TheBloke commited on
Commit
cab5bd7
1 Parent(s): 72664a4

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +354 -0
README.md ADDED
@@ -0,0 +1,354 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - Composer
5
+ - MosaicML
6
+ - llm-foundry
7
+ - StreamingDatasets
8
+ datasets:
9
+ - allenai/c4
10
+ - mc4
11
+ - togethercomputer/RedPajama-Data-1T
12
+ - bigcode/the-stack-dedup
13
+ - allenai/s2orc
14
+ inference: false
15
+ ---
16
+
17
+ <!-- header start -->
18
+ <div style="width: 100%;">
19
+ <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
20
+ </div>
21
+ <div style="display: flex; justify-content: space-between; width: 100%;">
22
+ <div style="display: flex; flex-direction: column; align-items: flex-start;">
23
+ <p><a href="https://discord.gg/Jq4vkcDakD">Chat & support: my new Discord server</a></p>
24
+ </div>
25
+ <div style="display: flex; flex-direction: column; align-items: flex-end;">
26
+ <p><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p>
27
+ </div>
28
+ </div>
29
+ <!-- header end -->
30
+
31
+ # MosaicML's MPT-30B GGML
32
+
33
+ These files are GGML format model files for [MosaicML's MPT-30B](https://huggingface.co/mosaicml/mpt-30b).
34
+
35
+ Please note that these GGMLs are **not compatible with llama.cpp, or currently with text-generation-webui**. Please see below for a list of tools known to work with these model files.
36
+
37
+ [KoboldCpp](https://github.com/LostRuins/koboldcpp) just added GPU accelerated (OpenCL) support for MPT models, so that is the client I recommend using for these models.
38
+
39
+ **Note**: Please use version 1.32.1 or later of KoboldCpp, as there was a bug in 1.32 that affected loading of MPT models.
40
+
41
+ ## Repositories available
42
+
43
+ * [2, 3, 4, 5, 6 and 8-bit GGML models for CPU+GPU inference](https://huggingface.co/TheBloke/mpt-30B-GGML)
44
+ * [Unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/mosaicml/mpt-30b)
45
+
46
+ ## Prompt template
47
+
48
+ This is a non-fine-tuned base model. So it is designed for text completion, not following instructions, such as in the following example:
49
+ ```
50
+ The meaning of life is
51
+ ```
52
+
53
+ ## A note regarding context length: 8K
54
+
55
+ The base model has an 8K context length. It is not yet confirmed if the 8K context of this model works with the quantised files.
56
+
57
+ If it does, [KoboldCpp](https://github.com/LostRuins/koboldcpp) supports 8K context if you manually set it to 8K by adjusting the text box above the slider:
58
+ ![.](https://i.imgur.com/tEbpeJqm.png)
59
+
60
+ It is currently unknown as to increased context is compatible with other MPT GGML clients.
61
+
62
+ If you have feedback on this, please let me know.
63
+
64
+ <!-- compatibility_ggml start -->
65
+ ## Compatibilty
66
+
67
+ These files are **not** compatible with text-generation-webui, llama.cpp, or llama-cpp-python.
68
+
69
+ Currently they can be used with:
70
+ * KoboldCpp, a powerful inference engine based on llama.cpp, with good UI and GPU accelerated support for MPT models: [KoboldCpp](https://github.com/LostRuins/koboldcpp)
71
+ * The ctransformers Python library, which includes LangChain support: [ctransformers](https://github.com/marella/ctransformers)
72
+ * The LoLLMS Web UI which uses ctransformers: [LoLLMS Web UI](https://github.com/ParisNeo/lollms-webui)
73
+ * [rustformers' llm](https://github.com/rustformers/llm)
74
+ * The example `mpt` binary provided with [ggml](https://github.com/ggerganov/ggml)
75
+
76
+ As other options become available I will endeavour to update them here (do let me know in the Community tab if I've missed something!)
77
+
78
+ ## Tutorial for using LoLLMS Web UI
79
+
80
+ * [Text tutorial, written by **Lucas3DCG**](https://huggingface.co/TheBloke/MPT-7B-Storywriter-GGML/discussions/2#6475d914e9b57ce0caa68888)
81
+ * [Video tutorial, by LoLLMS Web UI's author **ParisNeo**](https://www.youtube.com/watch?v=ds_U0TDzbzI)
82
+
83
+ <!-- compatibility_ggml end -->
84
+
85
+ ## Provided files
86
+ | Name | Quant method | Bits | Size | Max RAM required | Use case |
87
+ | ---- | ---- | ---- | ---- | ---- | ----- |
88
+ | mpt-30b.ggmlv0.q4_0.bin | q4_0 | 4 | 16.85 GB | 19.35 GB | Original llama.cpp quant method, 4-bit. |
89
+ | mpt-30b.ggmlv0.q4_1.bin | q4_1 | 4 | 18.73 GB | 21.23 GB | Original llama.cpp quant method, 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
90
+ | mpt-30b.ggmlv0.q5_0.bin | q5_0 | 5 | 20.60 GB | 23.10 GB | Original llama.cpp quant method, 5-bit. Higher accuracy, higher resource usage and slower inference. |
91
+ | mpt-30b.ggmlv0.q5_1.bin | q5_1 | 5 | 22.47 GB | 24.97 GB | Original llama.cpp quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
92
+ | mpt-30b.ggmlv0.q8_0.bin | q8_0 | 8 | 31.83 GB | 34.33 GB | Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
93
+
94
+ **Note**: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
95
+
96
+ <!-- footer start -->
97
+ ## Discord
98
+
99
+ For further support, and discussions on these models and AI in general, join us at:
100
+
101
+ [TheBloke AI's Discord server](https://discord.gg/Jq4vkcDakD)
102
+
103
+ ## Thanks, and how to contribute.
104
+
105
+ Thanks to the [chirper.ai](https://chirper.ai) team!
106
+
107
+ 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.
108
+
109
+ 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.
110
+
111
+ Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
112
+
113
+ * Patreon: https://patreon.com/TheBlokeAI
114
+ * Ko-Fi: https://ko-fi.com/TheBlokeAI
115
+
116
+ **Special thanks to**: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.
117
+
118
+ **Patreon special mentions**: Mano Prime, Fen Risland, Derek Yates, Preetika Verma, webtim, Sean Connelly, Alps Aficionado, Karl Bernard, Junyu Yang, Nathan LeClaire, Chris McCloskey, Lone Striker, Asp the Wyvern, Eugene Pentland, Imad Khwaja, trip7s trip, WelcomeToTheClub, John Detwiler, Artur Olbinski, Khalefa Al-Ahmad, Trenton Dambrowitz, Talal Aujan, Kevin Schuppel, Luke Pendergrass, Pyrater, Joseph William Delisle, terasurfer , vamX, Gabriel Puliatti, David Flickinger, Jonathan Leane, Iucharbius , Luke, Deep Realms, Cory Kujawski, ya boyyy, Illia Dulskyi, senxiiz, Johann-Peter Hartmann, John Villwock, K, Ghost , Spiking Neurons AB, Nikolai Manek, Rainer Wilmers, Pierre Kircher, biorpg, Space Cruiser, Ai Maven, subjectnull, Willem Michiel, Ajan Kanaga, Kalila, chris gileta, Oscar Rangel.
119
+
120
+ Thank you to all my generous patrons and donaters!
121
+
122
+ <!-- footer end -->
123
+
124
+ # Original model card: MosaicML's MPT-30B
125
+
126
+ # MPT-30B
127
+
128
+ MPT-30B is a decoder-style transformer pretrained from scratch on 1T tokens of English text and code.
129
+ This model was trained by [MosaicML](https://www.mosaicml.com).
130
+
131
+ MPT-30B is part of the family of Mosaic Pretrained Transformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference.
132
+
133
+ MPT-30B comes with special features that differentiate it from other LLMs, including an 8k token context window (which can be further extended via finetuning; see [MPT-7B-StoryWriter](https://huggingface.co/mosaicml/mpt-7b-storywriter)), support for context-length extrapolation via [ALiBi](https://arxiv.org/abs/2108.12409), and efficient inference + training via FlashAttention. It also has strong coding abilities thanks to its pretraining mix. MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's [FasterTransformer](https://github.com/NVIDIA/FasterTransformer).
134
+ The size of MPT-30B was also specifically chosen to make it easy to deploy on a single GPU—either 1xA100-80GB in 16-bit precision or 1xA100-40GB in 8-bit precision.
135
+
136
+ This model uses the MosaicML LLM codebase, which can be found in the [llm-foundry repository](https://github.com/mosaicml/llm-foundry). It was trained by MosaicML’s NLP team on the [MosaicML platform](https://www.mosaicml.com/training) for LLM pretraining, finetuning, and inference.
137
+
138
+
139
+ ### How is this model different?
140
+
141
+ MPT-30B is:
142
+ * **Licensed for the possibility of commercial use** (unlike [LLaMA](https://arxiv.org/abs/2302.13971)).
143
+ * **Trained on a large amount of data** (1T tokens like [LLaMA](https://arxiv.org/abs/2302.13971) vs. 300B for [Pythia](https://github.com/EleutherAI/pythia), 300B for [OpenLLaMA](https://github.com/openlm-research/open_llama), and 800B for [StableLM](https://github.com/Stability-AI/StableLM)).
144
+ * **Prepared to handle extremely long inputs** thanks to [ALiBi](https://arxiv.org/abs/2108.12409).
145
+ * **Capable of fast training and inference** (via [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf) and [FasterTransformer](https://github.com/NVIDIA/FasterTransformer))
146
+ * **Equipped with highly efficient open-source training code** via the [llm-foundry repository](https://github.com/mosaicml/llm-foundry)
147
+
148
+ ### Models finetuned off MPT-30B:
149
+
150
+ The following models are finetuned on MPT-30B:
151
+
152
+ * [MPT-30B-Instruct](https://huggingface.co/mosaicml/mpt-30b-instruct): a model for short-form instruction following.
153
+ Built by finetuning MPT-30B on several carefully curated datasets.
154
+ * License: _CC-By-NC-SA-3.0_
155
+
156
+ * [MPT-30B-Chat](https://huggingface.co/mosaicml/mpt-30b-chat): a chatbot-like model for dialogue generation.
157
+ Built by finetuning MPT-30B on [ShareGPT-Vicuna](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered), [Camel-AI](https://huggingface.co/camel-ai),
158
+ [GPTeacher](https://github.com/teknium1/GPTeacher), [Guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco), [Baize](https://github.com/project-baize/baize-chatbot) and some generated datasets.
159
+ * License: _CC-By-NC-SA-4.0_
160
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-30b-chat)
161
+
162
+ ## Model Date
163
+
164
+ June 22, 2023
165
+
166
+ ## Model License
167
+
168
+ Apache-2.0
169
+
170
+ ## Documentation
171
+
172
+ * [Blog post: MPT-30B: Raising the bar for open-source foundation models](https://www.mosaicml.com/blog/mpt-30b)
173
+ * [Codebase (mosaicml/llm-foundry repo)](https://github.com/mosaicml/llm-foundry/)
174
+ * Questions: Feel free to contact us via the [MosaicML Community Slack](https://mosaicml.me/slack)!
175
+
176
+
177
+ ## How to Use
178
+
179
+ This model is best used with the MosaicML [llm-foundry repository](https://github.com/mosaicml/llm-foundry) for training and finetuning.
180
+
181
+ ```python
182
+ import transformers
183
+ model = transformers.AutoModelForCausalLM.from_pretrained(
184
+ 'mosaicml/mpt-30b',
185
+ trust_remote_code=True
186
+ )
187
+ ```
188
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
189
+ This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package.
190
+ `MPT` includes options for many training efficiency features such as [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi](https://arxiv.org/abs/2108.12409), [QK LayerNorm](https://arxiv.org/abs/2010.04245), and more.
191
+
192
+ To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model on GPU (`cuda:0`) with `attn_impl='triton'` and with `bfloat16` precision:
193
+ ```python
194
+ import torch
195
+ import transformers
196
+
197
+ name = 'mosaicml/mpt-30b'
198
+
199
+ config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
200
+ config.attn_config['attn_impl'] = 'triton' # change this to use triton-based FlashAttention
201
+ config.init_device = 'cuda:0' # For fast initialization directly on GPU!
202
+
203
+ model = transformers.AutoModelForCausalLM.from_pretrained(
204
+ name,
205
+ config=config,
206
+ torch_dtype=torch.bfloat16, # Load model weights in bfloat16
207
+ trust_remote_code=True
208
+ )
209
+ ```
210
+
211
+ The model was trained initially with a sequence length of 4096 with an additional pretraining stage for sequence length adapation up to 8192. However, ALiBi enables users to increase the maximum sequence length even further during finetuning and/or inference. For example:
212
+
213
+ ```python
214
+ import transformers
215
+
216
+ name = 'mosaicml/mpt-30b'
217
+
218
+ config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
219
+ config.max_seq_len = 16384 # (input + output) tokens can now be up to 16384
220
+
221
+ model = transformers.AutoModelForCausalLM.from_pretrained(
222
+ name,
223
+ config=config,
224
+ trust_remote_code=True
225
+ )
226
+ ```
227
+
228
+ This model was trained with the MPT-30B tokenizer which is identical to the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
229
+
230
+ ```python
231
+ from transformers import AutoTokenizer
232
+ tokenizer = AutoTokenizer.from_pretrained('mosaicml/mpt-30b')
233
+ ```
234
+
235
+ The model can then be used, for example, within a text-generation pipeline.
236
+ Note: when running Torch modules in lower precision, it is best practice to use the [torch.autocast context manager](https://pytorch.org/docs/stable/amp.html).
237
+
238
+ ```python
239
+ from transformers import pipeline
240
+
241
+ with torch.autocast('cuda', dtype=torch.bfloat16):
242
+ inputs = tokenizer('Here is a recipe for vegan banana bread:\n', return_tensors="pt").to('cuda')
243
+ outputs = model.generate(**inputs, max_new_tokens=100)
244
+ print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
245
+
246
+ # or using the HF pipeline
247
+ pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0')
248
+ with torch.autocast('cuda', dtype=torch.bfloat16):
249
+ print(
250
+ pipe('Here is a recipe for vegan banana bread:\n',
251
+ max_new_tokens=100,
252
+ do_sample=True,
253
+ use_cache=True))
254
+ ```
255
+
256
+ ## Model Description
257
+
258
+ The architecture is a modification of a standard decoder-only transformer.
259
+
260
+ The model has been modified from a standard transformer in the following ways:
261
+ * It uses [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf)
262
+ * It uses [ALiBi (Attention with Linear Biases)](https://arxiv.org/abs/2108.12409) and does not use positional embeddings
263
+ * It does not use biases
264
+
265
+
266
+ | Hyperparameter | Value |
267
+ |----------------|-------|
268
+ |n_parameters | 29.95B |
269
+ |n_layers | 48 |
270
+ | n_heads | 64 |
271
+ | d_model | 7168 |
272
+ | vocab size | 50432 |
273
+ | sequence length | 8192 |
274
+
275
+
276
+
277
+ ## Training Data
278
+
279
+ ### Streaming Datasets
280
+
281
+ Data was formatted using the MosaicML [StreamingDataset](https://github.com/mosaicml/streaming) library to host our data in object storage and efficiently stream it to our compute cluster during training.
282
+ StreamingDataset obviates the need to download the whole dataset before starting training, and allows instant resumption of training from any point in the dataset.
