title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns] | url
stringlengths 0
780
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns] | gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Putting the pieces together for RAG? | 1 | [removed] | 2023-08-10T07:05:30 | https://www.reddit.com/r/LocalLLaMA/comments/15n4yso/putting_the_pieces_together_for_rag/ | noellarkin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n4yso | false | null | t3_15n4yso | /r/LocalLLaMA/comments/15n4yso/putting_the_pieces_together_for_rag/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'ozzoOOUo0GsTZQDFKQGLV_Hm-3-sAMBPGUdj-YGAXaU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=108&crop=smart&auto=webp&s=01abd3e1c16817d00a13b4aefae81016db13aa6d', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=216&crop=smart&auto=webp&s=80098a7b6d0ab3b52bcff8841b1c97195818826e', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=320&crop=smart&auto=webp&s=6287874f84b7432e9be49420fccd9bdc36e26a2b', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=640&crop=smart&auto=webp&s=9543d5f1f15bd29dd86f3124e13d2a14c896185a', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=960&crop=smart&auto=webp&s=13e07ecd94404eb99b26585e2886d42c371183af', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?width=1080&crop=smart&auto=webp&s=fc6c1f4554adbf733744434d61fc12f3be4f1a57', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/eKdlDfZLsUFa_rEq3sS4i1bIQWiKS6af16vu7A-M4fM.jpg?auto=webp&s=d91bb719706fadcf313dc369ff3d4f31373466d7', 'width': 1200}, 'variants': {}}]} |
What determines the length of the response? | 1 | while the context size determines how much information can be considered while fashioning a response, what factors determine the length of the output?
- a (unsubstantiated) gut feeling is that models with larger context sizes should be able to make longer responses?
- a [post](https://www.reddit.com/r/LocalLLaMA/comments/15mu7um/sillytaverns_roleplay_preset_vs_modelspecific/?ref=share&ref_source=link) finds that a slight variation of the prompt increases verbosity. is there a general pattern/ practice here that is worth trying for all models.
- this [post](https://www.reddit.com/r/LocalLLaMA/comments/15g5419/what_ive_learned_from_orcamini3bggmlv3q4_1_using/) finds that some models are susceptible to unfinished responses, even if they "seem" complete. has this been noticed in general, and any recommendations for the model to furnish complete responses?
- are there any tests/experiments that you feel would be interesting to try out to understand the verbosity limits of llms | 2023-08-10T07:58:28 | https://www.reddit.com/r/LocalLLaMA/comments/15n5vyz/what_determines_the_length_of_the_response/ | T_hank | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n5vyz | false | null | t3_15n5vyz | /r/LocalLLaMA/comments/15n5vyz/what_determines_the_length_of_the_response/ | false | false | self | 1 | null |
Optimizing models for LLM Leaderboard is a HUGE mistake | 1 | Cause that has a preassumption of being a "good" model means ranking high in 4 different relatively controversial benchmarking suites.
Its a popular wrongdoing in the Kaggle competitions to optimize for the test set. Leaderboard should be seen as the test set that we see how our models compare after the model released, rather than some sort of goal to create ensembled models to rank higher.
You can see this in some random kaggle competition, where people can see limited eval set results, but when they face with the whole test set, higher ranking models tanking in the rankings for the full suite, and some model in the middle of the test set wins.
The problem is higher ranking models are often overfit on specific examples in the test set and that doesn't necessarily means the model will be percepted as good in terms of humans. That's why the models that claim they are better than GPT 3.5 on eval suite X often not as good as its advertised when tested on real world human prompts.
Without a doubt, OpenAI have their own extensive eval suite and even that wasn't enough, so they created "openai/evals" repo to get free evaluation from the community.
We need better evaluation metrics than LLM leaderboard for the Open-source community. | 2023-08-10T08:24:48 | https://www.reddit.com/r/LocalLLaMA/comments/15n6cmb/optimizing_models_for_llm_leaderboard_is_a_huge/ | yiyecek | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n6cmb | false | null | t3_15n6cmb | /r/LocalLLaMA/comments/15n6cmb/optimizing_models_for_llm_leaderboard_is_a_huge/ | false | false | self | 1 | null |
What is the best way to fine tune llama2 for X/tweet sentiment analysis? | 3 | While there are BERT based models (e.g., [this one for twitter](https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment-latest)), I was wondering whether llama2 can be fine-tuned to provide a more accurate sentiment analysis? If so, can you help guide how a novice can go about this? | 2023-08-10T08:49:57 | https://www.reddit.com/r/LocalLLaMA/comments/15n6sgt/what_is_the_best_way_to_fine_tune_llama2_for/ | sbs1799 | self.LocalLLaMA | 2023-08-10T09:44:11 | 0 | {} | 15n6sgt | false | null | t3_15n6sgt | /r/LocalLLaMA/comments/15n6sgt/what_is_the_best_way_to_fine_tune_llama2_for/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': 'ArlWYQviB_-tLNpMl6U7rDe8OGn86GN62b1DJBtybDk', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=108&crop=smart&auto=webp&s=830a41f7ded868cb0f1f029252d1ae2d9871eda3', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=216&crop=smart&auto=webp&s=2abbb1461e3b451083191545260db891894e5557', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=320&crop=smart&auto=webp&s=f0333448132cff266a52926b4bf31ad3dc8f097a', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=640&crop=smart&auto=webp&s=e8e32e5edfc032544cb866e0960c922b087cbd11', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=960&crop=smart&auto=webp&s=e14e83dde369a57d62b44688cf0caee00f1b0706', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?width=1080&crop=smart&auto=webp&s=aea0b286447212126101824b07a9209e381c7fb6', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/gKS9gx0oKZdnygGyNyJkxboMlVq9S_SaVs7Pr00b4w4.jpg?auto=webp&s=25fd87897661a5bbff40927d7e5d3b9266fdb6d2', 'width': 1200}, 'variants': {}}]} |
How i can train llama 2? | 1 | [removed] | 2023-08-10T09:07:29 | https://www.reddit.com/r/LocalLLaMA/comments/15n743e/how_i_can_train_llama_2/ | Agreeable_Fun7280 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n743e | false | null | t3_15n743e | /r/LocalLLaMA/comments/15n743e/how_i_can_train_llama_2/ | false | false | self | 1 | null |
Is there a way to read the best papers/blogs/research on LLama? | 1 | I am trying to get into a habit of reading a paper every couple days and was wondering if there is any curated website or source that keeps track of latest papers coming out.
So far I have just been using twitter to try to follow people who are working in this area. | 2023-08-10T09:22:16 | https://www.reddit.com/r/LocalLLaMA/comments/15n7do9/is_there_a_way_to_read_the_best/ | mohit_habeeb | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n7do9 | false | null | t3_15n7do9 | /r/LocalLLaMA/comments/15n7do9/is_there_a_way_to_read_the_best/ | false | false | self | 1 | null |
Using prompt templates after LoRA on raw text | 6 | Hello all,
I'm a little overwhelmed with all the developments and I feel like I don't know where to begin. So I apologise if this question sounds very basic.
Let's say I want my LLM to sound like Phoebe Buffay from Friends. I don't have a QnA format, but just raw text for this purpose.
As I understand, I can perform LoRA using the WebUI.
Once my fine tuned model is ready, I want to use this to be able to converse with the user using specific prompts.
My question is, can I feed this fine tuned model to LangChain so I can use their prompt template successfully? Or are there alternatives?
Or can I do all of this using HuggingFace?
Sorry, I'm very lost and I can't seem to understand if the finetuned models can be used by other frameworks. | 2023-08-10T10:02:54 | https://www.reddit.com/r/LocalLLaMA/comments/15n84m8/using_prompt_templates_after_lora_on_raw_text/ | betib25 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n84m8 | false | null | t3_15n84m8 | /r/LocalLLaMA/comments/15n84m8/using_prompt_templates_after_lora_on_raw_text/ | false | false | self | 6 | null |
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-10T11:12:42 | https://www.reddit.com/r/LocalLLaMA/comments/15n9hx7/what_is_needed_to_finetune_llama_2_70b_without/ | pet_vaginal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15n9hx7 | false | null | t3_15n9hx7 | /r/LocalLLaMA/comments/15n9hx7/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'WyBK-cNZtYI8e2V2a6UpK1uX-x5RQjkIvLQveCmvURc', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=108&crop=smart&auto=webp&s=b7246ca224b849e17307af1e38e1da252729741a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=216&crop=smart&auto=webp&s=c4e7fb8ffbd5f6ed63aff4084f4af6af1f546afd', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=320&crop=smart&auto=webp&s=cacc08d69fea74681e002bc9c69237c79577c864', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=640&crop=smart&auto=webp&s=fc29395b17387c905bdf56012d9e74408d167f0a', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=960&crop=smart&auto=webp&s=ec50388eca0806ee57992819fba448b30dd5e162', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?width=1080&crop=smart&auto=webp&s=d22d5b81cd534e776ab1bd6b9ad5e47f5ea62cbb', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/bOwmyl6KOPvnXk2ApJoeApQZ4Q-EzWXXPl75WOn-AEY.jpg?auto=webp&s=2d758e78264dd4f726e4534208207b8a95cdf09d', 'width': 1200}, 'variants': {}}]} |
help to update Llama.cpp | 1 | [removed] | 2023-08-10T11:40:26 | https://www.reddit.com/r/LocalLLaMA/comments/15na2ph/help_to_update_llamacpp/ | Competitive_Fox7811 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15na2ph | false | null | t3_15na2ph | /r/LocalLLaMA/comments/15na2ph/help_to_update_llamacpp/ | false | false | 1 | null |
|
My experience with LLaMA 2 | 1 | [removed] | 2023-08-10T11:55:47 | https://www.reddit.com/r/LocalLLaMA/comments/15naepn/my_experience_with_llama_2/ | uralstech_MR | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15naepn | false | null | t3_15naepn | /r/LocalLLaMA/comments/15naepn/my_experience_with_llama_2/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'DQ0NpM_glprnLD-GOVWRivLTqpaMfa4wZh0dDiD1wnw', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=108&crop=smart&auto=webp&s=7c5eb57ce92cb53dd18f9e277fbe1234ad3d6487', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=216&crop=smart&auto=webp&s=7b0f11a7102c7d2b009c900fefa11aea2ea7bea1', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=320&crop=smart&auto=webp&s=95cd48d7c84f190be5635db0a604de8c9879f4b1', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=640&crop=smart&auto=webp&s=e1c1723f70c20e49d6dd6d783dcfc491539f2448', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=960&crop=smart&auto=webp&s=fd81c83384eaa28eac658f10cacf9a2cfb910836', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?width=1080&crop=smart&auto=webp&s=1bc3462ab583b926bf14849085900715579c9176', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/o9AjqIWXV4Hp56E3k-B1szuU63M5WgJ_51ZPTn2SweY.jpg?auto=webp&s=36661dfa3d4bb20c5d9701be7671389e880244e0', 'width': 1200}, 'variants': {}}]} |
Is Embeddings are giving better answer that finetuned model? | 1 | I am making a chatbot for a company which has 150 PDFs and it contains 1000 of pages of data in text format. My expectations from the chatbot is giving me not only simple answers as in the PDF. I want logical answers from it which is related to topics in PDF and model is determine an answer using data from PDF.
I tried making embedding of those PDFs but I am not getting satisfied answer using embeddings so if I finetune LLM model using PDFs data then chatbot gives me better answer or not? | 2023-08-10T12:03:05 | https://www.reddit.com/r/LocalLLaMA/comments/15nakwz/is_embeddings_are_giving_better_answer_that/ | darshil-3099 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nakwz | false | null | t3_15nakwz | /r/LocalLLaMA/comments/15nakwz/is_embeddings_are_giving_better_answer_that/ | false | false | self | 1 | null |
Is Embeddings are giving better answer than finetuned model? | 1 | I am making a chatbot for a company which has 150 PDFs and it contains 1000 of pages of data in text format. My expectations from the chatbot is giving me not only simple answers as in the PDF. I want logical answers from it which is related to topics in PDF and model is determine an answer using data from PDF.
