40 lines
1.3 KiB
Markdown
40 lines
1.3 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- zh
|
||
|
|
- fr
|
||
|
|
- es
|
||
|
|
- de
|
||
|
|
- pt
|
||
|
|
- ru
|
||
|
|
- it
|
||
|
|
- ja
|
||
|
|
- ko
|
||
|
|
- vi
|
||
|
|
- ar
|
||
|
|
tags:
|
||
|
|
- text-generation
|
||
|
|
- causal-lm
|
||
|
|
- rwkv
|
||
|
|
- gguf
|
||
|
|
- llama.cpp
|
||
|
|
- lmstudio
|
||
|
|
base_model: BlinkDL/rwkv7-g1
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# RWKV7-G1G 7.2B - GGUF Quantizations
|
||
|
|
|
||
|
|
GGUF conversions of [BlinkDL/rwkv7-g1g-7.2b](https://huggingface.co/BlinkDL/rwkv7-g1) (Apache 2.0)
|
||
|
|
for use with [llama.cpp](https://github.com/ggml-org/llama.cpp), LM Studio, and other GGUF-compatible runtimes.
|
||
|
|
|
||
|
|
All quants converted from **BF16 base** for maximum lossless quality.
|
||
|
|
**Static quantizations** — no importance matrix (imatrix) was used.
|
||
|
|
Check out this for imatrix quantizations: [RemySkye/rwkv7-g1g-7.2b-i1-GGUF](https://huggingface.co/RemySkye/rwkv7-g1g-7.2b-i1-GGUF)
|
||
|
|
|
||
|
|
*S/M/L K-quant variants had **no effect on size** for RWKV architecture — only `_M` variants are included.*
|
||
|
|
|
||
|
|
See [UncheatableEval](https://huggingface.co/spaces/Jellyfish042/UncheatableEval) for benchmarks.
|
||
|
|
|
||
|
|
> **Note:** Q4_0, Q3_K_M, and Q2_K were tested but produced completely unintelligible / broken output without an importance matrix. For lower BPW quantizations (below ~4.5 BPW), an imatrix would be required — these quants currently use **no imatrix** at all. Stick to Q4_K_M or higher for usable results.
|