Files changed (1) hide show
  1. README.md +141 -129
README.md CHANGED
@@ -1,129 +1,141 @@
1
- ---
2
- base_model: Qwen/Qwen2.5-72B-Instruct
3
- language:
4
- - en
5
- license: other
6
- license_name: qwen
7
- license_link: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE
8
- pipeline_tag: text-generation
9
- tags:
10
- - chat
11
- quantized_by: bartowski
12
- ---
13
-
14
- ## Llamacpp imatrix Quantizations of Qwen2.5-72B-Instruct
15
-
16
- Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b3772">b3772</a> for quantization.
17
-
18
- Original model: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct
19
-
20
- All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
21
-
22
- Run them in [LM Studio](https://lmstudio.ai/)
23
-
24
- ## Prompt format
25
-
26
- ```
27
- <|im_start|>system
28
- {system_prompt}<|im_end|>
29
- <|im_start|>user
30
- {prompt}<|im_end|>
31
- <|im_start|>assistant
32
- ```
33
-
34
- ## What's new:
35
-
36
- Update context length settings and tokenizer
37
-
38
- ## Download a file (not the whole branch) from below:
39
-
40
- | Filename | Quant type | File Size | Split | Description |
41
- | -------- | ---------- | --------- | ----- | ----------- |
42
- | [Qwen2.5-72B-Instruct-Q8_0.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q8_0) | Q8_0 | 77.26GB | true | Extremely high quality, generally unneeded but max available quant. |
43
- | [Qwen2.5-72B-Instruct-Q6_K.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q6_K) | Q6_K | 64.35GB | true | Very high quality, near perfect, *recommended*. |
44
- | [Qwen2.5-72B-Instruct-Q5_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q5_K_M) | Q5_K_M | 54.45GB | true | High quality, *recommended*. |
45
- | [Qwen2.5-72B-Instruct-Q4_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q4_K_M.gguf) | Q4_K_M | 47.42GB | false | Good quality, default size for must use cases, *recommended*. |
46
- | [Qwen2.5-72B-Instruct-Q4_0.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q4_0.gguf) | Q4_0 | 41.38GB | false | Legacy format, generally not worth using over similarly sized formats |
47
- | [Qwen2.5-72B-Instruct-Q3_K_XL.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_XL.gguf) | Q3_K_XL | 40.60GB | false | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
48
- | [Qwen2.5-72B-Instruct-IQ4_XS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ4_XS.gguf) | IQ4_XS | 39.71GB | false | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
49
- | [Qwen2.5-72B-Instruct-Q3_K_L.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_L.gguf) | Q3_K_L | 39.51GB | false | Lower quality but usable, good for low RAM availability. |
50
- | [Qwen2.5-72B-Instruct-Q3_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_M.gguf) | Q3_K_M | 37.70GB | false | Low quality. |
51
- | [Qwen2.5-72B-Instruct-IQ3_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ3_M.gguf) | IQ3_M | 35.50GB | false | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
52
- | [Qwen2.5-72B-Instruct-Q3_K_S.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_S.gguf) | Q3_K_S | 34.49GB | false | Low quality, not recommended. |
53
- | [Qwen2.5-72B-Instruct-IQ3_XXS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ3_XXS.gguf) | IQ3_XXS | 31.85GB | false | Lower quality, new method with decent performance, comparable to Q3 quants. |
54
- | [Qwen2.5-72B-Instruct-Q2_K_L.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q2_K_L.gguf) | Q2_K_L | 31.03GB | false | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
55
- | [Qwen2.5-72B-Instruct-Q2_K.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q2_K.gguf) | Q2_K | 29.81GB | false | Very low quality but surprisingly usable. |
56
- | [Qwen2.5-72B-Instruct-IQ2_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_M.gguf) | IQ2_M | 29.34GB | false | Relatively low quality, uses SOTA techniques to be surprisingly usable. |
57
- | [Qwen2.5-72B-Instruct-IQ2_XS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_XS.gguf) | IQ2_XS | 27.06GB | false | Low quality, uses SOTA techniques to be usable. |
58
- | [Qwen2.5-72B-Instruct-IQ2_XXS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_XXS.gguf) | IQ2_XXS | 25.49GB | false | Very low quality, uses SOTA techniques to be usable. |
59
- | [Qwen2.5-72B-Instruct-IQ1_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ1_M.gguf) | IQ1_M | 23.74GB | false | Extremely low quality, *not* recommended. |
60
-
61
- ## Embed/output weights
62
-
63
- Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.
64
-
65
- Some say that this improves the quality, others don't notice any difference. If you use these models PLEASE COMMENT with your findings. I would like feedback that these are actually used and useful so I don't keep uploading quants no one is using.
66
-
67
- Thanks!