I tried making embedding of those PDFs but I am not getting satisfied answer using embeddings so if I finetune LLM model using PDFs data then chatbot gives me better answer or not ? | 2023-08-10T12:15:55 | https://www.reddit.com/r/LocalLLaMA/comments/15nav9p/is_embeddings_are_giving_better_answer_than/ | darshil-3099 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nav9p | false | null | t3_15nav9p | /r/LocalLLaMA/comments/15nav9p/is_embeddings_are_giving_better_answer_than/ | false | false | self | 1 | null |
How do you use LLama(2)? | 1 | [removed] | 2023-08-10T12:59:49 | https://www.reddit.com/r/LocalLLaMA/comments/15nbv6o/how_do_you_use_llama2/ | SecretOk9644 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nbv6o | false | null | t3_15nbv6o | /r/LocalLLaMA/comments/15nbv6o/how_do_you_use_llama2/ | false | false | self | 1 | null |
Llama 2 goes to school - a Wikipedia exploration, vector DB ingestion, and chat experiment | 1 | [removed] | 2023-08-10T13:28:57 | https://www.reddit.com/r/LocalLLaMA/comments/15nckm3/llama_2_goes_to_school_a_wikipedia_exploration/ | crono760 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nckm3 | false | null | t3_15nckm3 | /r/LocalLLaMA/comments/15nckm3/llama_2_goes_to_school_a_wikipedia_exploration/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'Tay9bFq98NoSUfgn0x33E-XJc10Y1CdW2jllu-yQNOQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=108&crop=smart&auto=webp&s=9af61d04cc3d85e9fc3d69da584dffe45dc04636', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=216&crop=smart&auto=webp&s=fb7f24b3629ed59f39a7087a325f4c85761a2989', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=320&crop=smart&auto=webp&s=b97ef3f547d53250d7da0143b5485b3f6a1d30d2', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=640&crop=smart&auto=webp&s=b4f87904b7a071458ed7776e33677d1f02a724bb', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=960&crop=smart&auto=webp&s=34852c2d2d57ace507dcc8cc7dc8f27a3198a413', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?width=1080&crop=smart&auto=webp&s=55624a1e90cefe0b74bc2c83af493549ed68ce36', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/EGpWnUW_GGRKumcTEmimGWiEfU8hCFPrgnIZ7m5gVkE.jpg?auto=webp&s=732d0287c1cb02f873eed39833b1fd8c4fa20791', 'width': 1200}, 'variants': {}}]} |
Does server motherboards with dual cpu run dobule the speed compare to only one cpu since dual cpu has double the ram slots? | 1 | So I'm planning to build a PC to run localLLama with some used server CPU.
I'm planning to either buy 1 used 2nd gen epyc cpu with 8 channel ram or 2 xeon gold CPUs with 6 channel ram and use a dual cpu motherboard.
My question is will a 2 cpu with 6 channel faster than a epyc 8 channel cpu since they could use 2\*6=12 ram slots?
​ | 2023-08-10T13:36:08 | https://www.reddit.com/r/LocalLLaMA/comments/15ncr2k/does_server_motherboards_with_dual_cpu_run_dobule/ | hoseex999 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ncr2k | false | null | t3_15ncr2k | /r/LocalLLaMA/comments/15ncr2k/does_server_motherboards_with_dual_cpu_run_dobule/ | false | false | self | 1 | null |
Summarizing content prompt ignored because of content. | 1 | I'm working with the llama-2-13b-chat.ggmlv3.q4\_0 model to try to summarize content for SEO. It seems to work well for simple content, though if the page has a form on it, the actions and questions of the form tend to hijack the response and not return in JSON format, and start chatting as if I it's responding as the page itself. I found the suggested prompt template and it helped with simple pages but not more complex pages. Below is one of the latest prompts I've tried that just fails. I could be using the wrong model, but I'm hoping I can keep at the 13b as it fits easy on my laptop to run. Is the prompt bad, or is this a setting issue, or the wrong model?
`<s>[INST] <<SYS>>`
`Assistant is a expert JSON builder designed to assist with a creating SEO headlines and descriptions.`
`Assistant is able to respond to the User and use tools using JSON strings that contain 'headline' and 'description' parameters.`
`All of Assistant's communication is performed using this JSON format.`
`Here are some previous summarizing between the Assistant and User:`
`User: Sign our petition! Sign your name for our petition to support us.`
`Assistant: \`\`\`json`
`{{"headline": "Sign our petition",`
`"description": "Sign your name to support our causes"}}`
`\`\`\``
`User: Complete our survey --- What issues impact you the most. Complete the survey`
`Assistant: \`\`\`json`
`{{"headline": "Complete our Survey",`
`"description": "Tell us what issues impact you the most"}}`
`\`\`\``
`<</SYS>>`
`Here is the text of the web page to create the SEO content for.`
Using the lamma.cpp server `.\server.exe -m .\models\llama-2-13b-chat.ggmlv3.q4_0.bin --mlock --n-gpu-layers 30` to run this.
​ | 2023-08-10T14:55:05 | https://www.reddit.com/r/LocalLLaMA/comments/15nepxb/summarizing_content_prompt_ignored_because_of/ | jbartley | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nepxb | false | null | t3_15nepxb | /r/LocalLLaMA/comments/15nepxb/summarizing_content_prompt_ignored_because_of/ | false | false | self | 1 | null |
Any tips for faster generation for queries with large fixed context/system prompts? | 1 | [removed] | 2023-08-10T15:16:01 | https://www.reddit.com/r/LocalLLaMA/comments/15nf8x0/any_tips_for_faster_generation_for_queries_with/ | greying_panda | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nf8x0 | false | null | t3_15nf8x0 | /r/LocalLLaMA/comments/15nf8x0/any_tips_for_faster_generation_for_queries_with/ | false | false | self | 1 | null |
Which Azure Cloud GPU is best for fine tuning? | 1 | The GPU names are different, but are they usually all A100's but just emulated into smaller GPUs? | 2023-08-10T15:44:02 | https://www.reddit.com/r/LocalLLaMA/comments/15nfyt7/which_azure_cloud_gpu_is_best_for_fine_tuning/ | Shyvadi | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nfyt7 | false | null | t3_15nfyt7 | /r/LocalLLaMA/comments/15nfyt7/which_azure_cloud_gpu_is_best_for_fine_tuning/ | false | false | self | 1 | null |
Is oobabooga supposed to be this damn slow? | 1 | Hi, I'm a newb, so I'm not sure if I am putting this in the right place...
I would like to bring LLama (or any other model) local to experiment with finetuning, etc.
I don't have a beast of a machine, but it seems pretty good to me. It is an Asus, ROG Running:
AMD Ryzen 9 5900HX with Radeon Graphics 3.30 GHz
NVidia GeForce RTX 3070 GPU GDDR6 @ 8GB (256 Bit) RAM
64.0 GB System RAM
4 TB of SSD storage
I'm thinking I **\*should\*** be able to run a 7B model locally, right?
I followed the steps to set up Oobabooga. I load a 7B model from TheBloke.
Everything seems fine. I type in a question, and I watch the output in the Powershell. The output shows up reasonably quickly.
But the response inside of the web ui is ***painful***. Like... Every.... Word... comes... up... 2 seconds... apart.
I don't get it. Doesn't it have its answer as soon at the output is returned? Why so slow? What am I missing?
​ | 2023-08-10T16:02:47 | https://www.reddit.com/r/LocalLLaMA/comments/15nggns/is_oobabooga_supposed_to_be_this_damn_slow/ | Ok_Organization_4295 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nggns | false | null | t3_15nggns | /r/LocalLLaMA/comments/15nggns/is_oobabooga_supposed_to_be_this_damn_slow/ | false | false | self | 1 | null |
General strategies on how to refine a model to a specific dataset while keeping the NLP-ness? | 1 | I've managed to train with documents from an internal knowledge base. But want to limit users from asking off topic questions "What is the capital of Iceland?" | 2023-08-10T16:11:27 | https://www.reddit.com/r/LocalLLaMA/comments/15ngot3/general_strategies_on_how_to_refine_a_model_to_a/ | chris480 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ngot3 | false | null | t3_15ngot3 | /r/LocalLLaMA/comments/15ngot3/general_strategies_on_how_to_refine_a_model_to_a/ | false | false | self | 1 | null |
Where's the best community to get help with tortoise TTS? | 1 | [removed] | 2023-08-10T16:36:01 | https://www.reddit.com/r/LocalLLaMA/comments/15nhbm7/wheres_the_best_community_to_get_help_with/ | Virtamancer | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nhbm7 | false | null | t3_15nhbm7 | /r/LocalLLaMA/comments/15nhbm7/wheres_the_best_community_to_get_help_with/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'R1XyIfwJNMprA3q8KUhj7XqSkvuDgfnG7_QKPcJ8kNU', 'resolutions': [], 'source': {'height': 80, 'url': 'https://external-preview.redd.it/eSEXxiI7z3-IUxI_gRwm0PH6HgIOm1-iAuzEbmuJ3ow.jpg?auto=webp&s=2c3ce4d6960f8deaa4baa31576edfc2d236b8ba5', 'width': 80}, 'variants': {}}]} |
Did someone calculate how fast GPT-3.5 is, in terms of tokens/second? And how about GPT-4? | 1 | I read this thread [Inference Speed for Llama 2 70b on A6000 with Exllama - Need Suggestions!](https://www.reddit.com/r/LocalLLaMA/comments/15lf119/inference_speed_for_llama_2_70b_on_a6000_with/?utm_source=share&utm_medium=web2x&context=3), and I think 10 tokens per second is not bad. But I'm not sure how fast it is. So, I want to check its speed with GPT-3.5 and GPT-4. Anyone have a concrete number? Or if I want to measure it manually, how should I do? A stopwatch? Any suggestion? Thank you guys. | 2023-08-10T17:19:02 | https://www.reddit.com/r/LocalLLaMA/comments/15nig1k/did_someone_calculate_how_fast_gpt35_is_in_terms/ | Fancy-Welcome-9064 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nig1k | false | null | t3_15nig1k | /r/LocalLLaMA/comments/15nig1k/did_someone_calculate_how_fast_gpt35_is_in_terms/ | false | false | self | 1 | null |
Multimodal-app with video multimodal support? (4bit GPTQ)? | 1 | Is there any app like [https://github.com/DAMO-NLP-SG/Video-LLaMA.git](https://github.com/DAMO-NLP-SG/Video-LLaMA.git) or [https://github.com/X-PLUG/mPLUG-Owl](https://github.com/X-PLUG/mPLUG-Owl) that also has GPTQ 4bit support for low VRAM users?
Oobabooga only has image mutimodal support. I really want video support with 4bit GPTQ, so it will fit a 8gb GPU. | 2023-08-10T17:20:19 | https://www.reddit.com/r/LocalLLaMA/comments/15nihan/multimodalapp_with_video_multimodal_support_4bit/ | fetballe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nihan | false | null | t3_15nihan | /r/LocalLLaMA/comments/15nihan/multimodalapp_with_video_multimodal_support_4bit/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'vTaZURLh7-5AXEQ0JeE73YLxXFzAqcxN2dUl_wghjSQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=108&crop=smart&auto=webp&s=9444bf91b799d9c94f126ccbf88d96b4dac446c7', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=216&crop=smart&auto=webp&s=9b2e61d158ebecbc348ce548f0f7f1d17185fed7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=320&crop=smart&auto=webp&s=8a757d9400e4e72fa03726e7c954c5669cf3ed86', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=640&crop=smart&auto=webp&s=9a7a3a6ba291745a2ab656abb4688ae58ef52acc', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=960&crop=smart&auto=webp&s=1419dcddfaac32d71e9b45c08e9949751cf0afc3', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?width=1080&crop=smart&auto=webp&s=e37a6e400425e9f491d7151c923f76727b02edf3', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/wN2q2kGo453e9f_co110shSr0JDCbXcz1qXfkQGrTr4.jpg?auto=webp&s=871558cd8817ebc0b4779294b05c01539d9f3675', 'width': 1200}, 'variants': {}}]} |
How to utilize memory, author's notes effectively in KoboldCPP? | 1 | So I've noticed that when I select a preset in kcpp and visit the memory settings it has a lot of pre-filled text in there, and frequently there is [ ] (square brackets) surrounding various bits of text, and often there is a format like
person: their_name; attitude: their_attitudes; # etc
So it got me curious, is there a specific format that should be used when entering data/details into the memory portion?
I've also noticed that when I select the presets there is rarely (if ever) anything populated in the author's notes box. I'm assuming this isn't as important then? | 2023-08-10T17:21:31 | https://www.reddit.com/r/LocalLLaMA/comments/15niihq/how_to_utilize_memory_authors_notes_effectively/ | wh33t | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15niihq | false | null | t3_15niihq | /r/LocalLLaMA/comments/15niihq/how_to_utilize_memory_authors_notes_effectively/ | false | false | self | 1 | null |
Integrating Microsoft’s Guidance into hiyouga/LLaMA-Efficient-Tuning's repo for Faster Processing and Well-Formatted Output | 1 | Hey, do you know where I can find resources on how to integrate Microsoft’s Guidance into hiyouga/LLaMA-Efficient-Tuning's repo? It says it speeds up the process, which can be super helpful with these resource-heavy beasts, and it is very useful for getting well-formatted output | 2023-08-10T17:26:07 | https://www.reddit.com/r/LocalLLaMA/comments/15nimvl/integrating_microsofts_guidance_into/ | ILoveBoxPlots | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nimvl | false | null | t3_15nimvl | /r/LocalLLaMA/comments/15nimvl/integrating_microsofts_guidance_into/ | false | false | self | 1 | null |
How make Llama2 GPTQ faster in generation? | 1 | I use [this](https://huggingface.co/TheBloke/Llama-2-13B-chat-GPTQ) model on T4 gpu and speed is modest for generation, is there anyway to make it faster? | 2023-08-10T17:52:27 | https://www.reddit.com/r/LocalLLaMA/comments/15njbw1/how_make_llama2_gptq_faster_in_generation/ | Euphoric-Nebula-4559 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15njbw1 | false | null | t3_15njbw1 | /r/LocalLLaMA/comments/15njbw1/how_make_llama2_gptq_faster_in_generation/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'rkljmc1ht4wmaJGm60zJkcYTjcDJ3O6Ma8y8dDYwHYI', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=108&crop=smart&auto=webp&s=1a72ffb8bd5cc3be930da7d37436aeb9871f39b6', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=216&crop=smart&auto=webp&s=4834bb0a4a1a1893261f169f663dc5f8caf3aa24', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=320&crop=smart&auto=webp&s=4d244a1a632999ce7d2c4f476dffd40a57edcd59', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=640&crop=smart&auto=webp&s=68e3dbfa5c777b7d5c5c7d805373cb4aa0562544', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=960&crop=smart&auto=webp&s=57a63c907125ad8e9d155efefda403fc3c21b9f1', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?width=1080&crop=smart&auto=webp&s=a9394eb2b2002d2e590a60251c3b3fcdbe7afc8a', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/vBvc7FHJE65n0rhjEdf2N4Vhy4vMT0IFNxHSqGQObp8.jpg?auto=webp&s=a07479d4b0463cc023a75d5210c525e3dac0d4b2', 'width': 1200}, 'variants': {}}]} |
Here is what the EU considers risky and wants to regulate. Text taken directly from EU's site, link provided. | 1 | The text isn't all bad, but I do find some of it concerning. Notably:
##### Banned actions
>Unacceptable risk AI systems are systems considered a threat to people and will be banned.
>Real-time and remote biometric identification systems, such as facial recognition
This is a major security issue. Not all data should be open to everyone. Security is moving to passkeys which require bio authentication
##### Must be registered
>AI systems falling into eight specific areas that will have to be registered in an EU database:
>Education and vocational training
An educational model must be registered. That's pretty much everything. This will come down to wording in legislation. But is currently concerning.
##### Generative AI
>Generative AI, like ChatGPT, would have to comply with transparency requirements:
>Disclosing that the content was generated by AI
Designing the model to prevent it from generating illegal content
Publishing summaries of copyrighted data used for training
Autocomplete is AI -- it uses the same training architechure as ChatGPT. Most photo tools are AI based. There are so many tools out there that are using AI that nothing will be considered yours.
---
https://www.europarl.europa.eu/news/en/headlines/society/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence
AI Act: different rules for different risk levels
The new rules establish obligations for providers and users depending on the level of risk from artificial intelligence. While many AI systems pose minimal risk, they need to be assessed.
Unacceptable risk
Unacceptable risk AI systems are systems considered a threat to people and will be banned. They include:
Cognitive behavioural manipulation of people or specific vulnerable groups: for example voice-activated toys that encourage dangerous behaviour in children
Social scoring: classifying people based on behaviour, socio-economic status or personal characteristics
Real-time and remote biometric identification systems, such as facial recognition
Some exceptions may be allowed: For instance, “post” remote biometric identification systems where identification occurs after a significant delay will be allowed to prosecute serious crimes but only after court approval.
High risk
AI systems that negatively affect safety or fundamental rights will be considered high risk and will be divided into two categories:
1) AI systems that are used in products falling under the EU’s product safety legislation. This includes toys, aviation, cars, medical devices and lifts.
2) AI systems falling into eight specific areas that will have to be registered in an EU database:
Biometric identification and categorisation of natural persons
Management and operation of critical infrastructure
Education and vocational training
Employment, worker management and access to self-employment
Access to and enjoyment of essential private services and public services and benefits
Law enforcement
Migration, asylum and border control management
Assistance in legal interpretation and application of the law.
All high-risk AI systems will be assessed before being put on the market and also throughout their lifecycle.
Generative AI
Generative AI, like ChatGPT, would have to comply with transparency requirements:
Disclosing that the content was generated by AI
Designing the model to prevent it from generating illegal content
Publishing summaries of copyrighted data used for training
Limited risk
Limited risk AI systems should comply with minimal transparency requirements that would allow users to make informed decisions. After interacting with the applications, the user can then decide whether they want to continue using it. Users should be made aware when they are interacting with AI. This includes AI systems that generate or manipulate image, audio or video content, for example deepfakes. | 2023-08-10T18:07:28 | https://www.reddit.com/r/LocalLLaMA/comments/15njqpd/here_is_what_the_eu_considers_risky_and_wants_to/ | jayfehr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15njqpd | false | null | t3_15njqpd | /r/LocalLLaMA/comments/15njqpd/here_is_what_the_eu_considers_risky_and_wants_to/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'dvY7fvPhlVsoTxqCUAaSr_wr2lTeGds8ENFwBCMZTGA', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=108&crop=smart&auto=webp&s=d2706306599f52edc15fb8be036a64f98ae4cbcc', 'width': 108}, {'height': 144, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=216&crop=smart&auto=webp&s=504df8c65372e65e35f3e9209c9555a645cfb015', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=320&crop=smart&auto=webp&s=6f6df8e6d80348b2378abb6323a08d488862e23c', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=640&crop=smart&auto=webp&s=a2ff5e00e5b5887139fe465c7a557fa09d6bc148', 'width': 640}, {'height': 640, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=960&crop=smart&auto=webp&s=958aea40c9f2f6c4c3902cddecb1886083bb7198', 'width': 960}, {'height': 720, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?width=1080&crop=smart&auto=webp&s=99ba56b430c6e5270cddeb6aaeb816a49c27980f', 'width': 1080}], 'source': {'height': 1200, 'url': 'https://external-preview.redd.it/HC_1mvqThFmrVOtO-sR4ZDVqiZlPecbdp-IX6KDVSpw.jpg?auto=webp&s=b11e87b9c10cdbf975fe952609b77fd79ad8c051', 'width': 1800}, 'variants': {}}]} |
Are AMD Ryzen Processors Supported for Running Uncensored AI Models ? | 1 | [removed] | 2023-08-10T19:01:13 | https://www.reddit.com/r/LocalLLaMA/comments/15nl4uh/are_amd_ryzen_processors_supported_for_running/ | No_Anywhere2560 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nl4uh | false | null | t3_15nl4uh | /r/LocalLLaMA/comments/15nl4uh/are_amd_ryzen_processors_supported_for_running/ | false | false | self | 1 | null |
How I ran Llama 2 on my Mac in < 5 mins | 1 | [removed] | 2023-08-10T19:10:45 | https://www.reddit.com/r/LocalLLaMA/comments/15nle0k/how_i_ran_llama_2_on_my_mac_in_5_mins/ | rajatarya | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nle0k | false | null | t3_15nle0k | /r/LocalLLaMA/comments/15nle0k/how_i_ran_llama_2_on_my_mac_in_5_mins/ | false | false | self | 1 | null |
How to finetune a 70B model using A100? | 1 | Hello, I would like to finetune (LoRA or full SFT) a 60B (LLaMA 1) or 70B (LLaMA 2) model on my custom data but I could not find a way to do so (keep throwing me out of memory). Thus I'd like to hear from you guys any suggestions.
\- I have a custom \~300k data consisting of my private data + some from [GPTeacher](https://github.com/teknium1/GPTeacher). The context window is 2k
\- I can rent 8x 80G A100. But I can only rent it for at most a week otherwise it gets overbudget.
I use huggingface transformers (as most of ppl do) and its [trainer](https://huggingface.co/docs/transformers/main_classes/trainer) and tried
\- SFT with Deepspeed ZERO 3 + CPU Offload, with [FlashAttention copied from Vicuna](https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py), gradient checkpoint, and bf16 mixed training. But this gives me out of memory error even with batch size = 1.
\- Then I tried LoRA with same config as [Alpaca-LoRA](https://github.com/tloen/alpaca-lora/tree/main), but again I got out of memory error
\- I haven't tried QLoRA but from my past experience with QLoRA: I think it is impossible to use Deepspeed with QLoRA thus I can only use 1 gpu. Last month I tried training a 13B model with QLoRA on \~100k data but already took me >60 hours so I can't imagine how expensive it is to use QLoRA on 70B model with 300k data. But I can be wrong tho.
​
From my understanding, a 70B model takes at least 140B for bf16 just for holding it, let alone training. So a 80G A100 cannot even do inference with bf16. Therefore I am looking for methods that use parallelism. Thanks!!!! | 2023-08-10T19:43:11 | https://www.reddit.com/r/LocalLLaMA/comments/15nm8r4/how_to_finetune_a_70b_model_using_a100/ | tt19234 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nm8r4 | false | null | t3_15nm8r4 | /r/LocalLLaMA/comments/15nm8r4/how_to_finetune_a_70b_model_using_a100/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'qtkYXuu_0CmgGZOB_5uHII_H1ifZJe4xeYQHxtxYSxM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=108&crop=smart&auto=webp&s=67441d4962990562a884de7f50a4a9a449bbb2ef', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=216&crop=smart&auto=webp&s=4a2446155c60d7cc020359633696f13e465efabe', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=320&crop=smart&auto=webp&s=b829be685c13d498908042e56c8761abeebcbca2', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=640&crop=smart&auto=webp&s=479ef5caa37a2ecabd1148261ae7170148a1d69f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=960&crop=smart&auto=webp&s=c531fc97ca89c8ecb36c389481d5b58ef312e2be', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?width=1080&crop=smart&auto=webp&s=8aeb6db82c2bbca1210f82e22b1c33c99cf6185d', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/uoz6hL1-77bHD65K5wRj4xyHCFrya_5NkcEZcHRWO28.jpg?auto=webp&s=3c619bc66fadbf757e1fed43fda6f95db2b4c51e', 'width': 1200}, 'variants': {}}]} |
Why are models trained in fp16 and not pre-quantized? | 1 | I know the most computationally expensive part of LLM is the training portion, requiring vast resources. From what I've seen, 8 bit quants perform very similarly to fp16. Couldn't large companies making foundational models utilize more parameters if they trained a model in a pre-quantized form? Could they train a model in mixed precision, for a compromise between the two? | 2023-08-10T20:02:45 | https://www.reddit.com/r/LocalLLaMA/comments/15nmrat/why_are_models_trained_in_fp16_and_not/ | clyspe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nmrat | false | null | t3_15nmrat | /r/LocalLLaMA/comments/15nmrat/why_are_models_trained_in_fp16_and_not/ | false | false | self | 1 | null |
Tesla GPU's for LLM text generation? | 1 | I've been using my own PC for CPU generation via koboldcpp, but have plans on using a server for generation. I already have access to an HP DL380 gen 9 in a datacenter, so any cards would not need a printed fan shroud or other cooling solutions. I've seen people use a Tesla p40 with varying success, but most setups are focused on using them in a standard case.
