TheBloke commited on
Commit
2f79dee
1 Parent(s): dbfafad

New GGMLv3 format for breaking llama.cpp change May 19th commit 2d5db48

Browse files
Files changed (1) hide show
  1. README.md +13 -9
README.md CHANGED
@@ -21,27 +21,29 @@ This repo is the result of quantising to 4bit and 5bit GGML for CPU inference us
21
  * [4-bit, 5-bit and 8-bit GGML models for CPU (+CUDA) inference](https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-GGML).
22
  * [float16 HF format model for GPU inference and further conversions](https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-HF).
23
 
24
- ## THE FILES IN MAIN BRANCH REQUIRES LATEST LLAMA.CPP (May 12th 2023 - commit b9fd7ee)!
25
 
26
- llama.cpp recently made a breaking change to its quantisation methods.
27
 
28
- I have quantised the GGML files in this repo with the latest version. Therefore you will require llama.cpp compiled on May 12th or later (commit `b9fd7ee` or later) to use them.
 
 
29
 
30
  ## Provided files
31
  | Name | Quant method | Bits | Size | RAM required | Use case |
32
  | ---- | ---- | ---- | ---- | ---- | ----- |
33
- `Wizard-Vicuna-7B-Uncensored.ggmlv2.q4_0.bin` | q4_0 | 4bit | 4.21GB | 7.0GB | 4-bit. |
34
- `Wizard-Vicuna-7B-Uncensored.ggmlv2.q4_1.bin` | q4_1 | 4bit | 4.63GB | 7.5GB | 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
35
- `Wizard-Vicuna-7B-Uncensored.ggmlv2.q5_0.bin` | q5_0 | 5bit | 4.63GB | 7.5GB | 5-bit. Higher accuracy, higher resource usage and slower inference. |
36
- `Wizard-Vicuna-7B-Uncensored.ggmlv2.q5_1.bin` | q5_1 | 5bit | 5.06GB | 7.5GB | 5-bit. Even higher accuracy, and higher resource usage and slower inference. |
37
- `Wizard-Vicuna-7B-Uncensored.ggmlv2.q8_0.bin` | q8_0 | 8bit | 7.58GB | 9.0GB | 8-bit. Almost indistinguishable from float16. Huge resource use and slow. Not recommended for normal use. |
38
 
39
  ## How to run in `llama.cpp`
40
 
41
  I use the following command line; adjust for your tastes and needs:
42
 
43
  ```
44
- ./main -t 8 -m Wizard-Vicuna-7B-Uncensored.ggmlv2.q5_0.bin --color -c 2048 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "### Instruction: write a story about llamas ### Response:"
45
  ```
46
 
47
  Change `-t 8` to the number of physical CPU cores you have.
@@ -52,6 +54,8 @@ GGML models can be loaded into text-generation-webui by installing the llama.cpp
52
 
53
  Further instructions here: [text-generation-webui/docs/llama.cpp-models.md](https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md).
54
 
 
 
55
  # Original model card
56
 
57
  This is [wizard-vicuna-13b](https://huggingface.co/junelee/wizard-vicuna-13b) trained against LLaMA-7B with a subset of the dataset - responses that contained alignment / moralizing were removed. The intent is to train a WizardLM that doesn't have alignment built-in, so that alignment (of any sort) can be added separately with for example with a RLHF LoRA.
 
21
  * [4-bit, 5-bit and 8-bit GGML models for CPU (+CUDA) inference](https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-GGML).
22
  * [float16 HF format model for GPU inference and further conversions](https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-HF).
23
 
24
+ ## THE FILES IN MAIN BRANCH REQUIRES LATEST LLAMA.CPP (May 19th 2023 - commit 2d5db48)!
25
 
26
+ llama.cpp recently made another breaking change to its quantisation methods - https://github.com/ggerganov/llama.cpp/pull/1508
27
 
28
+ I have quantised the GGML files in this repo with the latest version. Therefore you will require llama.cpp compiled on May 19th or later (commit `2d5db48` or later) to use them.
29
+
30
+ For files compatible with the previous version of llama.cpp, please see branch `previous_llama_ggmlv2`.
31
 
32
  ## Provided files
33
  | Name | Quant method | Bits | Size | RAM required | Use case |
34
  | ---- | ---- | ---- | ---- | ---- | ----- |
35
+ `Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_0.bin` | q4_0 | 4bit | 4.21GB | 7.0GB | 4-bit. |
36
+ `Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin` | q4_1 | 4bit | 4.63GB | 7.5GB | 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
37
+ `Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_0.bin` | q5_0 | 5bit | 4.63GB | 7.5GB | 5-bit. Higher accuracy, higher resource usage and slower inference. |
38
+ `Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_1.bin` | q5_1 | 5bit | 5.06GB | 7.5GB | 5-bit. Even higher accuracy, and higher resource usage and slower inference. |
39
+ `Wizard-Vicuna-7B-Uncensored.ggmlv3.q8_0.bin` | q8_0 | 8bit | 7.58GB | 9.0GB | 8-bit. Almost indistinguishable from float16. Huge resource use and slow. Not recommended for normal use. |
40
 
41
  ## How to run in `llama.cpp`
42
 
43
  I use the following command line; adjust for your tastes and needs:
44
 
45
  ```
46
+ ./main -t 8 -m Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_0.bin --color -c 2048 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "### Instruction: write a story about llamas ### Response:"
47
  ```
48
 
49
  Change `-t 8` to the number of physical CPU cores you have.
 
54
 
55
  Further instructions here: [text-generation-webui/docs/llama.cpp-models.md](https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md).
56
 
57
+ Note: at this time text-generation-webui may not support the new May 19th llama.cpp quantisation methods for q4_0, q4_1 and q8_0 files.
58
+
59
  # Original model card
60
 
61
  This is [wizard-vicuna-13b](https://huggingface.co/junelee/wizard-vicuna-13b) trained against LLaMA-7B with a subset of the dataset - responses that contained alignment / moralizing were removed. The intent is to train a WizardLM that doesn't have alignment built-in, so that alignment (of any sort) can be added separately with for example with a RLHF LoRA.