283
+
284
+
285
+ ### Data Mix
286
+
287
+ The model was trained for 1T tokens on the following data mix:
288
+
289
+ | Data Source | Number of Tokens in Source | Proportion | Effective Number of Tokens | Epochs |
290
+ |-------------|----------------------------|------------|----------------------------|--------|
291
+ | mC4 3.1.0 - English (200+ words) | 2417.99 B | 33.50% | 335 B | 0.14 |
292
+ | c4 - English - SemDedup 80% | 100.42 B | 29.90% | 299 B | 2.98 |
293
+ | RedPajama - CommonCrawl | 878.45 B | 8.50% | 85 B | 0.097 |
294
+ | The Stack - Selected Languages | 463.78 B | 10.00% | 100 B | 0.22 |
295
+ | RedPajama - Wikipedia | 4.87 B | 4.00% | 40 B | 8.21 |
296
+ | The Stack - Markdown | 107.07 B | 4.50% | 45 B | 0.42 |
297
+ | Semantic Scholar ORC | 48.95 B | 3.30% | 33 B | 0.67 |
298
+ | RedPajama - Books | 26.02 B | 3.00% | 30 B | 1.15 |
299
+ | RedPajama - arXiv | 28.10 B | 1.90% | 19 B | 0.68 |
300
+ | RedPajama - StackExchange | 20.54 B | 1.40% | 14 B |0.68 |
301
+
302
+ Samples for each batch were selected from one of the datasets with the probability specified above. The examples were shuffled within each dataset, and each example was constructed from as many sequences from that dataset as were necessary to fill the sequence length. To build 8k support into MPT-30B efficiently, we first pre-trained on 1T tokens using sequences that were 2k tokens long, and then trained for an additional 50B tokens using sequences that were 8k tokens long.
303
+
304
+ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer. This BPE tokenizer has a number of desirable characteristics,
305
+ most of which are relevant for tokenizing code:
306
+ (1) It was trained on a diverse mix of data that includes code (The Pile)
307
+ (2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces
308
+ (3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters.
309
+
310
+ The model vocabulary size of 50432 was set to be a multiple of 128 (as in [MEGATRON-LM](https://arxiv.org/abs/1909.08053)).
311
+
312
+ ### Training Configuration
313
+
314
+ The model was trained in three stages using the [MosaicML Platform](https://www.mosaicml.com/platform):
315
+ (i) First it was trained on 440 A100-40GBs with a batch size of 1760.
316
+ (ii) Then, on 216 A100-40GBs with a batch size of 1728.
317
+ (iii) Training was completed on 256 H100-80GBs with a batch size of 512 with 8k context length and 50B tokens.
318
+ The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
319
+
320
+ ## Limitations and Biases
321
+
322
+ _The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_
323
+
324
+ MPT-30B (Base) is **not** intended for deployment without finetuning.
325
+ It should not be used for human-facing interactions without further guardrails and user consent.
326
+
327
+ MPT-30B can produce factually incorrect output, and should not be relied on to produce factually accurate information.
328
+ MPT-30B was trained on various public datasets.
329
+ While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
330
+
331
+
332
+ ## MosaicML Platform
333
+
334
+ If you're interested in [training](https://www.mosaicml.com/training) and [deploying](https://www.mosaicml.com/inference) your own MPT or LLMs on the MosaicML Platform, [sign up here](https://forms.mosaicml.com/demo?utm_source=huggingface&utm_medium=referral&utm_campaign=mpt-30b).
335
+
336
+ ## Disclaimer
337
+
338
+ The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.
339
+
340
+ ## Citation
341
+
342
+ Please cite this model using the following format:
343
+
344
+ ```
345
+ @online{MosaicML2023Introducing,
346
+ author = {MosaicML NLP Team},
347
+ title = {Introducing MPT-30B: Raising the bar
348
+ for open-source foundation models},
349
+ year = {2023},
350
+ url = {www.mosaicml.com/blog/mpt-30b},
351
+ note = {Accessed: 2023-06-22},
352
+ urldate = {2023-06-22}
353
+ }
354
+ ```