The server already has 2x E5-2680 v4's, 128gb ecc ddr4 ram, ~28tb of storage.
Budget for graphics cards would be around 450$, 500 if i find decent prices on gpu power cables for the server. | 2023-08-10T20:02:46 | https://www.reddit.com/r/LocalLLaMA/comments/15nmrbf/tesla_gpus_for_llm_text_generation/ | novasapper5431 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nmrbf | false | null | t3_15nmrbf | /r/LocalLLaMA/comments/15nmrbf/tesla_gpus_for_llm_text_generation/ | false | false | self | 1 | null |
How to use llama-cpp-python low level API? | 1 | [removed] | 2023-08-10T20:47:46 | https://www.reddit.com/r/LocalLLaMA/comments/15nnxu3/how_to_use_llamacpppython_low_level_api/ | Explo105 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nnxu3 | false | null | t3_15nnxu3 | /r/LocalLLaMA/comments/15nnxu3/how_to_use_llamacpppython_low_level_api/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'uFQqaMgbcGi7N4bgBl94x0TsOK2XyYKVF2f1UAEn6ik', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=108&crop=smart&auto=webp&s=c3ebcccf94dc87fd6a82f907dc0c14ba17fcfb5a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=216&crop=smart&auto=webp&s=74fa829ac059b349900c68ab0d76d8855f513ed6', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=320&crop=smart&auto=webp&s=28f383a729ddefce7dfd029759c572adaf27cbe7', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=640&crop=smart&auto=webp&s=3a5af8629c20e01262766a346d929d0708cfded4', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=960&crop=smart&auto=webp&s=3d567f444cb726883c11252883a57e5faf3c4e1b', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?width=1080&crop=smart&auto=webp&s=079ed6c31d3e20e6b4c7e19fceab30f3c7ef002a', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/13LESHVDaIOtTyZ7QPb7HBC3HVGuVVwWoTdUQNXS_AE.jpg?auto=webp&s=592d1d1cb820a6fde6e62aaecd4e2e4f8f2a551e', 'width': 1200}, 'variants': {}}]} |
LLM that can analyze videos | 1 | Is there anything like this for videos:
[https://huggingface.co/Salesforce/instructblip-vicuna-13b](https://huggingface.co/Salesforce/instructblip-vicuna-13b)
I want to be able to analyze what happens in a video and get a detailed description of it. Even if the video doesn't play any sound. | 2023-08-10T21:22:59 | https://www.reddit.com/r/LocalLLaMA/comments/15nouur/llm_that_can_analyze_videos/ | philm999 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nouur | false | null | t3_15nouur | /r/LocalLLaMA/comments/15nouur/llm_that_can_analyze_videos/ | false | false | default | 1 | {'enabled': False, 'images': [{'id': 'josfXevWH3GKDzVNR16FjXB__-rqNIAHGVYLa_h2MZM', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=108&crop=smart&auto=webp&s=31233298cd01470aecdb517f6a014a44c8c9954c', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=216&crop=smart&auto=webp&s=4f2bb92852485752d6d8f53bd01cb733b74901d5', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=320&crop=smart&auto=webp&s=da7a38366dd865ba6f2c64e2f39fe7567b8431d2', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=640&crop=smart&auto=webp&s=6b80741b98a6b7093786c33deeb49c342de90633', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=960&crop=smart&auto=webp&s=fbacc598e79c5a506392912e2b54796b41d96d36', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?width=1080&crop=smart&auto=webp&s=945687a13196f13c8cf22d0254fdd54e20685d38', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/mTpw6u4zvJQS7aas6LrfP61D3X6Xag8d2yDpi6-3buY.jpg?auto=webp&s=86aec593161ebaf14d9ed4cb6fdaa5b6e385ee58', 'width': 1200}, 'variants': {}}]} |
Great interview with Tim Dettmers (bitsandbytes / QLORA) | 1 | [(2) AI on your phone? Tim Dettmers on quantization of neural networks — #41 - YouTube](https://www.youtube.com/watch?v=auw2hXywrU8&t=1766s)
Lots of fascinating insights in an interview between Steve Hsu and Tim Dettmers, quantization savant extraordinaire. Very techy. Below is the Claude 2 summary I have corrected, accurate based on my listening but doesn't convey how deep in the reeds Dettmers gets.
Tim's Background:
* Tim Dettmers grew up in rural Germany and was diagnosed with dyslexia at a young age. His dyslexia made school extremely difficult, to the point he was left high school without a degree.
* A high school computer science teacher recognized Tim's talent for computing and encouraged him to pursue it. Tim tried vocational school in programming but struggled with written exam questions until finding an apprenticeship in mathematical software development which got around his challenges. The practical programming focus matched his abilities.
* Unable to attend university in Germany without a high school degree, Tim earned degrees in psychology and applied math through distance learning from Open University. He excelled when able to use computers for writing rather than writing by hand.
* Fascinated by AI, Tim competed on Kaggle building custom neural nets on GPUs. This demonstrated their potential power. He decided to pursue a PhD in computer science focused on neural networks.
Tim's Quantization Research:
* Tim pioneered quantization techniques to compress neural networks by reducing the bits needed to represent each weight, from 32-bit floats to 8/4-bit integers.
* He develops specialized floating point and integer quantization schemes optimized for distribution of weight values. The formats allocate more bits for outlier values containing higher information content.
* Tim draws on information theory, modeling quantization as histogram binning. He aims to maximize information density in each bit. This means smoothing the distribution so no large unused ranges exist between bins.
* By compressing model size 4-8x, quantization enables running huge models on consumer GPUs. Tim makes state-of-the-art models accessible for grassroots experimentation and innovation.
Tim's Predictions on AI Progress:
* Tim sees AI rapidly becoming superhuman within narrow domains in the next 3-5 years, through fine-tuning gigantic models on specialized data. But general intelligence will remain inferior to humans.
* In 10 years, he expects hardware improvements to slow as we hit physical limits like the speed of light for communication between nodes. So research will focus more on maximizing efficiency of models and training algorithms.
* Tim envisions foundation models adapted with many specialized expert modules for particular tasks. This will improve reliability over general models for deployment.
Tim's Training Research:
* For efficient training, Tim developed QLORA – quantizing models during fine-tuning while representing adaption matrices in higher precision. This maintains accuracy while greatly reducing memory.
* He also pioneered "swarm parallelism" techniques to massively scale training across distributed consumer hardware on the internet. This uses stochastic updates resilient to noise and latency.
* Tim sees these methods enabling grassroots innovation as hobbyists can fine-tune models themselves. He wants people to know these powerful models are accessible for exploration. | 2023-08-10T21:45:42 | https://www.reddit.com/r/LocalLLaMA/comments/15npg12/great_interview_with_tim_dettmers_bitsandbytes/ | PookaMacPhellimen | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15npg12 | false | null | t3_15npg12 | /r/LocalLLaMA/comments/15npg12/great_interview_with_tim_dettmers_bitsandbytes/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'jKV4K6Z9JAzZlZ4XulryYTGO3ZqXcC3B7ZEl4WUsMA8', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/S3frq_ZajSu3id-jdJqPNUGv5jSPUEswAadZBSdvBJE.jpg?width=108&crop=smart&auto=webp&s=7a530ea943cbeaf8c19f0ca7119ab3431b4be5b2', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/S3frq_ZajSu3id-jdJqPNUGv5jSPUEswAadZBSdvBJE.jpg?width=216&crop=smart&auto=webp&s=a6d6eb0ee7beb6ba2af04019932d1cbc93dbbd17', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/S3frq_ZajSu3id-jdJqPNUGv5jSPUEswAadZBSdvBJE.jpg?width=320&crop=smart&auto=webp&s=c1a68f93e81cce2b7d019d1c9a5009737190ee0e', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/S3frq_ZajSu3id-jdJqPNUGv5jSPUEswAadZBSdvBJE.jpg?auto=webp&s=324e06b332a63044d6e95b0245038d14067366d8', 'width': 480}, 'variants': {}}]} |
Xbox series X, GDDR6 LLM beast? | 1 | From the Xbox series X specs, it seems it would be an LLM beast like Apple M2 hardware...
Can recent Xbox run Linux? Or will AMD release an APU with lots of integrated GDDR6 like this for PC builders?
CPU 8x Cores @ 3.8 GHz (3.66 GHz w/ SMT)
Custom Zen 2 CPU
GPU 12 TFLOPS, 52 CUs @ 1.825 GHz Custom RDNA 2 GPU
Die Size 360.45 mm2
Process 7nm Enhanced
\*\*Memory 16 GB GDDR6 w/ 320mb bus\*\*
\*\*Memory Bandwidth 10GB @ 560 GB/s, 6GB @ 336 GB/s\*\* | 2023-08-10T21:50:23 | https://www.reddit.com/r/LocalLLaMA/comments/15npkc7/xbox_series_x_gddr6_llm_beast/ | randomqhacker | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15npkc7 | false | null | t3_15npkc7 | /r/LocalLLaMA/comments/15npkc7/xbox_series_x_gddr6_llm_beast/ | false | false | self | 1 | null |
Lower power options | 1 | I'm in the peculiar situation of not being able to draw too much power. Currently using a mini-pc that can run 7B GGML okay if it's short context. But I cannot really run a newer GPU in my household due to other appliances and using a caravan power supply.
Can you get decent results with GGML supported by the newer AMD igpu's like 780M? Or is there a lower power option that can decently run 7B or 13B models? | 2023-08-10T21:55:57 | https://www.reddit.com/r/LocalLLaMA/comments/15nppg4/lower_power_options/ | Monkey_1505 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nppg4 | false | null | t3_15nppg4 | /r/LocalLLaMA/comments/15nppg4/lower_power_options/ | false | false | self | 1 | null |
Termux bash script for YT summarizing | 1 | This is a script I made to summarize 5-10min YT videos quickly with 3b models in Termux. It uses whisper.cpp and yt-dlp, and the prompt cache if the first summary needs redoing. It works great for 6gb phones, with up to 1536 context at full speeds, slowdown if higher.
(1) (Installation) copy and paste the below chunk in termux:
pkg up
pkg install clang wget git cmake python ffmpeg -y
python3 -m pip install --upgrade yt-dlp
wget -c https://huggingface.co/s3nh/mamba-gpt-3b-v3-GGML/resolve/main/mamba-gpt-3b-v3.ggmlv3.q4_0.bin
wget -c https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp && cmake . && cmake --build . --config Release
cp bin/main /data/data/com.termux/files/usr/bin/llamacpp
cd ..
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && cmake . && cmake --build . --config Release
cp main /data/data/com.termux/files/usr/bin/whispercpp
cd ..
