TheBloke commited on
Commit
10dcb2d
1 Parent(s): b94eeb0

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +550 -0
README.md ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: brucethemoose/Yi-34B-200K-DARE-megamerge-v8
3
+ inference: false
4
+ language:
5
+ - en
6
+ library_name: transformers
7
+ license: other
8
+ license_link: https://huggingface.co/01-ai/Yi-34B/blob/main/LICENSE
9
+ license_name: yi-license
10
+ model_creator: brucethemoose
11
+ model_name: Yi 34B 200K DARE MegaMerge V8
12
+ model_type: yi
13
+ prompt_template: 'SYSTEM: {system_message}
14
+
15
+ USER: {prompt}
16
+
17
+ ASSISTANT:
18
+
19
+ '
20
+ quantized_by: TheBloke
21
+ tags:
22
+ - mergekit
23
+ - merge
24
+ - Yi
25
+ ---
26
+ <!-- markdownlint-disable MD041 -->
27
+
28
+ <!-- header start -->
29
+ <!-- 200823 -->
30
+ <div style="width: auto; margin-left: auto; margin-right: auto">
31
+ <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
32
+ </div>
33
+ <div style="display: flex; justify-content: space-between; width: 100%;">
34
+ <div style="display: flex; flex-direction: column; align-items: flex-start;">
35
+ <p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://discord.gg/theblokeai">Chat & support: TheBloke's Discord server</a></p>
36
+ </div>
37
+ <div style="display: flex; flex-direction: column; align-items: flex-end;">
38
+ <p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p>
39
+ </div>
40
+ </div>
41
+ <div style="text-align:center; margin-top: 0em; margin-bottom: 0em"><p style="margin-top: 0.25em; margin-bottom: 0em;">TheBloke's LLM work is generously supported by a grant from <a href="https://a16z.com">andreessen horowitz (a16z)</a></p></div>
42
+ <hr style="margin-top: 1.0em; margin-bottom: 1.0em;">
43
+ <!-- header end -->
44
+
45
+ # Yi 34B 200K DARE MegaMerge V8 - GPTQ
46
+ - Model creator: [brucethemoose](https://huggingface.co/brucethemoose)
47
+ - Original model: [Yi 34B 200K DARE MegaMerge V8](https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-megamerge-v8)
48
+
49
+ <!-- description start -->
50
+ # Description
51
+
52
+ This repo contains GPTQ model files for [brucethemoose's Yi 34B 200K DARE MegaMerge V8](https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-megamerge-v8).
53
+
54
+ Multiple GPTQ parameter permutations are provided; see Provided Files below for details of the options provided, their parameters, and the software used to create them.
55
+
56
+ <!-- description end -->
57
+ <!-- repositories-available start -->
58
+ ## Repositories available
59
+
60
+ * [AWQ model(s) for GPU inference.](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-AWQ)
61
+ * [GPTQ models for GPU inference, with multiple quantisation parameter options.](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ)
62
+ * [2, 3, 4, 5, 6 and 8-bit GGUF models for CPU+GPU inference](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GGUF)
63
+ * [brucethemoose's original unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-megamerge-v8)
64
+ <!-- repositories-available end -->
65
+
66
+ <!-- prompt-template start -->
67
+ ## Prompt template: Orca-Vicuna
68
+
69
+ ```
70
+ SYSTEM: {system_message}
71
+ USER: {prompt}
72
+ ASSISTANT:
73
+
74
+ ```
75
+
76
+ <!-- prompt-template end -->
77
+
78
+
79
+
80
+ <!-- README_GPTQ.md-compatible clients start -->
81
+ ## Known compatible clients / servers
82
+
83
+ GPTQ models are currently supported on Linux (NVidia/AMD) and Windows (NVidia only). macOS users: please use GGUF models.
84
+
85
+ These GPTQ models are known to work in the following inference servers/webuis.
