Hugging Face's logo

bartowski
/
burtenshaw_GemmaCoder3-12B-GGUF

Text Generation
GGUF
Generated from Trainer
trl
sft
conversational
Model card Files
xet
Community
bartowski commited on
Commit
01056b6
·
verified ·
1 Parent(s): 7155020

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +171 -0
README.md ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ quantized_by: bartowski
3
+ pipeline_tag: text-generation
4
+ ---
5
+
6
+ ## Llamacpp imatrix Quantizations of GemmaCoder3-12B by burtenshaw
7
+
8
+ Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b5010">b5010</a> for quantization.
9
+
10
+ Original model: https://huggingface.co/burtenshaw/GemmaCoder3-12B
11
+
12
+ All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
13
+
14
+ Run them in [LM Studio](https://lmstudio.ai/)
15
+
16
+ Run them directly with [llama.cpp](https://github.com/ggerganov/llama.cpp), or any other llama.cpp based project
17
+
18
+ ## Prompt format
19
+
20
+ ```
21
+ <bos><start_of_turn>user
22
+ {system_prompt}
23
+
24
+ {prompt}<end_of_turn>
25
+ <start_of_turn>model
26
+ <end_of_turn>
27
+ <start_of_turn>model
28
+ ```
29
+
30
+ ## Download a file (not the whole branch) from below:
31
+
32
+ | Filename | Quant type | File Size | Split | Description |
33
+ | -------- | ---------- | --------- | ----- | ----------- |
34
+ | [GemmaCoder3-12B-bf16.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-bf16.gguf) | bf16 | 23.54GB | false | Full BF16 weights. |
35
+ | [GemmaCoder3-12B-Q8_0.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q8_0.gguf) | Q8_0 | 12.51GB | false | Extremely high quality, generally unneeded but max available quant. |
36
+ | [GemmaCoder3-12B-Q6_K_L.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q6_K_L.gguf) | Q6_K_L | 9.90GB | false | Uses Q8_0 for embed and output weights. Very high quality, near perfect, *recommended*. |
37
+ | [GemmaCoder3-12B-Q6_K.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q6_K.gguf) | Q6_K | 9.66GB | false | Very high quality, near perfect, *recommended*. |
38
+ | [GemmaCoder3-12B-Q5_K_L.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q5_K_L.gguf) | Q5_K_L | 8.69GB | false | Uses Q8_0 for embed and output weights. High quality, *recommended*. |
39
+ | [GemmaCoder3-12B-Q5_K_M.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q5_K_M.gguf) | Q5_K_M | 8.45GB | false | High quality, *recommended*. |
40
+ | [GemmaCoder3-12B-Q5_K_S.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q5_K_S.gguf) | Q5_K_S | 8.23GB | false | High quality, *recommended*. |
41
+ | [GemmaCoder3-12B-Q4_1.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q4_1.gguf) | Q4_1 | 7.56GB | false | Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon. |
42
+ | [GemmaCoder3-12B-Q4_K_L.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q4_K_L.gguf) | Q4_K_L | 7.54GB | false | Uses Q8_0 for embed and output weights. Good quality, *recommended*. |
43
+ | [GemmaCoder3-12B-Q4_K_M.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q4_K_M.gguf) | Q4_K_M | 7.30GB | false | Good quality, default size for most use cases, *recommended*. |
44
+ | [GemmaCoder3-12B-Q4_K_S.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q4_K_S.gguf) | Q4_K_S | 6.94GB | false | Slightly lower quality with more space savings, *recommended*. |
45
+ | [GemmaCoder3-12B-Q4_0.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q4_0.gguf) | Q4_0 | 6.91GB | false | Legacy format, offers online repacking for ARM and AVX CPU inference. |
46
+ | [GemmaCoder3-12B-IQ4_NL.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ4_NL.gguf) | IQ4_NL | 6.89GB | false | Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference. |
47
+ | [GemmaCoder3-12B-Q3_K_XL.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q3_K_XL.gguf) | Q3_K_XL | 6.72GB | false | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
48
+ | [GemmaCoder3-12B-IQ4_XS.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ4_XS.gguf) | IQ4_XS | 6.55GB | false | Decent quality, smaller than Q4_K_S with similar performance, *recommended*. |
49
+ | [GemmaCoder3-12B-Q3_K_L.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q3_K_L.gguf) | Q3_K_L | 6.48GB | false | Lower quality but usable, good for low RAM availability. |
50
+ | [GemmaCoder3-12B-Q3_K_M.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q3_K_M.gguf) | Q3_K_M | 6.01GB | false | Low quality. |
51
+ | [GemmaCoder3-12B-IQ3_M.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ3_M.gguf) | IQ3_M | 5.66GB | false | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
52
+ | [GemmaCoder3-12B-Q3_K_S.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q3_K_S.gguf) | Q3_K_S | 5.46GB | false | Low quality, not recommended. |
53
+ | [GemmaCoder3-12B-IQ3_XS.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ3_XS.gguf) | IQ3_XS | 5.21GB | false | Lower quality, new method with decent performance, slightly better than Q3_K_S. |
54
+ | [GemmaCoder3-12B-Q2_K_L.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q2_K_L.gguf) | Q2_K_L | 5.01GB | false | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
55
+ | [GemmaCoder3-12B-IQ3_XXS.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ3_XXS.gguf) | IQ3_XXS | 4.78GB | false | Lower quality, new method with decent performance, comparable to Q3 quants. |
56
+ | [GemmaCoder3-12B-Q2_K.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-Q2_K.gguf) | Q2_K | 4.77GB | false | Very low quality but surprisingly usable. |
57
+ | [GemmaCoder3-12B-IQ2_M.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ2_M.gguf) | IQ2_M | 4.31GB | false | Relatively low quality, uses SOTA techniques to be surprisingly usable. |
58
+ | [GemmaCoder3-12B-IQ2_S.gguf](https://huggingface.co/bartowski/burtenshaw_GemmaCoder3-12B-GGUF/blob/main/burtenshaw_GemmaCoder3-12B-IQ2_S.gguf) | IQ2_S | 4.02GB | false | Low quality, uses SOTA techniques to be usable. |
59
+
60
+ ## Embed/output weights
61
+
62
+ 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.
63
+
64
+ ## Downloading using huggingface-cli
65
+
66
+ <details>
67
+ <summary>Click to view download instructions</summary>
68
+
69
+ First, make sure you have hugginface-cli installed:
70
+
71
+ ```
72
+ pip install -U "huggingface_hub[cli]"
73
+ ```
74
+
75
+ Then, you can target the specific file you want:
76
+
77
+ ```
78
+ huggingface-cli download bartowski/burtenshaw_GemmaCoder3-12B-GGUF --include "burtenshaw_GemmaCoder3-12B-Q4_K_M.gguf" --local-dir ./
79
+ ```
80
+
81
+ 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:
82
+
83
+ ```
84
+ huggingface-cli download bartowski/burtenshaw_GemmaCoder3-12B-GGUF --include "burtenshaw_GemmaCoder3-12B-Q8_0/*" --local-dir ./
85
+ ```
86
+
87
+ You can either specify a new local-dir (burtenshaw_GemmaCoder3-12B-Q8_0) or download them all in place (./)
88
+
89
+ </details>
90
+
91
+ ## ARM/AVX information
92
+
93
+ Previously, you would download Q4_0_4_4/4_8/8_8, and these would have their weights interleaved in memory in order to improve performance on ARM and AVX machines by loading up more data in one pass.
94
+
95
+ Now, however, there is something called "online repacking" for weights. details in [this PR](https://github.com/ggerganov/llama.cpp/pull/9921). If you use Q4_0 and your hardware would benefit from repacking weights, it will do it automatically on the fly.
96
+
97
+ As of llama.cpp build [b4282](https://github.com/ggerganov/llama.cpp/releases/tag/b4282) you will not be able to run the Q4_0_X_X files and will instead need to use Q4_0.
98
+
99
+ Additionally, if you want to get slightly better quality for , you can use IQ4_NL thanks to [this PR](https://github.com/ggerganov/llama.cpp/pull/10541) which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed incrase.
100
+
101
+ <details>
102
+ <summary>Click to view Q4_0_X_X information (deprecated</summary>
103
+
104
+ I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.
105
+
106
+ <details>
107
+ <summary>Click to view benchmarks on an AVX2 system (EPYC7702)</summary>
108
+
109
+ | model | size | params | backend | threads | test | t/s | % (vs Q4_0) |
110
+ | ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |-------------: |
111
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp512 | 204.03 ± 1.03 | 100% |
112
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp1024 | 282.92 ± 0.19 | 100% |
113
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp2048 | 259.49 ± 0.44 | 100% |
114
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg128 | 39.12 ± 0.27 | 100% |
115
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg256 | 39.31 ± 0.69 | 100% |
116
+ | qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg512 | 40.52 ± 0.03 | 100% |
117
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp512 | 301.02 ± 1.74 | 147% |
118
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp1024 | 287.23 ± 0.20 | 101% |
119
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp2048 | 262.77 ± 1.81 | 101% |
120
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg128 | 18.80 ± 0.99 | 48% |
121
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg256 | 24.46 ± 3.04 | 83% |
122
+ | qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg512 | 36.32 ± 3.59 | 90% |
123
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp512 | 271.71 ± 3.53 | 133% |
124
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp1024 | 279.86 ± 45.63 | 100% |
125
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp2048 | 320.77 ± 5.00 | 124% |
126
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg128 | 43.51 ± 0.05 | 111% |
127
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg256 | 43.35 ± 0.09 | 110% |
128
+ | qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg512 | 42.60 ± 0.31 | 105% |
129
+
130
+ Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation
131
+
132
+ </details>
133
+
134
+ </details>
135
+
136
+ ## Which file should I choose?
137
+
138
+ <details>
139
+ <summary>Click here for details</summary>
140
+
141
+ A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
142
+
143
+ 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.
144
+
145
+ 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.
146
+
147
+ 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.
148
+
149
+ Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
150
+
151
+ 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.
152
+
153
+ If you want to get more into the weeds, you can check out this extremely useful feature chart:
154
+
155
+ [llama.cpp feature matrix](https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix)
156
+
157
+ 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.
158
+
159
+ These I-quants can also be used on CPU, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
160
+
161
+ </details>
162
+
163
+ ## Credits
164
+
165
+ Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset.
166
+
167
+ Thank you ZeroWw for the inspiration to experiment with embed/output.
168
+
169
+ Thank you to LM Studio for sponsoring my work.
170
+
171
+ Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski