TheBloke commited on
Commit
0d92380
1 Parent(s): 7696422

Update for Transformers GPTQ support

Browse files
README.md CHANGED
@@ -7,17 +7,20 @@ task_categories:
7
  ---
8
 
9
  <!-- header start -->
10
- <div style="width: 100%;">
11
- <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
 
12
  </div>
13
  <div style="display: flex; justify-content: space-between; width: 100%;">
14
  <div style="display: flex; flex-direction: column; align-items: flex-start;">
15
- <p><a href="https://discord.gg/theblokeai">Chat & support: my new Discord server</a></p>
16
  </div>
17
  <div style="display: flex; flex-direction: column; align-items: flex-end;">
18
- <p><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p>
19
  </div>
20
  </div>
 
 
21
  <!-- header end -->
22
 
23
  # Dolphin Llama 13B - GPTQ
@@ -52,13 +55,13 @@ Each separate quant is in a different branch. See below for instructions on fet
52
 
53
  | Branch | Bits | Group Size | Act Order (desc_act) | File Size | ExLlama Compatible? | Made With | Description |
54
  | ------ | ---- | ---------- | -------------------- | --------- | ------------------- | --------- | ----------- |
55
- | main | 4 | 128 | False | 7.26 GB | True | AutoGPTQ | Most compatible option. Good inference speed in AutoGPTQ and GPTQ-for-LLaMa. Lower inference quality than other options. |
56
- | gptq-4bit-32g-actorder_True | 4 | 32 | True | 8.00 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 32g. Gives highest possible inference quality, with maximum VRAM usage. Poor AutoGPTQ CUDA speed. |
57
- | gptq-4bit-64g-actorder_True | 4 | 64 | True | 7.51 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 64g. Uses less VRAM than 32g, but with slightly lower accuracy. Poor AutoGPTQ CUDA speed. |
58
- | gptq-4bit-128g-actorder_True | 4 | 128 | True | 7.26 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 128g. Uses even less VRAM than 64g, but with slightly lower accuracy. Poor AutoGPTQ CUDA speed. |
59
- | gptq-8bit--1g-actorder_True | 8 | None | True | 13.36 GB | False | AutoGPTQ | 8-bit, with Act Order. No group size, to lower VRAM requirements and to improve AutoGPTQ speed. |
60
- | gptq-8bit-128g-actorder_False | 8 | 128 | False | 13.65 GB | False | AutoGPTQ | 8-bit, with group size 128g for higher inference quality and without Act Order to improve AutoGPTQ speed. |
61
- | gptq-8bit-128g-actorder_True | 8 | 128 | True | 13.65 GB | False | AutoGPTQ | 8-bit, with group size 128g for higher inference quality and with Act Order for even higher accuracy. Poor AutoGPTQ CUDA speed. |
62
  | gptq-8bit-64g-actorder_True | 8 | 64 | True | 13.95 GB | False | AutoGPTQ | 8-bit, with group size 64g and Act Order for maximum inference quality. Poor AutoGPTQ CUDA speed. |
63
 
64
  ## How to download from branches
@@ -102,7 +105,7 @@ from transformers import AutoTokenizer, pipeline, logging
102
  from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
103
 
104
  model_name_or_path = "TheBloke/Dolphin-Llama-13B-GPTQ"
105
- model_basename = "gptq_model-4bit-128g"
106
 
107
  use_triton = False
108
 
@@ -166,6 +169,7 @@ The files provided will work with AutoGPTQ (CUDA and Triton modes), GPTQ-for-LLa
166
  ExLlama works with Llama models in 4-bit. Please see the Provided Files table above for per-file compatibility.
167
 
168
  <!-- footer start -->
 
169
  ## Discord
170
 
171
  For further support, and discussions on these models and AI in general, join us at:
@@ -185,13 +189,15 @@ Donaters will get priority support on any and all AI/LLM/model questions and req
185
  * Patreon: https://patreon.com/TheBlokeAI
186
  * Ko-Fi: https://ko-fi.com/TheBlokeAI
187
 
188
- **Special thanks to**: Luke from CarbonQuill, Aemon Algiz.
189
 
190
- **Patreon special mentions**: Slarti, Chadd, John Detwiler, Pieter, zynix, K, Mano Prime, ReadyPlayerEmma, Ai Maven, Leonard Tan, Edmond Seymore, Joseph William Delisle, Luke @flexchar, Fred von Graf, Viktor Bowallius, Rishabh Srivastava, Nikolai Manek, Matthew Berman, Johann-Peter Hartmann, ya boyyy, Greatston Gnanesh, Femi Adebogun, Talal Aujan, Jonathan Leane, terasurfer, David Flickinger, William Sang, Ajan Kanaga, Vadim, Artur Olbinski, Raven Klaugh, Michael Levine, Oscar Rangel, Randy H, Cory Kujawski, RoA, Dave, Alex, Alexandros Triantafyllidis, Fen Risland, Eugene Pentland, vamX, Elle, Nathan LeClaire, Khalefa Al-Ahmad, Rainer Wilmers, subjectnull, Junyu Yang, Daniel P. Andersen, SuperWojo, LangChain4j, Mandus, Kalila, Illia Dulskyi, Trenton Dambrowitz, Asp the Wyvern, Derek Yates, Jeffrey Morgan, Deep Realms, Imad Khwaja, Pyrater, Preetika Verma, biorpg, Gabriel Tamborski, Stephen Murray, Spiking Neurons AB, Iucharbius, Chris Smitley, Willem Michiel, Luke Pendergrass, Sebastain Graf, senxiiz, Will Dee, Space Cruiser, Karl Bernard, Clay Pascal, Lone Striker, transmissions 11, webtim, WelcomeToTheClub, Sam, theTransient, Pierre Kircher, chris gileta, John Villwock, Sean Connelly, Willian Hasse
191
 
192
 
193
  Thank you to all my generous patrons and donaters!
194
 
 
 
195
  <!-- footer end -->
196
 
197
  # Original model card: Eric Hartford's Dolphin Llama 13B
@@ -224,7 +230,7 @@ We also filtered out duplicates and cleaned the data.
224
  We trained with the flan5m (gpt3.5 completions) dataset in its entirety for 3 epochs at a learning rate of 2e-5 before we stopped training to avoid overfit.
225
  We trained with the flan1m (gpt4 completions) dataset in its entirety for 2.5 epochs at a learning rate of 1e-5 before we stopped training to avoid overfit.
226
  It took about 600 hours to train on 8x H100s
227
- We used a prompt format similar to Vicuna, but we added the SYSTEM: field.
228
 