68
-
69
- ## Downloading using huggingface-cli
70
-
71
- First, make sure you have hugginface-cli installed:
72
-
73
- ```
74
- pip install -U "huggingface_hub[cli]"
75
- ```
76
-
77
- Then, you can target the specific file you want:
78
-
79
- ```
80
- huggingface-cli download bartowski/Qwen2.5-72B-Instruct-GGUF --include "Qwen2.5-72B-Instruct-Q4_K_M.gguf" --local-dir ./
81
- ```
82
-
83
- If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:
84
-
85
- ```
86
- huggingface-cli download bartowski/Qwen2.5-72B-Instruct-GGUF --include "Qwen2.5-72B-Instruct-Q8_0/*" --local-dir ./
87
- ```
88
-
89
- You can either specify a new local-dir (Qwen2.5-72B-Instruct-Q8_0) or download them all in place (./)
90
-
91
- ## Q4_0_X_X
92
-
93
- These are *NOT* for Metal (Apple) offloading, only ARM chips.
94
-
95
- If you're using an ARM chip, the Q4_0_X_X quants will have a substantial speedup. Check out Q4_0_4_4 speed comparisons [on the original pull request](https://github.com/ggerganov/llama.cpp/pull/5780#pullrequestreview-21657544660)
96
-
97
- To check which one would work best for your ARM chip, you can check [AArch64 SoC features](https://gpages.juszkiewicz.com.pl/arm-socs-table/arm-socs.html) (thanks EloyOn!).
98
-
99
- ## Which file should I choose?
100
-
101
- A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
102
-
103
- The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.
104
-
105
- If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.
106
-
107
- If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.
108
-
109
- Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
110
-
111
- If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.
112
-
113
- If you want to get more into the weeds, you can check out this extremely useful feature chart:
114
-
115
- [llama.cpp feature matrix](https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix)
116
-
117
- But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.
118
-
119
- These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
120
-
121
- The I-quants are *not* compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.
122
-
123
- ## Credits
124
-
125
- Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset
126
-
127
- Thank you ZeroWw for the inspiration to experiment with embed/output
128
-
129
- Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: Qwen/Qwen2.5-72B-Instruct
3
+ language:
4
+ - zho
5
+ - eng
6
+ - fra
7
+ - spa
8
+ - por
9
+ - deu
10
+ - ita
11
+ - rus
12
+ - jpn
13
+ - kor
14
+ - vie
15
+ - tha
16
+ - ara
17
+ license: other
18
+ license_name: qwen
19
+ license_link: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE
20
+ pipeline_tag: text-generation
21
+ tags:
22
+ - chat
23
+ quantized_by: bartowski
24
+ ---
25
+
26
+ ## Llamacpp imatrix Quantizations of Qwen2.5-72B-Instruct
27
+
28
+ Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b3772">b3772</a> for quantization.
29
+
30
+ Original model: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct
31
+
32
+ All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
33
+
34
+ Run them in [LM Studio](https://lmstudio.ai/)
35
+
36
+ ## Prompt format
37
+
38
+ ```
39
+ <|im_start|>system
40
+ {system_prompt}<|im_end|>
41
+ <|im_start|>user
42
+ {prompt}<|im_end|>
43
+ <|im_start|>assistant
44
+ ```
45
+
46
+ ## What's new:
47
+
48
+ Update context length settings and tokenizer
49
+
50
+ ## Download a file (not the whole branch) from below:
51
+
52
+ | Filename | Quant type | File Size | Split | Description |
53
+ | -------- | ---------- | --------- | ----- | ----------- |
54
+ | [Qwen2.5-72B-Instruct-Q8_0.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q8_0) | Q8_0 | 77.26GB | true | Extremely high quality, generally unneeded but max available quant. |
55
+ | [Qwen2.5-72B-Instruct-Q6_K.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q6_K) | Q6_K | 64.35GB | true | Very high quality, near perfect, *recommended*. |
56
+ | [Qwen2.5-72B-Instruct-Q5_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/tree/main/Qwen2.5-72B-Instruct-Q5_K_M) | Q5_K_M | 54.45GB | true | High quality, *recommended*. |
57
+ | [Qwen2.5-72B-Instruct-Q4_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q4_K_M.gguf) | Q4_K_M | 47.42GB | false | Good quality, default size for must use cases, *recommended*. |
58
+ | [Qwen2.5-72B-Instruct-Q4_0.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q4_0.gguf) | Q4_0 | 41.38GB | false | Legacy format, generally not worth using over similarly sized formats |
59
+ | [Qwen2.5-72B-Instruct-Q3_K_XL.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_XL.gguf) | Q3_K_XL | 40.60GB | false | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
60
+ | [Qwen2.5-72B-Instruct-IQ4_XS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ4_XS.gguf) | IQ4_XS | 39.71GB | false | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