86
+
87
+ - [text-generation-webui](https://github.com/oobabooga/text-generation-webui)
88
+ - [KoboldAI United](https://github.com/henk717/koboldai)
89
+ - [LoLLMS Web UI](https://github.com/ParisNeo/lollms-webui)
90
+ - [Hugging Face Text Generation Inference (TGI)](https://github.com/huggingface/text-generation-inference)
91
+
92
+ This may not be a complete list; if you know of others, please let me know!
93
+ <!-- README_GPTQ.md-compatible clients end -->
94
+
95
+ <!-- README_GPTQ.md-provided-files start -->
96
+ ## Provided files, and GPTQ parameters
97
+
98
+ Multiple quantisation parameters are provided, to allow you to choose the best one for your hardware and requirements.
99
+
100
+ Each separate quant is in a different branch. See below for instructions on fetching from different branches.
101
+
102
+ Most GPTQ files are made with AutoGPTQ. Mistral models are currently made with Transformers.
103
+
104
+ <details>
105
+ <summary>Explanation of GPTQ parameters</summary>
106
+
107
+ - Bits: The bit size of the quantised model.
108
+ - GS: GPTQ group size. Higher numbers use less VRAM, but have lower quantisation accuracy. "None" is the lowest possible value.
109
+ - Act Order: True or False. Also known as `desc_act`. True results in better quantisation accuracy. Some GPTQ clients have had issues with models that use Act Order plus Group Size, but this is generally resolved now.
110
+ - Damp %: A GPTQ parameter that affects how samples are processed for quantisation. 0.01 is default, but 0.1 results in slightly better accuracy.
111
+ - GPTQ dataset: The calibration dataset used during quantisation. Using a dataset more appropriate to the model's training can improve quantisation accuracy. Note that the GPTQ calibration dataset is not the same as the dataset used to train the model - please refer to the original model repo for details of the training dataset(s).
112
+ - Sequence Length: The length of the dataset sequences used for quantisation. Ideally this is the same as the model sequence length. For some very long sequence models (16+K), a lower sequence length may have to be used. Note that a lower sequence length does not limit the sequence length of the quantised model. It only impacts the quantisation accuracy on longer inference sequences.
113
+ - ExLlama Compatibility: Whether this file can be loaded with ExLlama, which currently only supports Llama and Mistral models in 4-bit.
114
+
115
+ </details>
116
+
117
+ | Branch | Bits | GS | Act Order | Damp % | GPTQ Dataset | Seq Len | Size | ExLlama | Desc |
118
+ | ------ | ---- | -- | --------- | ------ | ------------ | ------- | ---- | ------- | ---- |
119
+ | [main](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/main) | 4 | None | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 18.60 GB | Yes | 4-bit, with Act Order. No group size, to lower VRAM requirements. |
120
+ | [gptq-4bit-128g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-4bit-128g-actorder_True) | 4 | 128 | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 19.25 GB | Yes | 4-bit, with Act Order and group size 128g. Uses even less VRAM than 64g, but with slightly lower accuracy. |
121
+ | [gptq-4bit-32g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-4bit-32g-actorder_True) | 4 | 32 | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 21.21 GB | Yes | 4-bit, with Act Order and group size 32g. Gives highest possible inference quality, with maximum VRAM usage. |
122
+ | [gptq-3bit-128g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-3bit-128g-actorder_True) | 3 | 128 | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 15.03 GB | No | 3-bit, with group size 128g and act-order. Higher quality than 128g-False. |
123
+ | [gptq-8bit--1g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-8bit--1g-actorder_True) | 8 | None | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 35.34 GB | No | 8-bit, with Act Order. No group size, to lower VRAM requirements. |
124
+ | [gptq-3bit-32g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-3bit-32g-actorder_True) | 3 | 32 | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 16.90 GB | No | 3-bit, with group size 64g and act-order. Highest quality 3-bit option. |
125
+ | [gptq-8bit-128g-actorder_True](https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ/tree/gptq-8bit-128g-actorder_True) | 8 | 128 | Yes | 0.1 | [VMware Open Instruct](https://huggingface.co/datasets/VMware/open-instruct/viewer/) | 4096 | 36.12 GB | No | 8-bit, with group size 128g for higher inference quality and with Act Order for even higher accuracy. |
126
+
127
+ <!-- README_GPTQ.md-provided-files end -->
128
+
129
+ <!-- README_GPTQ.md-download-from-branches start -->
130
+ ## How to download, including from branches
131
+
132
+ ### In text-generation-webui
133
+
134
+ To download from the `main` branch, enter `TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ` in the "Download model" box.
135
+
136
+ To download from another branch, add `:branchname` to the end of the download name, eg `TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ:gptq-4bit-128g-actorder_True`
137
+
138
+ ### From the command line
139
+
140
+ I recommend using the `huggingface-hub` Python library:
141
+
142
+ ```shell
143
+ pip3 install huggingface-hub
144
+ ```
145
+
146
+ To download the `main` branch to a folder called `Yi-34B-200K-DARE-megamerge-v8-GPTQ`:
147
+
148
+ ```shell
149
+ mkdir Yi-34B-200K-DARE-megamerge-v8-GPTQ
150
+ huggingface-cli download TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ --local-dir Yi-34B-200K-DARE-megamerge-v8-GPTQ --local-dir-use-symlinks False
151
+ ```
152
+
153
+ To download from a different branch, add the `--revision` parameter:
154
+
155
+ ```shell
156
+ mkdir Yi-34B-200K-DARE-megamerge-v8-GPTQ
157
+ huggingface-cli download TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ --revision gptq-4bit-128g-actorder_True --local-dir Yi-34B-200K-DARE-megamerge-v8-GPTQ --local-dir-use-symlinks False
158
+ ```
159
+
160
+ <details>
161
+ <summary>More advanced huggingface-cli download usage</summary>
162
+
163
+ If you remove the `--local-dir-use-symlinks False` parameter, the files will instead be stored in the central Hugging Face cache directory (default location on Linux is: `~/.cache/huggingface`), and symlinks will be added to the specified `--local-dir`, pointing to their real location in the cache. This allows for interrupted downloads to be resumed, and allows you to quickly clone the repo to multiple places on disk without triggering a download again. The downside, and the reason why I don't list that as the default option, is that the files are then hidden away in a cache folder and it's harder to know where your disk space is being used, and to clear it up if/when you want to remove a download model.
164
+
165
+ The cache location can be changed with the `HF_HOME` environment variable, and/or the `--cache-dir` parameter to `huggingface-cli`.
166
+
167
+ For more documentation on downloading with `huggingface-cli`, please see: [HF -> Hub Python Library -> Download files -> Download from the CLI](https://huggingface.co/docs/huggingface_hub/guides/download#download-from-the-cli).
168
+
169
+ To accelerate downloads on fast connections (1Gbit/s or higher), install `hf_transfer`:
170
+
171
+ ```shell
172
+ pip3 install hf_transfer
173
+ ```
174
+
175
+ And set environment variable `HF_HUB_ENABLE_HF_TRANSFER` to `1`:
176
+
177
+ ```shell
178
+ mkdir Yi-34B-200K-DARE-megamerge-v8-GPTQ
179
+ HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ --local-dir Yi-34B-200K-DARE-megamerge-v8-GPTQ --local-dir-use-symlinks False
180
+ ```
181
+
182
+ Windows Command Line users: You can set the environment variable by running `set HF_HUB_ENABLE_HF_TRANSFER=1` before the download command.
183
+ </details>
184
+
185
+ ### With `git` (**not** recommended)
186
+
187
+ To clone a specific branch with `git`, use a command like this:
188
+
189
+ ```shell
190
+ git clone --single-branch --branch gptq-4bit-128g-actorder_True https://huggingface.co/TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ
191
+ ```
192
+
193
+ Note that using Git with HF repos is strongly discouraged. It will be much slower than using `huggingface-hub`, and will use twice as much disk space as it has to store the model files twice (it stores every byte both in the intended target folder, and again in the `.git` folder as a blob.)