229
  Prompt format:
230
  ```
 
7
  ---
8
 
9
  <!-- header start -->
10
+ <!-- 200823 -->
11
+ <div style="width: auto; margin-left: auto; margin-right: auto">
12
+ <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
13
  </div>
14
  <div style="display: flex; justify-content: space-between; width: 100%;">
15
  <div style="display: flex; flex-direction: column; align-items: flex-start;">
16
+ <p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://discord.gg/theblokeai">Chat & support: TheBloke's Discord server</a></p>
17
  </div>
18
  <div style="display: flex; flex-direction: column; align-items: flex-end;">
19
+ <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>
20
  </div>
21
  </div>
22
+ <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>
23
+ <hr style="margin-top: 1.0em; margin-bottom: 1.0em;">
24
  <!-- header end -->
25
 
26
  # Dolphin Llama 13B - GPTQ
 
55
 
56
  | Branch | Bits | Group Size | Act Order (desc_act) | File Size | ExLlama Compatible? | Made With | Description |
57
  | ------ | ---- | ---------- | -------------------- | --------- | ------------------- | --------- | ----------- |
58
+ | main | 4 | 128 | False | 7.26 GB | True | AutoGPTQ | Most compatible option. Good inference speed in AutoGPTQ and GPTQ-for-LLaMa. Lower inference quality than other options. |
59
+ | gptq-4bit-32g-actorder_True | 4 | 32 | True | 8.00 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 32g. Gives highest possible inference quality, with maximum VRAM usage. Poor AutoGPTQ CUDA speed. |
60
+ | gptq-4bit-64g-actorder_True | 4 | 64 | True | 7.51 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 64g. Uses less VRAM than 32g, but with slightly lower accuracy. Poor AutoGPTQ CUDA speed. |
61
+ | gptq-4bit-128g-actorder_True | 4 | 128 | True | 7.26 GB | True | AutoGPTQ | 4-bit, with Act Order and group size 128g. Uses even less VRAM than 64g, but with slightly lower accuracy. Poor AutoGPTQ CUDA speed. |
62
+ | gptq-8bit--1g-actorder_True | 8 | None | True | 13.36 GB | False | AutoGPTQ | 8-bit, with Act Order. No group size, to lower VRAM requirements and to improve AutoGPTQ speed. |
63
+ | gptq-8bit-128g-actorder_False | 8 | 128 | False | 13.65 GB | False | AutoGPTQ | 8-bit, with group size 128g for higher inference quality and without Act Order to improve AutoGPTQ speed. |
64
+ | gptq-8bit-128g-actorder_True | 8 | 128 | True | 13.65 GB | False | AutoGPTQ | 8-bit, with group size 128g for higher inference quality and with Act Order for even higher accuracy. Poor AutoGPTQ CUDA speed. |
65
  | gptq-8bit-64g-actorder_True | 8 | 64 | True | 13.95 GB | False | AutoGPTQ | 8-bit, with group size 64g and Act Order for maximum inference quality. Poor AutoGPTQ CUDA speed. |
66
 
67
  ## How to download from branches
 
105
  from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
106
 
107
  model_name_or_path = "TheBloke/Dolphin-Llama-13B-GPTQ"
108
+ model_basename = "model"
109
 
110
  use_triton = False
111
 
 
169
  ExLlama works with Llama models in 4-bit. Please see the Provided Files table above for per-file compatibility.
170
 
171
  <!-- footer start -->
172
+ <!-- 200823 -->
173
  ## Discord
174
 
175
  For further support, and discussions on these models and AI in general, join us at:
 
189
  * Patreon: https://patreon.com/TheBlokeAI
190
  * Ko-Fi: https://ko-fi.com/TheBlokeAI
191
 
192
+ **Special thanks to**: Aemon Algiz.
193
 
194
+ **Patreon special mentions**: Sam, theTransient, Jonathan Leane, Steven Wood, webtim, Johann-Peter Hartmann, Geoffrey Montalvo, Gabriel Tamborski, Willem Michiel, John Villwock, Derek Yates, Mesiah Bishop, Eugene Pentland, Pieter, Chadd, Stephen Murray, Daniel P. Andersen, terasurfer, Brandon Frisco, Thomas Belote, Sid, Nathan LeClaire, Magnesian, Alps Aficionado, Stanislav Ovsiannikov, Alex, Joseph William Delisle, Nikolai Manek, Michael Davis, Junyu Yang, K, J, Spencer Kim, Stefan Sabev, Olusegun Samson, transmissions 11, Michael Levine, Cory Kujawski, Rainer Wilmers, zynix, Kalila, Luke @flexchar, Ajan Kanaga, Mandus, vamX, Ai Maven, Mano Prime, Matthew Berman, subjectnull, Vitor Caleffi, Clay Pascal, biorpg, alfie_i, 阿明, Jeffrey Morgan, ya boyyy, Raymond Fosdick, knownsqashed, Olakabola, Leonard Tan, ReadyPlayerEmma, Enrico Ros, Dave, Talal Aujan, Illia Dulskyi, Sean Connelly, senxiiz, Artur Olbinski, Elle, Raven Klaugh, Fen Risland, Deep Realms, Imad Khwaja, Fred von Graf, Will Dee, usrbinkat, SuperWojo, Alexandros Triantafyllidis, Swaroop Kallakuri, Dan Guido, John Detwiler, Pedro Madruga, Iucharbius, Viktor Bowallius, Asp the Wyvern, Edmond Seymore, Trenton Dambrowitz, Space Cruiser, Spiking Neurons AB, Pyrater, LangChain4j, Tony Hughes, Kacper Wikieł, Rishabh Srivastava, David Ziegler, Luke Pendergrass, Andrey, Gabriel Puliatti, Lone Striker, Sebastain Graf, Pierre Kircher, Randy H, NimbleBox.ai, Vadim, danny, Deo Leter
195
 
196
 
197
  Thank you to all my generous patrons and donaters!
198
 
199
+ And thank you again to a16z for their generous grant.
200
+
201
  <!-- footer end -->
202
 
203
  # Original model card: Eric Hartford's Dolphin Llama 13B
 
230
  We trained with the flan5m (gpt3.5 completions) dataset in its entirety for 3 epochs at a learning rate of 2e-5 before we stopped training to avoid overfit.
231
  We trained with the flan1m (gpt4 completions) dataset in its entirety for 2.5 epochs at a learning rate of 1e-5 before we stopped training to avoid overfit.
232
  It took about 600 hours to train on 8x H100s
233
+ We used a prompt format similar to Vicuna, but we added the SYSTEM: field.
234
 
235
  Prompt format:
236
  ```
config.json CHANGED
@@ -1,26 +1,37 @@
1
  {
2
- "architectures": [
3
- "LlamaForCausalLM"
4
- ],
5
- "bos_token_id": 1,
6
- "eos_token_id": 2,
7
- "hidden_act": "silu",
8
- "hidden_size": 5120,
9
- "initializer_range": 0.02,
10
- "intermediate_size": 13824,
11
- "max_position_embeddings": 2048,
12
- "max_sequence_length": 2048,
13
- "model_type": "llama",
14
- "num_attention_heads": 40,
15
- "num_hidden_layers": 40,
16
- "num_key_value_heads": 40,
17
- "pad_token_id": 0,
18
- "pretraining_tp": 1,
19
- "rms_norm_eps": 1e-06,
20
- "rope_scaling": null,
21
- "tie_word_embeddings": false,
22
- "torch_dtype": "float16",
23
- "transformers_version": "4.32.0.dev0",
24
- "use_cache": true,
25
- "vocab_size": 32000
 