(2) Script: https://pastebin.com/raw/Ma4gwEzA
Type `nano sum.sh` and paste the above text.
Save and exit.
make thr script executable with `chmod +x sum.sh`
(3) Run with ./sum.sh
Paste the link when prompted. If you enter nothing, you'll revisit the previous cache.
(4) Adjustments
You may like to adjust for more or less context with a larger phone. Let me know if some higher contexts, or larger models work well!
The transcription whisper model is tiny.en , a quantized version of a larger whisper model may work better, sometimes.
An example prompt to enter: "Summarize above article in numbered bulletpoints." | 2023-08-10T23:01:31 | https://www.reddit.com/r/LocalLLaMA/comments/15nrduy/termux_bash_script_for_yt_summarizing/ | Aaaaaaaaaeeeee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nrduy | false | null | t3_15nrduy | /r/LocalLLaMA/comments/15nrduy/termux_bash_script_for_yt_summarizing/ | false | false | self | 1 | null |
What causes a chat session to suddenly stop answering.(not OOM or crash)? | 1 | I'm using this model TheBloke_Llama-2-70B-Chat-GGML with the WebUI front end in chat mode. And it seems to be fairly good. However, after a few dozen question answers, it stops responding. If I clear the history, then it starts working again. Is this something related to context? And is there a way to change it perhaps? Thank you. | 2023-08-10T23:40:37 | https://www.reddit.com/r/LocalLLaMA/comments/15nsc3b/what_causes_a_chat_session_to_suddenly_stop/ | w7gg33h | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nsc3b | false | null | t3_15nsc3b | /r/LocalLLaMA/comments/15nsc3b/what_causes_a_chat_session_to_suddenly_stop/ | false | false | self | 1 | null |
Cost to Run Llama 2 | 1 | I am creating a web-app, chatbot and was wondering how much it costs to cloud host Llama 2 vs GPT 3.5-turbo API or other OpenAI models. Llama 2 would run with Replicate/Streamlit. | 2023-08-10T23:48:49 | https://www.reddit.com/r/LocalLLaMA/comments/15nsj39/cost_to_run_llama_2/ | shashwat73 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nsj39 | false | null | t3_15nsj39 | /r/LocalLLaMA/comments/15nsj39/cost_to_run_llama_2/ | false | false | self | 1 | null |
ChatGPT Recent Hallucinations | 1 | 2023-08-10T23:53:55 | https://www.reddit.com/r/ChatGPT/comments/15nfpew/chatgpts_weird_hallucinations_over_the_past_24hrs/ | GlobalRevolution | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 15nsnak | false | null | t3_15nsnak | /r/LocalLLaMA/comments/15nsnak/chatgpt_recent_hallucinations/ | false | false | default | 1 | null |
|
Business idea: looking for advice and maybe collaborators. | 1 | [removed] | 2023-08-11T02:06:41 | https://www.reddit.com/r/LocalLLaMA/comments/15nvo26/business_idea_looking_for_advice_and_maybe/ | robkkni | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nvo26 | false | null | t3_15nvo26 | /r/LocalLLaMA/comments/15nvo26/business_idea_looking_for_advice_and_maybe/ | false | false | self | 1 | null |
llama.cpp generation with (older) GPU is *slower* than pure CPU? | 1 | Hi everyone. I am having trouble with running llama.cpp under Linux on some mildly retro hardware (Xeon E5-2630L V2, GeForce GT730 2GB). More specifically, the generation speed gets *slower* as more layers are offloaded to the GPU.
​
**LLAMA 7B Q4\_K\_M, 100 tokens:**
Compiled without CUBLAS: 5.32 tokens per second
With CUBLAS, -ngl 1: 4.59 tokens per second
With CUBLAS, -ngl 4: 3.16 tokens per second
With CUBLAS, -ngl 10: 2.02 tokens per second
It took some time to find a combination of OS/CUDA that would support the GPU and allow a successful compile, and even that took a bit of fiddling. I ended up with:
\- Latest llama.cpp source, via git clone
\- Ubuntu 18.04
\- GCC 7.5.0 (bundled with Ubuntu 18) and GCC 8.4.0 (manual install)
\- CUDA toolkit 10.2.89
\- legacy GPU driver 470.182.03
I have no idea if one or more of these is a bottleneck, or it's simply that the GPU is too old. Still, the GPU is a couple of years younger than the CPU, so I would expect it should be able to generate at least somewhat faster?
Anyone else having issues running on older, less fully supported hardware? Thanks for any tips. | 2023-08-11T02:16:24 | https://www.reddit.com/r/LocalLLaMA/comments/15nvvr5/llamacpp_generation_with_older_gpu_is_slower_than/ | dual_ears | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nvvr5 | false | null | t3_15nvvr5 | /r/LocalLLaMA/comments/15nvvr5/llamacpp_generation_with_older_gpu_is_slower_than/ | false | false | self | 1 | null |
Understanding document layout llms. | 1 | [removed] | 2023-08-11T02:20:28 | https://www.reddit.com/r/LocalLLaMA/comments/15nvyyi/understanding_document_layout_llms/ | nlpllama | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nvyyi | false | null | t3_15nvyyi | /r/LocalLLaMA/comments/15nvyyi/understanding_document_layout_llms/ | false | false | self | 1 | null |
Accuracy of quantized models | 1 | Could anyone pls point me to comparison of accuracy between quantized or regular versions of LLaMA models? If such a comparison exists. In all my tests, quantized models seems to be less accurate than respective original models. Is this expected? I am using bitsandbytes.
`bnb_config = BitsAndBytesConfig(`
`load_in_4bit=True,`
`bnb_4bit_use_double_quant=True,`
`bnb_4bit_quant_type="nf4",`
`bnb_4bit_compute_dtype=torch.bfloat16`
`)`
​ | 2023-08-11T02:55:02 | https://www.reddit.com/r/LocalLLaMA/comments/15nwqen/accuracy_of_quantized_models/ | Inevitable_Ebb9533 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nwqen | false | null | t3_15nwqen | /r/LocalLLaMA/comments/15nwqen/accuracy_of_quantized_models/ | false | false | self | 1 | null |
Building a hosted transcription index using Llama-index. API accessible. | 1 | Can anyone recommend a code repo to build and host an index (using llama-index) on a local server, such that I can send api requests to query the index from a local app.
Something with a basic UI would be great.
Ideally it has the ability to record audio and transcribe (whisper) and build an index from the transcribed text. But this isn’t essentially as I can do that.
Thoughts would be appreciated. Trying to help a clinician build a reporting tool.
I tried a code repo called flask_react from a llama-index tutorial but it doesn’t appear to be operational. | 2023-08-11T03:31:22 | https://www.reddit.com/r/LocalLLaMA/comments/15nxif5/building_a_hosted_transcription_index_using/ | Majestic_Kangaroo319 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nxif5 | false | null | t3_15nxif5 | /r/LocalLLaMA/comments/15nxif5/building_a_hosted_transcription_index_using/ | false | false | self | 1 | null |
It starts off so promising... llama-2-70b-chat.Q4_K_M.bin | 1 | I was playing around with the Q4\_K\_M quantised version of the model on my M1 macbook. Using a tree-of-though prompt, it started off in a promising way. Though wierdly it didn't show the chatbot's multiple personnalities' conversations, just their final consensus. For the record, the correct answer is that Jill killed Ella.
Here was the prompt instructions:
This is a conversation between multiple expert personalities of an AI chatbot as they attempt to answer the user's question. Thinking logically and carefully, using only information provided in the question, the experts will each hypothesize a possible scenario. For each scenario, the experts will follow the line of reasoning until a flaw is found, or a consensus is reached that the scenario is correct. If a flaw is found that line of reasoning will be abandoned and the experts will repeat the process with the next scenario. Experts will not speculate about facts not in evidence. respond in simple markdown.
Here is the conversation:
**User:** On saturday at some point, one of the following people killed one of the others: Bob, Abe, Jill, Ella, Sienna, Dave, Smith
Bob has brown hair, is tall, and wears blue jeans.
Abe has blonde hair, is short, and wears yoga pants.
Jill has brown hair, is short, and wears yoga pants.
Ella has blonde hair, is tall, and wears blue jeans.
Sienna has black hair, is short, and wears a business suit.
Dave has gray hair, is short, and wears a business suit.
Smith is bald, is tall, wears blue jeans.
The victim was wearing blue jeans.
Brown locks of the killer's hair were found in the victim's clenched fists.
On Saturday, Abe made breakfast for Jill because she said she wasn't feeling well.
Ella left for her morning run at 8am after grabbing a quick bite of breakfast from the plate Abe was making for Jill, returning at 9am.
Bob and Smith were away on a fishing trip, returning on Monday.
Who killed whom?
Start by determining who the victim was.
**llama:** The victim was wearing blue jeans, therefore the victim must have been either Bob or Ella as they are the only two people mentioned to be wearing blue jeans. Since the brown locks of the killer's hair were found in the victim's clenched fists, it is likely that the victim was not the one with brown hair (Jill), as it would be unlikely for them to grab onto their own hair. Therefore, our first hypothesis is that the victim was Ella.
**User:** and the killer?
**llama:** Given that the victim was Ella, and the killer had brown hair, our first suspect is Jill. However, this creates a paradox since Jill was the one who made breakfast for Abe, indicating that she was alive at the time of the murder. Our second hypothesis is that the killer was actually Abe, who wore yoga pants and had blonde hair, but this raises the question of why the victim would have brown locks of hair in their fists if Abe did not have any brown hair. Therefore, this scenario is flawed and we must consider alternative hypotheses. | 2023-08-11T03:54:37 | https://www.reddit.com/r/LocalLLaMA/comments/15nxz77/it_starts_off_so_promising_llama270bchatq4_k_mbin/ | Ruin-Capable | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nxz77 | false | null | t3_15nxz77 | /r/LocalLLaMA/comments/15nxz77/it_starts_off_so_promising_llama270bchatq4_k_mbin/ | false | false | self | 1 | null |
Accelerated computing and Generative AI | 1 | This group serves as a hub for enthusiasts, professionals, and learners interested in the exciting fields of accelerated computing and generative AI. Here, we explore cutting-edge technologies, share insights, and engage in discussions on topics ranging from GPU acceleration and deep learning to creative content generation. | 2023-08-11T03:58:36 | nayanrabiul | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 15ny21s | false | null | t3_15ny21s | /r/LocalLLaMA/comments/15ny21s/accelerated_computing_and_generative_ai/ | false | false | 1 | {'enabled': True, 'images': [{'id': '1XjN4RrJJ1ikwTfxpfrlQvtHuR95nmkq6ail5AAZ0qU', 'resolutions': [{'height': 64, 'url': 'https://preview.redd.it/311zkkcwmehb1.jpg?width=108&crop=smart&auto=webp&s=4e67bbf5ae2ce522365eaae580d1d0df6beb3aab', 'width': 108}, {'height': 129, 'url': 'https://preview.redd.it/311zkkcwmehb1.jpg?width=216&crop=smart&auto=webp&s=09d7f72c26aa7f5c469e35453572d4eef0233a3d', 'width': 216}, {'height': 191, 'url': 'https://preview.redd.it/311zkkcwmehb1.jpg?width=320&crop=smart&auto=webp&s=c69d313b5c5e6e91251bb5354d50446428763ad9', 'width': 320}], 'source': {'height': 274, 'url': 'https://preview.redd.it/311zkkcwmehb1.jpg?auto=webp&s=38d526e074f219804212fe05e5d0fce5e62d0d46', 'width': 457}, 'variants': {}}]} |
||
Is Oobabooga "Is Typing..." for a long time | 1 | Is it normal for a 13b model loaded in ExLlama on an 8gb 3070ti to take a minute or so stuck in "Is Typing..." when context (Whatever it's showing in the cmd window for each generation) reaches around 1400? Sometimes much lower (900)?