61
+ | [Qwen2.5-72B-Instruct-Q3_K_L.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_L.gguf) | Q3_K_L | 39.51GB | false | Lower quality but usable, good for low RAM availability. |
62
+ | [Qwen2.5-72B-Instruct-Q3_K_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_M.gguf) | Q3_K_M | 37.70GB | false | Low quality. |
63
+ | [Qwen2.5-72B-Instruct-IQ3_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ3_M.gguf) | IQ3_M | 35.50GB | false | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
64
+ | [Qwen2.5-72B-Instruct-Q3_K_S.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q3_K_S.gguf) | Q3_K_S | 34.49GB | false | Low quality, not recommended. |
65
+ | [Qwen2.5-72B-Instruct-IQ3_XXS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ3_XXS.gguf) | IQ3_XXS | 31.85GB | false | Lower quality, new method with decent performance, comparable to Q3 quants. |
66
+ | [Qwen2.5-72B-Instruct-Q2_K_L.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q2_K_L.gguf) | Q2_K_L | 31.03GB | false | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
67
+ | [Qwen2.5-72B-Instruct-Q2_K.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-Q2_K.gguf) | Q2_K | 29.81GB | false | Very low quality but surprisingly usable. |
68
+ | [Qwen2.5-72B-Instruct-IQ2_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_M.gguf) | IQ2_M | 29.34GB | false | Relatively low quality, uses SOTA techniques to be surprisingly usable. |
69
+ | [Qwen2.5-72B-Instruct-IQ2_XS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_XS.gguf) | IQ2_XS | 27.06GB | false | Low quality, uses SOTA techniques to be usable. |
70
+ | [Qwen2.5-72B-Instruct-IQ2_XXS.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ2_XXS.gguf) | IQ2_XXS | 25.49GB | false | Very low quality, uses SOTA techniques to be usable. |
71
+ | [Qwen2.5-72B-Instruct-IQ1_M.gguf](https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF/blob/main/Qwen2.5-72B-Instruct-IQ1_M.gguf) | IQ1_M | 23.74GB | false | Extremely low quality, *not* recommended. |
72
+
73
+ ## Embed/output weights
74
+
75
+ Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.
76
+
77
+ Some say that this improves the quality, others don't notice any difference. If you use these models PLEASE COMMENT with your findings. I would like feedback that these are actually used and useful so I don't keep uploading quants no one is using.
78
+
79
+ Thanks!
80
+
81
+ ## Downloading using huggingface-cli
82
+
83
+ First, make sure you have hugginface-cli installed:
84
+
85
+ ```
86
+ pip install -U "huggingface_hub[cli]"
87
+ ```
88
+
89
+ Then, you can target the specific file you want:
90
+
91
+ ```
92
+ huggingface-cli download bartowski/Qwen2.5-72B-Instruct-GGUF --include "Qwen2.5-72B-Instruct-Q4_K_M.gguf" --local-dir ./
93
+ ```
94
+
95
+ If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:
96
+
97
+ ```
98
+ huggingface-cli download bartowski/Qwen2.5-72B-Instruct-GGUF --include "Qwen2.5-72B-Instruct-Q8_0/*" --local-dir ./
99
+ ```
100
+
101
+ You can either specify a new local-dir (Qwen2.5-72B-Instruct-Q8_0) or download them all in place (./)
102
+
103
+ ## Q4_0_X_X
104
+
105
+ These are *NOT* for Metal (Apple) offloading, only ARM chips.
106
+
107
+ If you're using an ARM chip, the Q4_0_X_X quants will have a substantial speedup. Check out Q4_0_4_4 speed comparisons [on the original pull request](https://github.com/ggerganov/llama.cpp/pull/5780#pullrequestreview-21657544660)
108
+
109
+ To check which one would work best for your ARM chip, you can check [AArch64 SoC features](https://gpages.juszkiewicz.com.pl/arm-socs-table/arm-socs.html) (thanks EloyOn!).
110
+
111
+ ## Which file should I choose?
112
+
113
+ A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
114
+
115
+ The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.
116
+
117
+ If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.
118
+
119
+ If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.
120
+
121
+ Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
122
+
123
+ If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.
124
+
125
+ If you want to get more into the weeds, you can check out this extremely useful feature chart:
126
+
127
+ [llama.cpp feature matrix](https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix)
128
+
129
+ But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.
130
+
131
+ These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
132
+
133
+ The I-quants are *not* compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.
134
+
135
+ ## Credits
136
+
137
+ Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset
138
+
139
+ Thank you ZeroWw for the inspiration to experiment with embed/output
140
+
141
+ Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski