Document RKLLM artifacts and checksums
Browse files
README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: openbmb/MiniCPM4-0.5B
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
library_name: rkllm
|
| 6 |
+
tags:
|
| 7 |
+
- rkllm
|
| 8 |
+
- rockchip
|
| 9 |
+
- rk3576
|
| 10 |
+
- rk3588
|
| 11 |
+
- minicpm4
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# MiniCPM4-0.5B-RKLLM
|
| 15 |
+
|
| 16 |
+
RKLLM-converted MiniCPM4-0.5B language-model artifacts for Rockchip RK3576 and RK3588 NPUs.
|
| 17 |
+
|
| 18 |
+
These files are hardware-specific `.rkllm` binaries. They are not Transformers checkpoints and require a compatible Rockchip RKLLM runtime; they cannot be loaded directly with Transformers, llama.cpp, or Ollama.
|
| 19 |
+
|
| 20 |
+
## Base model
|
| 21 |
+
|
| 22 |
+
- Upstream model: [openbmb/MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B)
|
| 23 |
+
- License: Apache-2.0
|
| 24 |
+
- Model type: LLM (text only)
|
| 25 |
+
|
| 26 |
+
## Conversion and variants
|
| 27 |
+
|
| 28 |
+
Converted with RKLLM Toolkit 1.3.0. Use a file built for the exact target SoC.
|
| 29 |
+
|
| 30 |
+
| Target | Quantization | File | SHA256 |
|
| 31 |
+
| --- | --- | --- | --- |
|
| 32 |
+
| RK3576 | W4A16 (g128) | [MiniCPM4-0.5B_RK3576_w4a16_g128.rkllm](RK3576/MiniCPM4-0.5B_RK3576_w4a16_g128.rkllm) | `34060c78f36ee6ac2a772d5353d32d700fe940ea48564f083b8665a85fdc499c` |
|
| 33 |
+
| RK3576 | W8A8 | [MiniCPM4-0.5B_RK3576_w8a8.rkllm](RK3576/MiniCPM4-0.5B_RK3576_w8a8.rkllm) | `10f2bfa298b5a79077fc52f339ead0446e9e834632226a8bfcc3d003d42a0ba8` |
|
| 34 |
+
| RK3588 | W8A8 | [MiniCPM4-0.5B_RK3588_w8a8.rkllm](RK3588/MiniCPM4-0.5B_RK3588_w8a8.rkllm) | `fef1e215dfa69391d2c5da53eb918020419a71e6b3539d9ef91d56d956b602ea` |
|
| 35 |
+
|
| 36 |
+
The repository also includes `MiniCPM4-0.5B_data_quant.json`, used as calibration data during conversion.
|
| 37 |
+
|
| 38 |
+
## Usage
|
| 39 |
+
|
| 40 |
+
Download a matching artifact with the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli):
|
| 41 |
+
|
| 42 |
+
```bash
|
| 43 |
+
hf download HanzoHuang/MiniCPM4-0.5B-RKLLM \
|
| 44 |
+
RK3576/MiniCPM4-0.5B_RK3576_w4a16_g128.rkllm \
|
| 45 |
+
--local-dir MiniCPM4-0.5B-RKLLM
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
Run it with the RKLLM runtime and the upstream MiniCPM4 chat template. For a Docker deployment, see [Hanzo-Huang/rkllm-docker](https://github.com/Hanzo-Huang/rkllm-docker).
|
| 49 |
+
|
| 50 |
+
## Limitations
|
| 51 |
+
|
| 52 |
+
Conversion can change output quality relative to the upstream model. Runtime, driver, memory, and toolkit compatibility depend on the Rockchip software stack installed on the device.
|
| 53 |
+
|
| 54 |
+
## Acknowledgements
|
| 55 |
+
|
| 56 |
+
Thanks to OpenBMB, Rockchip, and the RKLLM community.
|