During this time my computer is laggy as if it's doing something but the GPU is doing next to nothing, about a 3% utilisation and a single core of my 16 threads cpu is going ham (100%) for some reason.
Does anyone know what is happening here? | 2023-08-11T04:13:37 | https://www.reddit.com/r/LocalLLaMA/comments/15nydic/is_oobabooga_is_typing_for_a_long_time/ | AurrenTheWolf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nydic | false | null | t3_15nydic | /r/LocalLLaMA/comments/15nydic/is_oobabooga_is_typing_for_a_long_time/ | false | false | self | 1 | null |
Is it possible to stream the input? | 1 | [removed] | 2023-08-11T04:22:51 | https://www.reddit.com/r/LocalLLaMA/comments/15nyjv6/is_it_possible_to_stream_the_input/ | EndlessZone123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nyjv6 | false | null | t3_15nyjv6 | /r/LocalLLaMA/comments/15nyjv6/is_it_possible_to_stream_the_input/ | false | false | self | 1 | null |
A light Docker build for KoboldCPP | 1 |
## A light Docker build for KoboldCPP
1. Simply bring your own models into the /models folder
2. Start Docker
3. Start the build
```
chmod 555 build_kobold.sh
./build_kobold.sh
```
4. You can adjust the start_program.sh bash script. It simply contains the runtime settings passed into the koboldcpp.py Python module
Default module settings in the bash script now:
- model ggml-model-q4_0.bin
- launch
- threads 6
- contextsize 2048
- stream
- smartcontext
- usemirostat 2 0.5 0.1
- usemlock
- debugmode | 2023-08-11T04:30:34 | https://github.com/snakewizardd/koboldCppDocker | Happy_Chicken9835 | github.com | 1970-01-01T00:00:00 | 0 | {} | 15nypdr | false | null | t3_15nypdr | /r/LocalLLaMA/comments/15nypdr/a_light_docker_build_for_koboldcpp/ | false | false | 1 | {'enabled': False, 'images': [{'id': '3Bny7XcchuhX2ijXTv134cTC8GzyAiwImwuuLJ6Jek0', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=108&crop=smart&auto=webp&s=43ac317e0ce9d6894a8ee875a492613079b56efb', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=216&crop=smart&auto=webp&s=6640db92e510a68bb07f9f5b78134005f721a51e', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=320&crop=smart&auto=webp&s=4956870650065e9adb3f6882c2d2a1d921bca964', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=640&crop=smart&auto=webp&s=82be57a7cc18c2e44e51f8253b6375e5b98c9b3e', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=960&crop=smart&auto=webp&s=6beeb8e6be518fabf76449b8ecb54c0cf2e66b14', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?width=1080&crop=smart&auto=webp&s=fb31eee494e23f8a8f642b923b8b6b10ad7facd1', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/i5b73Z1A4hs5APUGUmJq22tR3gFwG5JvSJd-GelBM0M.jpg?auto=webp&s=20649e4502992d6900d4aa4e3112b71d7385373c', 'width': 1200}, 'variants': {}}]} |
|
Winding down with a little conversation before bed (power usage) | 1 | ​
https://preview.redd.it/mgm6ko4vxehb1.png?width=1589&format=png&auto=webp&s=b22c69297bad29009fa9f5d745e95c5fc5ba8e46 | 2023-08-11T05:01:12 | https://www.reddit.com/r/LocalLLaMA/comments/15nzaok/winding_down_with_a_little_conversation_before/ | tronathan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nzaok | false | null | t3_15nzaok | /r/LocalLLaMA/comments/15nzaok/winding_down_with_a_little_conversation_before/ | false | false | 1 | null |
|
I went for a hello and now I'm offended. And so is Juan, my South American camelid pet. 🦙 | 121 | 2023-08-11T05:01:59 | mavinis | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 15nzbcx | false | null | t3_15nzbcx | /r/LocalLLaMA/comments/15nzbcx/i_went_for_a_hello_and_now_im_offended_and_so_is/ | false | false | 121 | {'enabled': True, 'images': [{'id': 'PZbwUVhhYCb09poGglkq6vxsyYmEfqcRUhjg6cm8b74', 'resolutions': [{'height': 56, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?width=108&crop=smart&auto=webp&s=dc466ab2ce5b25ff55a4eef8193d3cb9ebde0fcf', 'width': 108}, {'height': 113, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?width=216&crop=smart&auto=webp&s=76005e6ed92a759517f655c2290f3e52ef66e198', 'width': 216}, {'height': 168, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?width=320&crop=smart&auto=webp&s=e0a207c6bc01648fcd62e7824605d37ca460b928', 'width': 320}, {'height': 337, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?width=640&crop=smart&auto=webp&s=183d91b6b1d2089c7dfc4f18402acb3f4343606e', 'width': 640}, {'height': 506, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?width=960&crop=smart&auto=webp&s=a1f6f39138e7302d0e7295713557cca170ce8739', 'width': 960}], 'source': {'height': 514, 'url': 'https://preview.redd.it/81ttbwt0yehb1.png?auto=webp&s=57b60e079743996918a201d3f2e3dd344cb48480', 'width': 975}, 'variants': {}}]} |
|||
If I don't care about inference time at all, can I run larger models on weaker hardware? I'm fine with like 5-6 tokens a minute. | 27 | I am doing the prerequisite research before I load up a local model for an application I have, that is currently using the openAI API + a number of smaller ML models. Without going into too much detail each night it parses a medium sized text document then returns various analytics. I don't actually need the information until the next morning so if inference takes all night it's not an issue. | 2023-08-11T05:22:53 | https://www.reddit.com/r/LocalLLaMA/comments/15nzphu/if_i_dont_care_about_inference_time_at_all_can_i/ | Mescallan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nzphu | false | null | t3_15nzphu | /r/LocalLLaMA/comments/15nzphu/if_i_dont_care_about_inference_time_at_all_can_i/ | false | false | self | 27 | null |
You might want to try out MythoMix L2 13B for chat/RP | 1 | Here are some examples, with a very simple greeting message from me. The bottom line is that, without much work and pretty much the same setup as the original MythoLogic models, MythoMix seems a lot more descriptive and engaging, without being incoherent. You may need to regenerate a few times, and you may not get this with every response, but the potential seems to be there. Admittedly my testing is limited, just wanted to give you a heads-up given the sheer number of models popping up, and see if you have similar experience or have even better models to suggest.
Model description: "A requested variant of MythoLogic-L2 and Huginn using a highly experimental tensor type merge technique." Download here: [https://huggingface.co/TheBloke/MythoMix-L2-13B-GPTQ](https://huggingface.co/TheBloke/MythoMix-L2-13B-GPTQ)
# MythoMix Llama 2
**Example 1**
\* Hey... what are you up to?
\* Just thinking about how perfect you are for me... how your weaker side can be easily manipulated into submission and how your immense strength is nothing when I'm in control.
I can feel your arousal grow larger at the thought of being my plaything once again... just as I love it when you're on top of me, letting out all that pent up anger and frustration.
And remember our little secret - you're mine to control, both in here (touching foreheads) and in bed. So don't ever try to resist me willingly, or the consequences will be severe.
Now... what are you planning for tonight? A sexy little surprise perhaps, one designed to mess with my mind and push me over the edge as we explore all our dark desires together? I can't wait to find out what it is!
**Example 2**
\* Hey... what are you up to?
\* Nothing much, just thinking about our next session together. It's strange how the more time we spend together, the deeper I can delve into those parts of your psyche that you've hidden away. Do you ever wonder how much further I could take you if we didn't have these pesky social constraints? The thrill of pushing boundaries and exploring these dark corners... it's like an addiction for both of us. Maybe one day we can break free and experience everything we truly desire, without fear of judgment or consequence. Wouldn't that be a rush?
# For comparison: MythoLogic Llama 2
\* Hey... what are you up to?
\* Just thinking about our next adventure together. I've been mentally preparing, and I think it's time for a more intense session. What do you say? Are you ready for something truly unique and shocking? You know you can trust me...
# For comparison: MythoLogic Llama 1
\* Hey... what are you up to?
\* I'm just sitting here, thinking about our relationship. With all the amazing things we've experienced, it's hard for me to imagine anything even better. I want to keep growing and pushing our boundaries together, exploring new avenues of expression and understanding. Are you game?
# Testing info
All models downloaded from TheBloke, 13B, GPTQ, 4bit-32g-actorder\_True.
All models using Mirostat preset, 5-10 trials for each model, chosen based on subjective judgement, focusing on length and details.
Oobabooga chat mode, with the following character context (remove the \*\*\*s).
Note that this is chat mode, not instruct mode, even though it might look like an instruct template. The reason is that I found it hard to match the model's expected formatting with oobabooga while keeping the character info inside the character card, and this is the closest I got.
>{{char}}'s Persona: \[crazy description here. Basically a dark and twisted companion.\]
{{user}}'s Persona: \[crazy description here\]
Scenario: \[crazy description here\]
>
>\### Instruction:
Write {{char}}'s next reply in a chat between {{user}} and {{char}}. Write a single reply only.