194
+
195
+ <!-- README_GPTQ.md-download-from-branches end -->
196
+ <!-- README_GPTQ.md-text-generation-webui start -->
197
+ ## How to easily download and use this model in [text-generation-webui](https://github.com/oobabooga/text-generation-webui)
198
+
199
+ Please make sure you're using the latest version of [text-generation-webui](https://github.com/oobabooga/text-generation-webui).
200
+
201
+ It is strongly recommended to use the text-generation-webui one-click-installers unless you're sure you know how to make a manual install.
202
+
203
+ 1. Click the **Model tab**.
204
+ 2. Under **Download custom model or LoRA**, enter `TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ`.
205
+
206
+ - To download from a specific branch, enter for example `TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ:gptq-4bit-128g-actorder_True`
207
+ - see Provided Files above for the list of branches for each option.
208
+
209
+ 3. Click **Download**.
210
+ 4. The model will start downloading. Once it's finished it will say "Done".
211
+ 5. In the top left, click the refresh icon next to **Model**.
212
+ 6. In the **Model** dropdown, choose the model you just downloaded: `Yi-34B-200K-DARE-megamerge-v8-GPTQ`
213
+ 7. The model will automatically load, and is now ready for use!
214
+ 8. If you want any custom settings, set them and then click **Save settings for this model** followed by **Reload the Model** in the top right.
215
+
216
+ - Note that you do not need to and should not set manual GPTQ parameters any more. These are set automatically from the file `quantize_config.json`.
217
+
218
+ 9. Once you're ready, click the **Text Generation** tab and enter a prompt to get started!
219
+
220
+ <!-- README_GPTQ.md-text-generation-webui end -->
221
+
222
+ <!-- README_GPTQ.md-use-from-tgi start -->
223
+ ## Serving this model from Text Generation Inference (TGI)
224
+
225
+ It's recommended to use TGI version 1.1.0 or later. The official Docker container is: `ghcr.io/huggingface/text-generation-inference:1.1.0`
226
+
227
+ Example Docker parameters:
228
+
229
+ ```shell
230
+ --model-id TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ --port 3000 --quantize gptq --max-input-length 3696 --max-total-tokens 4096 --max-batch-prefill-tokens 4096
231
+ ```
232
+
233
+ Example Python code for interfacing with TGI (requires huggingface-hub 0.17.0 or later):
234
+
235
+ ```shell
236
+ pip3 install huggingface-hub
237
+ ```
238
+
239
+ ```python
240
+ from huggingface_hub import InferenceClient
241
+
242
+ endpoint_url = "https://your-endpoint-url-here"
243
+
244
+ prompt = "Tell me about AI"
245
+ prompt_template=f'''SYSTEM: {system_message}
246
+ USER: {prompt}
247
+ ASSISTANT:
248
+ '''
249
+
250
+ client = InferenceClient(endpoint_url)
251
+ response = client.text_generation(
252
+ prompt_template,
253
+ max_new_tokens=128,
254
+ do_sample=True,
255
+ temperature=0.7,
256
+ top_p=0.95,
257
+ top_k=40,
258
+ repetition_penalty=1.1
259
+ )
260
+
261
+ print(f"Model output: {response}")
262
+ ```
263
+ <!-- README_GPTQ.md-use-from-tgi end -->
264
+ <!-- README_GPTQ.md-use-from-python start -->
265
+ ## Python code example: inference from this GPTQ model
266
+
267
+ ### Install the necessary packages
268
+
269
+ Requires: Transformers 4.33.0 or later, Optimum 1.12.0 or later, and AutoGPTQ 0.4.2 or later.
270
+
271
+ ```shell
272
+ pip3 install --upgrade transformers optimum
273
+ # If using PyTorch 2.1 + CUDA 12.x:
274
+ pip3 install --upgrade auto-gptq
275
+ # or, if using PyTorch 2.1 + CUDA 11.x:
276
+ pip3 install --upgrade auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
277
+ ```
278
+
279
+ If you are using PyTorch 2.0, you will need to install AutoGPTQ from source. Likewise if you have problems with the pre-built wheels, you should try building from source:
280
+
281
+ ```shell
282
+ pip3 uninstall -y auto-gptq
283
+ git clone https://github.com/PanQiWei/AutoGPTQ
284
+ cd AutoGPTQ
285
+ git checkout v0.5.1
286
+ pip3 install .
287
+ ```
288
+
289
+ ### Example Python code
290
+
291
+ ```python
292
+ from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
293
+
294
+ model_name_or_path = "TheBloke/Yi-34B-200K-DARE-megamerge-v8-GPTQ"
295
+ # To use a different branch, change revision
296
+ # For example: revision="gptq-4bit-128g-actorder_True"
297
+ model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
298
+ device_map="auto",
299
+ trust_remote_code=False,
300
+ revision="main")
301
+
302
+ tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)
303
+
304
+ prompt = "Write a story about llamas"
305
+ system_message = "You are a story writing assistant"
306
+ prompt_template=f'''SYSTEM: {system_message}
307
+ USER: {prompt}
308
+ ASSISTANT:
309
+ '''
310
+
311
+ print("\n\n*** Generate:")
312
+
313
+ input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
314
+ output = model.generate(inputs=input_ids, temperature=0.7, do_sample=True, top_p=0.95, top_k=40, max_new_tokens=512)
315
+ print(tokenizer.decode(output[0]))
316
+
317
+ # Inference can also be done using transformers' pipeline
318
+
319
+ print("*** Pipeline:")
320
+ pipe = pipeline(
321
+ "text-generation",
322
+ model=model,
323
+ tokenizer=tokenizer,
324
+ max_new_tokens=512,
325
+ do_sample=True,
326
+ temperature=0.7,
327
+ top_p=0.95,
328
+ top_k=40,
329
+ repetition_penalty=1.1
330
+ )
331
+
332
+ print(pipe(prompt_template)[0]['generated_text'])
333
+ ```
334
+ <!-- README_GPTQ.md-use-from-python end -->
335
+
336
+ <!-- README_GPTQ.md-compatibility start -->
337
+ ## Compatibility
338
+
339
+ The files provided are tested to work with Transformers. For non-Mistral models, AutoGPTQ can also be used directly.
340
+
341
+ [ExLlama](https://github.com/turboderp/exllama) is compatible with Llama architecture models (including Mistral, Yi, DeepSeek, SOLAR, etc) in 4-bit. Please see the Provided Files table above for per-file compatibility.
342
+
343
+ For a list of clients/servers, please see "Known compatible clients / servers", above.
344
+ <!-- README_GPTQ.md-compatibility end -->
345
+
346
+ <!-- footer start -->
347
+ <!-- 200823 -->
348
+ ## Discord
349
+
350
+ For further support, and discussions on these models and AI in general, join us at:
351
+
352
+ [TheBloke AI's Discord server](https://discord.gg/theblokeai)
353
+
354
+ ## Thanks, and how to contribute
355
+
356
+ Thanks to the [chirper.ai](https://chirper.ai) team!
357
+
358
+ Thanks to Clay from [gpus.llm-utils.org](llm-utils)!
359
+
360
+ I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
361
+
362
+ If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
363
+
364
+ Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
365
+
366
+ * Patreon: https://patreon.com/TheBlokeAI
367
+ * Ko-Fi: https://ko-fi.com/TheBlokeAI
368
+
369
+ **Special thanks to**: Aemon Algiz.
