ed
/
rwkv.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Optimize Q4_1_O by moving outlier multiplication out of the dequantize+dot loop

This commit is contained in:
saharNooby 2023-04-18 09:47:20 +04:00
parent 0a8157d1ee
commit 2ef7ee0fac
2 changed files with 7 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
README.md
View File

@ -12,10 +12,9 @@ Loading LoRA checkpoints in [Blealtan's format](https://github.com/Blealtan/RWKV
**TODO (contributions welcome!)**: **TODO (contributions welcome!)**:
1. Optimize AVX2 implementation of `Q4_1_O` matmul — currently, it is 40% slower than `Q4_1` 1. Measure latency and perplexity of different model sizes (169M to 14B) and data types (`FP32`, `FP16`, `Q4_0`, `Q4_1`, `Q4_1_O`)
2. Measure latency and perplexity of different model sizes (169M to 14B) and data types (`FP32`, `FP16`, `Q4_0`, `Q4_1`, `Q4_1_O`) 2. Test on Linux (including Colab) and MacOS
3. Test on Linux (including Colab) and MacOS 3. Make required memory calculation more robust (see [#4](https://github.com/saharNooby/rwkv.cpp/issues/4))
4. Make required memory calculation more robust (see [#4](https://github.com/saharNooby/rwkv.cpp/issues/4))
## How to use ## How to use
@ -91,9 +90,9 @@ python rwkv/quantize.py ~/Downloads/rwkv.cpp-169M.bin ~/Downloads/rwkv.cpp-169M-
Formats available: Formats available:
- `4`: `Q4_1_O`, best quality, slow (30% slower than `FP16`). - `4`: `Q4_1_O`, OK quality, fast (comparable to `FP16`).
- `3`: `Q4_1`, poor quality, fast (comparable to `FP16`). - `3`: `Q4_1`, worst quality, fast (comparable to `FP16`).
- `2`: `Q4_0`, worst quality, breaks larger models, very fast. - `2`: `Q4_0`, poor quality, very fast.
### 4. Run the model ### 4. Run the model

2
ggml

@ -1 +1 @@
Subproject commit 4856b5d30a8985cde31efe4a5bfcc77ac1df61a6 Subproject commit fbf4d6052fd2df028169a5609a4f45fbbdf6eece