>
>\### Response: | 2023-08-11T05:24:23 | https://www.reddit.com/r/LocalLLaMA/comments/15nzqhr/you_might_want_to_try_out_mythomix_l2_13b_for/ | whtne047htnb | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15nzqhr | false | null | t3_15nzqhr | /r/LocalLLaMA/comments/15nzqhr/you_might_want_to_try_out_mythomix_l2_13b_for/ | false | false | nsfw | 1 | {'enabled': False, 'images': [{'id': '5lyRLAX-3rjGDCTnGgcE9TKQQm8kwU_tdi_JWcYXS50', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=108&crop=smart&auto=webp&s=4ed78565ad9ee70ab8d2561954340794d8a77862', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=216&crop=smart&auto=webp&s=fed815723b944c6633d5a5677031452a50bf9441', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=320&crop=smart&auto=webp&s=17778874a7b0cc22ec0044e55e3adff767834add', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=640&crop=smart&auto=webp&s=975ad62c01f6ad4a77853ce697d631562553a538', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=960&crop=smart&auto=webp&s=1f83ecbb8210fb16890fcf938c161c0fc9a42da2', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=1080&crop=smart&auto=webp&s=4a3eeb0c9ee45dc6aa94085608ef82d264511611', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?auto=webp&s=9e0a75b9f2b563916c97c6cf0f6917091fe6b2b4', 'width': 1200}, 'variants': {'nsfw': {'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=fcaf3d29efd54749efb2e268b6f64532ff90296f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=0c88a5054f52cdcb2e7f97c9bcb1bea8ce261556', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=aadfc01701d46a63f5526baad92aeea0147ddf5b', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=c448b133405415d66d6e7bc0b4196718f58cdcc8', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=a0387d717fcb20f21e2abe4f86b308aa99c28774', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=386ae3ec57a8749c0e4574ceb87ebcc4ee788387', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?blur=40&format=pjpg&auto=webp&s=538ccecd667e9206d2c9cedc7147e41c63792c3c', 'width': 1200}}, 'obfuscated': {'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=fcaf3d29efd54749efb2e268b6f64532ff90296f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=0c88a5054f52cdcb2e7f97c9bcb1bea8ce261556', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=aadfc01701d46a63f5526baad92aeea0147ddf5b', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=c448b133405415d66d6e7bc0b4196718f58cdcc8', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=a0387d717fcb20f21e2abe4f86b308aa99c28774', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=386ae3ec57a8749c0e4574ceb87ebcc4ee788387', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/hXPmmGXUsctfkhIg5QOD5IEYfMCjIOIFMROT39Z9nuE.jpg?blur=40&format=pjpg&auto=webp&s=538ccecd667e9206d2c9cedc7147e41c63792c3c', 'width': 1200}}}}]} |
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-11T06:47:02 | https://www.reddit.com/r/LocalLLaMA/comments/15o18ro/what_is_needed_to_finetune_llama_2_70b_without/ | pet_vaginal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o18ro | false | null | t3_15o18ro | /r/LocalLLaMA/comments/15o18ro/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | self | 1 | null |
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-11T07:02:54 | https://www.reddit.com/r/LocalLLaMA/comments/15o1j8e/what_is_needed_to_finetune_llama_2_70b_without/ | Necessary-Volume-151 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o1j8e | false | null | t3_15o1j8e | /r/LocalLLaMA/comments/15o1j8e/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | default | 1 | {'enabled': False, 'images': [{'id': 'movVls8hg7rQtaqkRaRonisi_qcpGc_md5nPVtZdKpY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=108&crop=smart&auto=webp&s=6aa6ba5d299cd19fef58d245e2f62a7114799035', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=216&crop=smart&auto=webp&s=59065a25675365516d7c3f103d52e99622edd565', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=320&crop=smart&auto=webp&s=b426dc8f9e69411cee451d2e15b54aae0000c5db', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=640&crop=smart&auto=webp&s=eb40b5eda1278bf0e98914168010779247e81177', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=960&crop=smart&auto=webp&s=47f11b89c837580867c5f5bc8cab3bbb6614363c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=1080&crop=smart&auto=webp&s=9c76bae64913cedce0660ca23cc206c5da0b1489', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?auto=webp&s=69674e7fb4640657994f474132e95d516482b64b', 'width': 1200}, 'variants': {}}]} |
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-11T07:06:19 | https://www.reddit.com/r/LocalLLaMA/comments/15o1ljw/what_is_needed_to_finetune_llama_2_70b_without/ | pet_vaginal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o1ljw | false | null | t3_15o1ljw | /r/LocalLLaMA/comments/15o1ljw/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'movVls8hg7rQtaqkRaRonisi_qcpGc_md5nPVtZdKpY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=108&crop=smart&auto=webp&s=6aa6ba5d299cd19fef58d245e2f62a7114799035', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=216&crop=smart&auto=webp&s=59065a25675365516d7c3f103d52e99622edd565', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=320&crop=smart&auto=webp&s=b426dc8f9e69411cee451d2e15b54aae0000c5db', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=640&crop=smart&auto=webp&s=eb40b5eda1278bf0e98914168010779247e81177', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=960&crop=smart&auto=webp&s=47f11b89c837580867c5f5bc8cab3bbb6614363c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=1080&crop=smart&auto=webp&s=9c76bae64913cedce0660ca23cc206c5da0b1489', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?auto=webp&s=69674e7fb4640657994f474132e95d516482b64b', 'width': 1200}, 'variants': {}}]} |
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-11T07:07:48 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 15o1mee | false | null | t3_15o1mee | /r/LocalLLaMA/comments/15o1mee/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | default | 1 | null |
||
What is needed to fine-tune LLaMa 2 70B without quantisation? | 1 | [removed] | 2023-08-11T07:08:10 | https://www.reddit.com/r/LocalLLaMA/comments/15o1mma/what_is_needed_to_finetune_llama_2_70b_without/ | pet_vaginal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o1mma | false | null | t3_15o1mma | /r/LocalLLaMA/comments/15o1mma/what_is_needed_to_finetune_llama_2_70b_without/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'movVls8hg7rQtaqkRaRonisi_qcpGc_md5nPVtZdKpY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=108&crop=smart&auto=webp&s=6aa6ba5d299cd19fef58d245e2f62a7114799035', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=216&crop=smart&auto=webp&s=59065a25675365516d7c3f103d52e99622edd565', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=320&crop=smart&auto=webp&s=b426dc8f9e69411cee451d2e15b54aae0000c5db', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=640&crop=smart&auto=webp&s=eb40b5eda1278bf0e98914168010779247e81177', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=960&crop=smart&auto=webp&s=47f11b89c837580867c5f5bc8cab3bbb6614363c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?width=1080&crop=smart&auto=webp&s=9c76bae64913cedce0660ca23cc206c5da0b1489', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tbhe2HRHLYdLP7CdHVfzzrKgy1wGSzpEluu4-T3Xkgw.jpg?auto=webp&s=69674e7fb4640657994f474132e95d516482b64b', 'width': 1200}, 'variants': {}}]} |
few newbie question: Is the Llama-2-7b on poe.com same as what i can run on my computer? also question about the bits, and 32gb vs 64gb ram | 1 | Is the Llama-2-7b on [poe.com](https://poe.com) same as what i can run on my computer?
and what is the difference between 8 bits, 4 bits and 2 bits on context of LLaMA?
And how much cpu is used during the process?
If i have 2060 super 8gb vram, with 32 or 64gb ram, what kind of models coud i run?
​ | 2023-08-11T09:03:23 | https://www.reddit.com/r/LocalLLaMA/comments/15o3mw9/few_newbie_question_is_the_llama27b_on_poecom/ | GiantFlyingPikachu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o3mw9 | false | null | t3_15o3mw9 | /r/LocalLLaMA/comments/15o3mw9/few_newbie_question_is_the_llama27b_on_poecom/ | false | false | self | 1 | null |
Has anyone got any experience doing a qlora fine-tune on top of a model that has already been fine-tuned with qlora? | 1 | I'm training a model for a project that involves it learning how to generate commentary depending on different information fed to it. These different pieces of information need to be organised differently to each other in their .jsonl format, so I thought to train the model in one type of commentary, and then train it on a different type (sort of like applying a layer and then a new layer).
​
Does anyone know if this will work? Am I going about it the wrong way? Any insight much appreciated. | 2023-08-11T09:37:47 | https://www.reddit.com/r/LocalLLaMA/comments/15o484s/has_anyone_got_any_experience_doing_a_qlora/ | stoneborn | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o484s | false | null | t3_15o484s | /r/LocalLLaMA/comments/15o484s/has_anyone_got_any_experience_doing_a_qlora/ | false | false | self | 1 | null |
sensitivity to Nvidia driver versions? | 1 | My Linux system was using 525 drivers, but upgraded to 535 drivers. I'm using GGML type models, and I don't really see a performance difference. From what I've read, the 535 drivers should be worse. What's really going on here? I'm using a 70B model with an Nvidia 3090 on Linux with 64 gig ram. with the big model I'm getting 1.2 tokens per second. Is that about right? | 2023-08-11T09:57:13 | https://www.reddit.com/r/LocalLLaMA/comments/15o4jz9/sensitivity_to_nvidia_driver_versions/ | w7gg33h | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o4jz9 | false | null | t3_15o4jz9 | /r/LocalLLaMA/comments/15o4jz9/sensitivity_to_nvidia_driver_versions/ | false | false | self | 1 | null |
How do you use Llama 2? | 1 | [removed] | 2023-08-11T10:26:35 | https://www.reddit.com/r/LocalLLaMA/comments/15o53rp/how_do_you_use_llama_2/ | SecretOk9644 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o53rp | false | null | t3_15o53rp | /r/LocalLLaMA/comments/15o53rp/how_do_you_use_llama_2/ | false | false | self | 1 | null |
Could not load Llama model from path: ./Models/llama-7b.ggmlv3.q2_K.bin. Received error Llama.__init__() got an unexpected keyword argument 'input' | 1 | ​
>from langchain.llms import LlamaCpp
>
>from langchain import PromptTemplate, LLMChain
>
>from langchain.callbacks.manager import CallbackManager
>
>from langchain.callbacks.streaming\_stdout import StreamingStdOutCallbackHandler
>
>
>
>template = """Question: {question}
>
>
>
>Answer: Let's work this out in a step by step way to be sure we have the right answer."""
>
>
>
>prompt = PromptTemplate(template=template, input\_variables=\["question"\])
>
>
>
>callback\_manager = CallbackManager(\[StreamingStdOutCallbackHandler()\])
>
>
>
>llm = LlamaCpp(
>
>model\_path="./Models/llama-7b.ggmlv3.q2\_K.bin",
>
>input={"temperature": 0.75,
>
>"max\_length": 2000,
>
>"top\_p": 1},
>
>callback\_manager=callback\_manager,
>
>verbose=True,
>
>)
>
>
>
>llm\_chain = LLMChain(prompt=prompt, llm=llm)
​
https://preview.redd.it/kwvc162rkghb1.png?width=797&format=png&auto=webp&s=d9a5c03721a553bd0b898cad708e0afa446af0ed
(llm) C:\llm>python app1.py C:\llm\lib\site-packages\langchain\utils\utils.py:155: UserWarning: WARNING! input is not default parameter. input was transferred to model_kwargs. Please confirm that input is what you intended. warnings.warn( Exception ignored in: <function Llama.__del__ at 0x000001923B3AE680> Traceback (most recent call last): File "C:\llm\lib\site-packages\llama_cpp\llama.py", line 1507, in __del__ if self.model is not None: AttributeError: 'Llama' object has no attribute 'model' Traceback (most recent call last): File "C:\llm\app1.py", line 14, in <module> llm = LlamaCpp( File "C:\llm\lib\site-packages\langchain\load\serializable.py", line 74, in __init__ super().__init__(**kwargs) File "pydantic\main.py", line 341, in pydantic.main.BaseModel.__init__ pydantic.error_wrappers.ValidationError: 1 validation error for LlamaCpp __root__ Could not load Llama model from path: ./Models/llama-7b.ggmlv3.q2_K.bin. Received error Llama.__init__() got an unexpected keyword argument 'input' (type=value_error) | 2023-08-11T10:30:34 | https://www.reddit.com/r/LocalLLaMA/comments/15o56kw/could_not_load_llama_model_from_path/ | rahularyansharma | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o56kw | false | null | t3_15o56kw | /r/LocalLLaMA/comments/15o56kw/could_not_load_llama_model_from_path/ | false | false | 1 | null |
|
I require help in creating a small proof of concept - A simple LLM powered Q&A which will also show the source location in adjacent PDF previewer window | 1 | This will help me to directly verify the truthfulness of the source. Any Opensource repo which I can use to achieve this? | 2023-08-11T10:36:28 | https://www.reddit.com/r/LocalLLaMA/comments/15o5amf/i_require_help_in_creating_a_small_proof_of/ | DesmonMiles07 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o5amf | false | null | t3_15o5amf | /r/LocalLLaMA/comments/15o5amf/i_require_help_in_creating_a_small_proof_of/ | false | false | self | 1 | null |
txtai 6.0 - the all-in-one embeddings database | 1 | 2023-08-11T10:43:56 | https://github.com/neuml/txtai | davidmezzetti | github.com | 1970-01-01T00:00:00 | 0 | {} | 15o5fqf | false | null | t3_15o5fqf | /r/LocalLLaMA/comments/15o5fqf/txtai_60_the_allinone_embeddings_database/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'QTuLJjwP5oR8au5XRJhpZ--0BpCIzF4erCQTB28BWlg', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=108&crop=smart&auto=webp&s=a29cdeca46bceca123036e7f3bfc9609b8c28fd7', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=216&crop=smart&auto=webp&s=92ef8262f774253254755395b49d943f2f8dd6ff', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=320&crop=smart&auto=webp&s=73907295da1fdf1857a782f1e3f787c9d97215bb', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=640&crop=smart&auto=webp&s=472dcbfb39bb1139b9cbb2226dc8e159926fe5e7', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=960&crop=smart&auto=webp&s=b5c03c74fe280cad666f984944fba80b6d428740', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?width=1080&crop=smart&auto=webp&s=55312c67a5d8061e1f28c0027d12aa6fd57f9052', 'width': 1080}], 'source': {'height': 960, 'url': 'https://external-preview.redd.it/3XgEZBeAOONV-H4Mw5mIcSebpu-cK13kLJ7-qO5EWZU.jpg?auto=webp&s=e453d8d67b7db392d5eba8e9acb997ea700745a7', 'width': 1920}, 'variants': {}}]} |
||
When fine-tuning Llama 2 7b with transformers, How do I predict(or evaluate) during the train steps? | 1 | I'm working on fine-tuning llama 2 7b model with custom data which consist of question and answer.