 
 
 
 
 
 
 
 
 
 
26
  }
 
1
  {
2
+ "architectures": [
3
+ "LlamaForCausalLM"
4
+ ],
5
+ "bos_token_id": 1,
6
+ "eos_token_id": 2,
7
+ "hidden_act": "silu",
8
+ "hidden_size": 5120,
9
+ "initializer_range": 0.02,
10
+ "intermediate_size": 13824,
11
+ "max_position_embeddings": 2048,
12
+ "max_sequence_length": 2048,
13
+ "model_type": "llama",
14
+ "num_attention_heads": 40,
15
+ "num_hidden_layers": 40,
16
+ "num_key_value_heads": 40,
17
+ "pad_token_id": 0,
18
+ "pretraining_tp": 1,
19
+ "rms_norm_eps": 1e-06,
20
+ "rope_scaling": null,
21
+ "tie_word_embeddings": false,
22
+ "torch_dtype": "float16",
23
+ "transformers_version": "4.32.0.dev0",
24
+ "use_cache": true,
25
+ "vocab_size": 32000,
26
+ "quantization_config": {
27
+ "bits": 4,
28
+ "group_size": 128,
29
+ "damp_percent": 0.1,
30
+ "desc_act": false,
31
+ "sym": true,
32
+ "true_sequential": true,
33
+ "model_name_or_path": null,
34
+ "model_file_base_name": "model",
35
+ "quant_method": "gptq"
36
+ }
37
  }
gptq_model-4bit-128g.safetensors → model.safetensors RENAMED
File without changes
quantize_config.json CHANGED
@@ -6,5 +6,5 @@
6
  "sym": true,
7
  "true_sequential": true,
8
  "model_name_or_path": null,
9
- "model_file_base_name": null
10
  }
 
6
  "sym": true,
7
  "true_sequential": true,
8
  "model_name_or_path": null,
9
+ "model_file_base_name": "model"
10
  }