370
+
371
+ **Patreon special mentions**: Michael Levine, 阿明, Trailburnt, Nikolai Manek, John Detwiler, Randy H, Will Dee, Sebastain Graf, NimbleBox.ai, Eugene Pentland, Emad Mostaque, Ai Maven, Jim Angel, Jeff Scroggin, Michael Davis, Manuel Alberto Morcote, Stephen Murray, Robert, Justin Joy, Luke @flexchar, Brandon Frisco, Elijah Stavena, S_X, Dan Guido, Undi ., Komninos Chatzipapas, Shadi, theTransient, Lone Striker, Raven Klaugh, jjj, Cap'n Zoog, Michel-Marie MAUDET (LINAGORA), Matthew Berman, David, Fen Risland, Omer Bin Jawed, Luke Pendergrass, Kalila, OG, Erik Bjäreholt, Rooh Singh, Joseph William Delisle, Dan Lewis, TL, John Villwock, AzureBlack, Brad, Pedro Madruga, Caitlyn Gatomon, K, jinyuan sun, Mano Prime, Alex, Jeffrey Morgan, Alicia Loh, Illia Dulskyi, Chadd, transmissions 11, fincy, Rainer Wilmers, ReadyPlayerEmma, knownsqashed, Mandus, biorpg, Deo Leter, Brandon Phillips, SuperWojo, Sean Connelly, Iucharbius, Jack West, Harry Royden McLaughlin, Nicholas, terasurfer, Vitor Caleffi, Duane Dunston, Johann-Peter Hartmann, David Ziegler, Olakabola, Ken Nordquist, Trenton Dambrowitz, Tom X Nguyen, Vadim, Ajan Kanaga, Leonard Tan, Clay Pascal, Alexandros Triantafyllidis, JM33133, Xule, vamX, ya boyyy, subjectnull, Talal Aujan, Alps Aficionado, wassieverse, Ari Malik, James Bentley, Woland, Spencer Kim, Michael Dempsey, Fred von Graf, Elle, zynix, William Richards, Stanislav Ovsiannikov, Edmond Seymore, Jonathan Leane, Martin Kemka, usrbinkat, Enrico Ros
372
+
373
+
374
+ Thank you to all my generous patrons and donaters!
375
+
376
+ And thank you again to a16z for their generous grant.
377
+
378
+ <!-- footer end -->
379
+
380
+ # Original model card: brucethemoose's Yi 34B 200K DARE MegaMerge V8
381
+
382
+ # Yi 34B 200K DARE Merge v8
383
+
384
+ A merge of many Yi 34B 200K models using the new DARE Ties method via mergekit. The goal is to create a merge model that excels at 32K+ context performance, without any additional finetuning.
385
+
386
+ ## Prompt template: Orca-Vicuna
387
+ ```
388
+ SYSTEM: {system_message}
389
+ USER: {prompt}
390
+ ASSISTANT:
391
+ ```
392
+ It might recognize ChatML, and possibly Alpaca-like formats. Raw prompting as described here is also effective: https://old.reddit.com/r/LocalLLaMA/comments/18zqy4s/the_secret_to_writing_quality_stories_with_llms/
393
+
394
+
395
+
396
+ ## Running
397
+ Being a Yi model, run a lower temperature with 0.05 or higher MinP, a little repetition penalty, maybe mirostat with a low tau, and no other samplers. Yi tends to run "hot" by default, and it really needs a low temperature + MinP to cull Yi's huge vocabulary. See the explanation here: https://github.com/ggerganov/llama.cpp/pull/3841
398
+
399
+ 24GB GPUs can efficiently run Yi-34B-200K models at **40K-90K context** with exllamav2, and performant UIs like [exui](https://github.com/turboderp/exui). I go into more detail in this [post](https://old.reddit.com/r/LocalLLaMA/comments/1896igc/how_i_run_34b_models_at_75k_context_on_24gb_fast/). 16GB GPUs can still run the high context with aggressive quantization.
