
||
- ollama
|
||
- indonesian
|
||
- aksarallm
|
||
- qwen2
|
||
---
|
||
# AksaraLLM-Qwen-1.5B-GGUF
|
||
|
||
GGUF quantizations of [`AksaraLLM/AksaraLLM-Qwen-1.5B`](https://huggingface.co/AksaraLLM/AksaraLLM-Qwen-1.5B) for inference with [llama.cpp](https://github.com/ggml-org/llama.cpp), [Ollama](https://ollama.ai), [LM Studio](https://lmstudio.ai), and other GGUF runtimes.
|
||
|
||
## Files
|
||
|
||
| File | Quant | Size | Recommended use |
|
||
|---|---|---|---|
|
||
| `AksaraLLM-Qwen-1.5B.f16.gguf` | F16 | 3.56 GB | lossless from safetensors |
|
||
| `AksaraLLM-Qwen-1.5B.q8_0.gguf` | Q8_0 | 1.89 GB | near-lossless, ~2× smaller |
|
||
| `AksaraLLM-Qwen-1.5B.q6_k.gguf` | Q6_K | 1.46 GB | high quality, ~2.5× smaller |
|
||
| `AksaraLLM-Qwen-1.5B.q5_k_m.gguf` | Q5_K_M | 1.29 GB | good quality, ~3× smaller |
|
||
| `AksaraLLM-Qwen-1.5B.q4_k_m.gguf` | Q4_K_M | 1.12 GB | recommended default, ~4× smaller |
|
||
|
||
## CPU benchmark (AMD EPYC 7763, 2 threads, AVX2)
|
||
|
||
| Quant | Prompt eval (32 tok) | Generation (16 tok) |
|
||
|---|---:|---:|
|
||
| `q4_k_m` | **23.7 tok/s** | **11.8 tok/s** |
|
||
|
||
So a 1.78B model at q4_k_m runs comfortably on a CPU laptop. Larger quants (q5_k_m, q6_k, q8_0) trade a bit of speed for better quality.
|
||
|
||
## Quick start — llama.cpp
|
||
|
||
```bash
|
||
huggingface-cli download AksaraLLM/AksaraLLM-Qwen-1.5B-GGUF AksaraLLM-Qwen-1.5B.q4_k_m.gguf --local-dir .
|
||
./llama-cli -m AksaraLLM-Qwen-1.5B.q4_k_m.gguf -p "Indonesia adalah" -n 64
|
||
```
|
||
|
||
## Quick start — Ollama
|
||
|
||
```bash
|
||
huggingface-cli download AksaraLLM/AksaraLLM-Qwen-1.5B-GGUF AksaraLLM-Qwen-1.5B.q4_k_m.gguf Modelfile --local-dir .
|
||
ollama create aksara-aksarallm-qwen-1.5b -f Modelfile
|
||
ollama run aksara-aksarallm-qwen-1.5b "Apa ibukota Indonesia?"
|
||
```
|
||
|
||
## Source model
|
||
|
||
See [`AksaraLLM/AksaraLLM-Qwen-1.5B`](https://huggingface.co/AksaraLLM/AksaraLLM-Qwen-1.5B) for architecture, training data, eval results, and limitations.
|
||
|
||
## Conversion provenance
|
||
|
||
- Converted with [`convert_hf_to_gguf.py`](https://github.com/ggml-org/llama.cpp/blob/master/convert_hf_to_gguf.py) from llama.cpp
|
||
- Quantized with `llama-quantize` from the same build
|
||
- Architecture detected as `qwen2`
|
||
- All files listed above are reproducible from the source HF safetensors
|