81 lines
2.6 KiB
Markdown
81 lines
2.6 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
library_name: gguf
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: yasserrmd/glm5.1-distill
|
||
|
|
tags:
|
||
|
|
- gguf
|
||
|
|
- llama.cpp
|
||
|
|
- lfm2
|
||
|
|
- liquid-ai
|
||
|
|
- edge
|
||
|
|
- text-generation-inference
|
||
|
|
- conversational
|
||
|
|
---
|
||
|
|
|
||
|
|
# glm5.1-distill-GGUF
|
||
|
|
|
||
|
|
GGUF quantizations of [`yasserrmd/glm5.1-distill`](https://huggingface.co/yasserrmd/glm5.1-distill),
|
||
|
|
produced with `convert_hf_to_gguf.py` and `llama-quantize` from
|
||
|
|
[`ggml-org/llama.cpp`](https://github.com/ggml-org/llama.cpp).
|
||
|
|
|
||
|
|
The quant ladder mirrors Liquid AI's own LFM2.5-1.2B GGUF releases
|
||
|
|
(e.g. [`LiquidAI/LFM2.5-1.2B-Base-GGUF`](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base-GGUF)).
|
||
|
|
|
||
|
|
## Files
|
||
|
|
|
||
|
|
| File | Quantization | Size |
|
||
|
|
|------|--------------|------|
|
||
|
|
| `glm5.1-distill-BF16.gguf` | BF16 | 2.18 GB |
|
||
|
|
| `glm5.1-distill-Q4_0.gguf` | Q4_0 | 664 MB |
|
||
|
|
| `glm5.1-distill-Q4_K_M.gguf` | Q4_K_M | 697 MB |
|
||
|
|
| `glm5.1-distill-Q5_K_M.gguf` | Q5_K_M | 804 MB |
|
||
|
|
| `glm5.1-distill-Q6_K.gguf` | Q6_K | 918 MB |
|
||
|
|
| `glm5.1-distill-Q8_0.gguf` | Q8_0 | 1.16 GB |
|
||
|
|
|
||
|
|
## Quickstart with `llama.cpp`
|
||
|
|
|
||
|
|
Run any quant directly from the Hub:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
llama-cli -hf yasserrmd/glm5.1-distill-GGUF:Q4_K_M --jinja --ctx-size 32768 --temp 0.1 --top-k 50 --top-p 0.1 --repeat-penalty 1.05
|
||
|
|
```
|
||
|
|
|
||
|
|
Or download manually and serve via `llama-server`:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
huggingface-cli download yasserrmd/glm5.1-distill-GGUF --include "*Q4_K_M*" --local-dir ./glm5.1-distill-GGUF
|
||
|
|
|
||
|
|
llama-server --model ./glm5.1-distill-GGUF/glm5.1-distill-Q4_K_M.gguf --alias "yasserrmd/glm5.1-distill" --threads -1 --n-gpu-layers 99 --ctx-size 32768 --port 8001 --temp 0.1 --top-k 50 --top-p 0.1 --repeat-penalty 1.05 --jinja
|
||
|
|
```
|
||
|
|
|
||
|
|
The recommended sampling parameters above are the official ones published by
|
||
|
|
Liquid AI for the LFM2.5 family.
|
||
|
|
|
||
|
|
## Quickstart with Ollama
|
||
|
|
|
||
|
|
```bash
|
||
|
|
ollama run hf.co/yasserrmd/glm5.1-distill-GGUF:Q4_K_M
|
||
|
|
```
|
||
|
|
|
||
|
|
## Choosing a quant
|
||
|
|
|
||
|
|
| Use case | Recommended |
|
||
|
|
|----------|-------------|
|
||
|
|
| Maximum quality, plenty of RAM | `Q8_0` or `Q6_K` |
|
||
|
|
| Balanced default | `Q4_K_M` (matches Liquid AI's recommendation) |
|
||
|
|
| Smallest footprint, mobile / IoT | `Q4_0` |
|
||
|
|
| Lossless reference | `BF16` (only if you need it for further re-quantization) |
|
||
|
|
|
||
|
|
> **Note**: imatrix-based quantization is currently not supported for the LFM2
|
||
|
|
> architecture in upstream llama.cpp ([issue #14979](https://github.com/ggml-org/llama.cpp/issues/14979)).
|
||
|
|
> These files are plain k-quants, the same scheme used in Liquid AI's official
|
||
|
|
> GGUF releases.
|
||
|
|
|
||
|
|
## Source model
|
||
|
|
|
||
|
|
For training data, hyperparameters, evaluation, and limitations see the source
|
||
|
|
repo: [`yasserrmd/glm5.1-distill`](https://huggingface.co/yasserrmd/glm5.1-distill).
|