400
+
401
+ I recommend exl2 quantizations profiled on data similar to the desired task. It is especially sensitive to the quantization data at low bpw. I've upload my own fiction-oriented quantizations here: https://huggingface.co/collections/brucethemoose/most-recent-merge-65742644ca03b6c514afa204
402
+
403
+ To load/train this in full-context backends like transformers, you *must* change `max_position_embeddings` in config.json to a lower value than 200,000, otherwise you will OOM! I do not recommend running high context without context-efficient backends like exllamav2, litellm or unsloth.
404
+
405
+
406
+ ## Testing Notes
407
+
408
+ See: https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-merge-v5#testing-notes
409
+
410
+ An intermediate merge model was created to try and extend the context of several 4k models before adding them to the main merge, as seen in the "megamerge" recipe below. I can upload this upon request
411
+
412
+ In addition, the weight gradients are biased towards Vicuna-format models in the first few layers to try and "emphasize" the Orca-Vicuna prompt template. How sucessful this is remains to be seen.
413
+
414
+
415
+
416
+ ## Merge Details
417
+ ### Merge Method
418
+
419
+ This model was merged using the [DARE](https://arxiv.org/abs/2311.03099) [TIES](https://arxiv.org/abs/2306.01708) merge method using /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama as a base.
420
+
421
+ ### Models Merged
422
+
423
+ The following models were included in the merge:
424
+ * https://huggingface.co/kyujinpy/PlatYi-34B-200k-Q-FastChat
425
+ * https://huggingface.co/jondurbin/bagel-34b-v0.2
426
+ * https://huggingface.co/migtissera/Tess-M-Creative-v1.0
427
+ * https://huggingface.co/brucethemoose/SUS-Bagel-200K-DARE-Test
428
+ * https://huggingface.co/Mihaiii/Pallas-0.5
429
+ * https://huggingface.co/bhenrym14/airoboros-3_1-yi-34b-200k
430
+ * https://huggingface.co/adamo1139/Yi-34B-200K-AEZAKMI-v2
431
+ * https://huggingface.co/migtissera/Tess-34B-v1.4
432
+ * https://huggingface.co/SUSTech/SUS-Chat-34B
433
+ * https://huggingface.co/jondurbin/bagel-dpo-34b-v0.2
434
+ * https://huggingface.co/bhenrym14/platypus-yi-34b
435
+ * https://huggingface.co/Weyaxi/Nous-Hermes-2-SUS-Chat-34B-Slerp
436
+ * https://huggingface.co/TriadParty/deepsex-34b
437
+ * https://huggingface.co/TriadParty/deepmoney-34b-200k-base
438
+ * https://huggingface.co/chargoddard/Yi-34B-200K-Llama
439
+ * https://huggingface.co/chargoddard/Yi-34B-Llama
440
+
441
+ ### Configuration
442
+
443
+ The following YAML configuration was used to produce this model:
444
+
445
+ ```yaml
446
+ models:
447
+ - model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
448
+ # No parameters necessary for base model
449
+ - model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
450
+ #200K base to extend the context of 4K models, max density as we *want* it to 'interfere'
451
+ parameters:
452
+ weight: 0.33
453
+ density: 1
454
+ - model: /home/alpha/Models/Raw/Weyaxi_Nous-Hermes-2-SUS-Chat-34B-Slerp
455
+ parameters:
456
+ weight: 0.15
457
+ density: 0.36
458
+ - model: /home/alpha/Models/Raw/jondurbin_bagel-dpo-34b-v0.2
459
+ #Mix dpo with sft to tone down dpo
460
+ parameters:
461
+ weight: 0.06
462
+ density: 0.36
463
+ - model: /home/alpha/Models/Raw/jondurbin_bagel-34b-v0.2
464
+ parameters:
465
+ weight: 0.06
466
+ density: 0.41
467
+ - model: /home/alpha/Models/Raw/bhenrym14_platypus-yi-34b
468
+ #Vicuna format
469
+ parameters:
470
+ weight: 0.19
471
+ density: 0.41
472
+ # - model: /home/alpha/Models/Raw/01-ai_Yi-34B-Chat #+/home/alpha/Models/Raw/Doctor-Shotgun_limarpv3-yi-llama-34b-lora
473
+ # #Can't get lora OR base model to work without erroring out?
474
+ # parameters:
475
+ # weight: 0.04
476
+ # density: 0.36
477
+ - model: /home/alpha/Models/Raw/TriadParty_deepsex-34b
478
+ #Base model with no prompt
479
+ parameters:
480
+ weight: 0.21
481
+ density: 0.39
482
+ merge_method: dare_ties
483
+ tokenizer_source: union
484
+ base_model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
485
+ parameters:
486
+ int8_mask: true
487
+ dtype: bfloat16
488
+ name: 4kmerge-v2
489
+ ---
490
+ models:
491
+ - model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
492
+ # No parameters necessary for base model
493
+ - model: /home/alpha/Storage/Models/Raw/migtissera_Tess-34B-v1.4
494
+ #Emphasize the beginning of Vicuna format models
495
+ parameters:
496
+ weight: [0.22, 0.113, 0.113, 0.113, 0.113, 0.113]
497
+ density: 0.61
498
+ - model: /home/alpha/Models/Raw/Mihaiii_Pallas-0.5
499
+ # Vicuna format
500
+ parameters:
501
+ weight: [0.22, 0.113, 0.113, 0.113, 0.113, 0.113]
502
+ density: 0.61
503
+ - model: /home/alpha//Storage/Models/Raw/bhenrym14_airoboros-3_1-yi-34b-200k
504
+ parameters:
505
+ weight: [0.02, 0.081, 0.081, 0.081, 0.081, 0.081]
506
+ density: 0.59
507
+ - model: /home/alpha/Storage/Models/Raw/jondurbin_bagel-34b-v0.2
508
+ #Only the SFT in the main merge since the DPO version seems to have no long context ability at all, and some overfitting(?) issues
509
+ parameters:
510
+ weight: [0.02, 0.093, 0.093, 0.093, 0.093, 0.093]
511
+ density: 0.4
512
+ - model: /home/alpha/Storage/Models/Raw/kyujinpy_PlatYi-34B-200k-Q-FastChat
513
+ parameters:
514
+ weight: [0.02, 0.081, 0.081, 0.081, 0.081, 0.081]
515
+ density: 0.59
516
+ #- model: /home/alpha/Storage/Models/Raw/ehartford_dolphin-2.2-yi-34b-200k
517
+ # Dolphin 200K seems to be funky according to multiple leaderboards and perplexity tests?
518
+ # parameters:
519
+ # weight: 0.15
520
+ # density: 0.6
521
+ - model: /home/alpha/Models/Raw/adamo1139_Yi-34B-200K-AEZAKMI-v2
522
+ parameters:
523
+ weight: [0.02, 0.096, 0.096, 0.096, 0.096, 0.096]
524
+ density: 0.59
525
+ - model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B
526
+ parameters:
527
+ weight: [0.21, 0.115, 0.115, 0.115, 0.115, 0.115]
528
+ density: 0.59
529
+ - model: 4kmerge-v2
530
+ #Previous merge
531
+ parameters:
532
+ weight: [0.02, 0.115, 0.115, 0.115, 0.115, 0.115]
533
+ density: 0.4
534
+ - model: /home/alpha/Models/Raw/migtissera_Tess-M-Creative-v1.0
535
+ # Vicuna format
536
+ parameters:
537
+ weight: [0.21, 0.09, 0.09, 0.09, 0.09, 0.09]
538
+ density: 0.61
539
+ - model: /home/alpha/Models/Raw/TriadParty_deepmoney-34b-200k-base
540
+ # No prompt format, native long context full finetune
541
+ parameters:
542
+ weight: [0.04, 0.103, 0.103, 0.103, 0.103, 0.103]
543
+ density: 0.61
544
+ merge_method: dare_ties
545
+ tokenizer_source: union
546
+ base_model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
547
+ parameters:
548
+ int8_mask: true
549
+ dtype: bfloat16
550
+ ```