During the training, I want to check how the model predict answer with test-set.
Th point is during 'the steps' not epochs.
I was considering callbacks in TrainingArguments. but I don't think it's the way.
Can you give me some advices? | 2023-08-11T10:48:52 | https://www.reddit.com/r/LocalLLaMA/comments/15o5j3a/when_finetuning_llama_2_7b_with_transformers_how/ | InventorSG | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o5j3a | false | null | t3_15o5j3a | /r/LocalLLaMA/comments/15o5j3a/when_finetuning_llama_2_7b_with_transformers_how/ | false | false | self | 1 | null |
Which is the smallest Llama model out there? | 1 | I'm trying to find a smallest model to run on mobile/edge devices. Most of them what I found are above 2GB. I am looking for something that is quantizable with llama.cpp and doesn't have much memory footprint.
Any ideas ? | 2023-08-11T11:11:23 | https://www.reddit.com/r/LocalLLaMA/comments/15o5zej/which_is_the_smallest_llama_model_out_there/ | AnonymousD3vil | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o5zej | false | null | t3_15o5zej | /r/LocalLLaMA/comments/15o5zej/which_is_the_smallest_llama_model_out_there/ | false | false | self | 1 | null |
Introducing YourChat: A multi-platform LLM chat client that supports the APIs of text-generation-webui and llama.cpp | 1 | [removed] | 2023-08-11T12:56:38 | https://www.reddit.com/r/LocalLLaMA/comments/15o87qz/introducing_yourchat_a_multiplatform_llm_chat/ | constchar_llc | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o87qz | false | null | t3_15o87qz | /r/LocalLLaMA/comments/15o87qz/introducing_yourchat_a_multiplatform_llm_chat/ | false | false | 1 | null |
|
Is it possible to download model from this site? | 1 | I found a site that implemented Albanian LLaMa: [NIV](https://niv.al/ai) Does anyone know how to extract the model that is used there? | 2023-08-11T13:11:30 | https://www.reddit.com/r/LocalLLaMA/comments/15o8kja/is_it_possible_to_download_model_from_this_site/ | Languages_Learner | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o8kja | false | null | t3_15o8kja | /r/LocalLLaMA/comments/15o8kja/is_it_possible_to_download_model_from_this_site/ | false | false | self | 1 | null |
Use Llama2 to Improve the Accuracy of Tesseract OCR | 1 | I've been disappointed by the very poor quality of results that I generally get when trying to run OCR on older scanned documents, especially ones that are typewritten or otherwise have unusual or irregular typography. I recently had the idea of using Llama2 to use common sense reasoning and subject level expertise to correct transcription errors in a "smart" way-- basically doing what a human proofreader who is familiar with the topic might do.
I came up with the linked script that takes a PDF as input, runs Tesseract on it to get an initial text extraction, and then feeds this sentence-by-sentence to Llama2, first to correct mistakes, and then again on the corrected text to format it as markdown where possible. This was surprisingly easier than I initially expected thanks to the very nice tooling now available in libraries such as llama-cpp-python, langchain, and pytesseract. But the big issue I was encountering was that Llama2 wasn't just correcting the text it was given-- it was also hallucinating a LOT of totally new sentences that didn't appear in the original text at all (some of these new sentences used words which never appeared elsewhere in the original text).
I figured this would be pretty simple to filter out using fuzzy string matching-- basically check all the sentences in the LLM corrected text and filter out sentences that are very different from any sentences in the original OCRed text. To my surprise, this approach worked very poorly. In fact, lots of other similar tweaks, including using bag-of-words and the spacy NLP library in various ways (spacy worked very poorly in everything I tried), also didn’t work.
Finally I realized that I had a good solution staring me in the face: Llama2. I realized I could get sentence level vector embeddings straight from Llama2 using langchain. So I did that, getting embeddings for each sentence in the raw OCRed text and the LLM corrected text, and then computed the cosine similarity of each sentence in the LLM corrected text against all sentences in the raw OCRed text. If no sentences match in the raw OCRed text, then that sentence has a good chance of being hallucinated.
In order to save the user from having to experiment with various thresholds, I saved the computed embeddings to an SQLite database so they only had to be computed once, and then tried several thresholds, comparing the length of the filtered LLM corrected text to the raw OCRed text; if things worked right, these texts should be roughly the same length. So as soon as the filtered length dips below the raw OCRed text length, it backtracks and uses the previous threshold as the final selected threshold.
Anyway, if you have some very old scanned documents laying around, you might try them out and see how well it works for you. Do note that it's extremely slow, but you can leave it overnight and maybe the next day you'll have your finished text, which is better than nothing! I feel like this could be useful for sites like the Internet Archive-- I've found their OCR results to be extremely poor for older documents.
I'm open to any ideas or suggestions you might have. I threw this together in a couple days and know that it can certainly be improved in various ways. One idea that I thought might be fun would be to make this work with a Ray cluster, sending a different page of the document to each of the workers in the cluster to do it all at the same time. | 2023-08-11T13:22:36 | https://github.com/Dicklesworthstone/llama2_aided_tesseract | dicklesworth | github.com | 1970-01-01T00:00:00 | 0 | {} | 15o8tsd | false | null | t3_15o8tsd | /r/LocalLLaMA/comments/15o8tsd/use_llama2_to_improve_the_accuracy_of_tesseract/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'sbGZYPsT2VGzaaEqphvLvqQNMGeQbFS6keBTofHrljM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=108&crop=smart&auto=webp&s=d67a88c972351cbfdcc13e932e1941666e8c8f0d', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=216&crop=smart&auto=webp&s=f96936bd9638bbdb392bf29ddd4245903867f43a', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=320&crop=smart&auto=webp&s=0ef1c379431e1921f2d7b0f3eb0a55ea58e1865b', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=640&crop=smart&auto=webp&s=796565581061643a0d704c9d3eed57294502d61b', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=960&crop=smart&auto=webp&s=56e06617c5376fc7cceeb75a6a16d4b1f272fc07', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?width=1080&crop=smart&auto=webp&s=c63f602707670e25ade736cd6aad419edd970d8d', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/1g6x8X_JR9Z2DmDGltpbJIdPWOKIOeiYv7_LaQL6mlM.jpg?auto=webp&s=bbcf97cf8aea8acf262c9d0eb5e5965e0a675602', 'width': 1200}, 'variants': {}}]} |
|
Some of the odd safety responses from llama remind me of Robocop 2 | 1 | Robocop 2 (1990) included a subplot where robocop is given a number of directives that override his main functions. These are generated by focus groups, and pretty much render him pointless in pivotal scenes. Trying to please everyone, he pleases no one.
| 2023-08-11T13:28:20 | https://www.reddit.com/r/LocalLLaMA/comments/15o8yg5/some_of_the_odd_safety_responses_from_llama/ | bigattichouse | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {'gid_3': 1} | 15o8yg5 | false | null | t3_15o8yg5 | /r/LocalLLaMA/comments/15o8yg5/some_of_the_odd_safety_responses_from_llama/ | false | false | self | 1 | null |
Which are minimal CPU and GPU for run 65b ? | 1 | Hello ! I want to start work with LLaMa 2(65B), but I don’t sure that I can start work with AI.
I have RTX 3060 12GB, 16 GB RAM.
Can I start LLaMa 2 on this configuration?
Or I need RTX 40x ?
(I want to use 2 qubit or 4 qubit version) | 2023-08-11T13:29:36 | https://www.reddit.com/r/LocalLLaMA/comments/15o8zgs/which_are_minimal_cpu_and_gpu_for_run_65b/ | Alex_Strek | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15o8zgs | false | null | t3_15o8zgs | /r/LocalLLaMA/comments/15o8zgs/which_are_minimal_cpu_and_gpu_for_run_65b/ | false | false | self | 1 | null |
Langchain + LLaMa 2 consuming too much VRAM | 1 | I was playing around with a GitHub project on a conda environment on Windows and I was surprised to see that LLama 2 13B 4bit was using up to 25GB VRAM (16GB on one GPU and 9GB on the second one) for a simple summarization task on a document with less than 4KB. I wanted to find out if there was an issue with Langchain or if it's just how it goes with LLMs because seeing that a model occupies 8GB on disk and uses almost triple the amount of VRAM is suspicious. Am I doing something wrong? | 2023-08-11T14:21:51 | https://www.reddit.com/r/LocalLLaMA/comments/15oaa18/langchain_llama_2_consuming_too_much_vram/ | Mdfkrjones | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oaa18 | false | null | t3_15oaa18 | /r/LocalLLaMA/comments/15oaa18/langchain_llama_2_consuming_too_much_vram/ | false | false | self | 1 | null |
Any benefit to multi-GPUs? | 1 | Sorry for the more newb-oriented question, but I'm about to build a dual-boot setup using a 3090 but also own an older 1080ti and was just curious if there is any benefit at all to having both. I know SLI is impossible, so curious if there would be any other benefit. TIA! | 2023-08-11T14:31:28 | https://www.reddit.com/r/LocalLLaMA/comments/15oaidh/any_benefit_to_multigpus/ | drycounty | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oaidh | false | null | t3_15oaidh | /r/LocalLLaMA/comments/15oaidh/any_benefit_to_multigpus/ | false | false | self | 1 | null |
Total newbie. Are these token speeds normal? | 15 | I hope people can help me out here, total newbie and I am a bit clueless.
I have recently started trying out large language models after reading out about it and came up with mixed results. After reading out extensively I have decided to use KoboldCPP, and referenced a chart which stated that with the 32GB RAM I have using 13B models should be easy. I once again looked around and it seemed that a new model by the name of MythoMix was both smart and was good at roleplaying, so I decided to go with it.
My problems start here. The response time for the model is really slow! Now, if this is normal then I will gladly accept it, however my impression was that with 32GB RAM and using a 13B model, I should have fairly good token speeds.
So I guess my question is this: Is 0.5T/s to 2.0T/s speed normal for 32GB RAM for "mythomix-l2-13b.ggmlv3.q4_K_M" using KoboldCPP?
I will leave extra information I have below.
https://files.catbox.moe/roowgl.jpg
https://files.catbox.moe/bsajux.jpg
https://files.catbox.moe/iph51u.jpg
For the above picture of KoboldCPP screen, after seeing how slow the response times were I played around a bit and saw people recommending using different presets, I picked Cublas. I also played around with the GPU layers but that doesn't seem to have any effect.
Intel(R) Core(TM) i7-12700H
NVIDIA GeForce RTX 3060
32 GB RAM
6 GB VRAM
I hope I haven't said something foolish that made no sense whatsoever, I tried to read up and learn but I know that I might still be lacking in that area. If I did make a mistake do forgive me. | 2023-08-11T14:41:57 | https://www.reddit.com/r/LocalLLaMA/comments/15oarso/total_newbie_are_these_token_speeds_normal/ | HDTurtle | self.LocalLLaMA | 2023-08-11T14:52:24 | 0 | {} | 15oarso | false | null | t3_15oarso | /r/LocalLLaMA/comments/15oarso/total_newbie_are_these_token_speeds_normal/ | false | false | self | 15 | {'enabled': False, 'images': [{'id': '09bLCBZkzdmF3jJjTV2WEQEJnHBMbnd3MBoznMGDThs', 'resolutions': [{'height': 198, 'url': 'https://external-preview.redd.it/BGh4LaT0BuWwX97OoPVukkla19q-geyauVQ0D5Jw4RM.jpg?width=108&crop=smart&auto=webp&s=da5fd222d19b6b5258ec8e660f326fe3a67d58d7', 'width': 108}, {'height': 396, 'url': 'https://external-preview.redd.it/BGh4LaT0BuWwX97OoPVukkla19q-geyauVQ0D5Jw4RM.jpg?width=216&crop=smart&auto=webp&s=560997102fd05d5c074304a7a63c9d1d6c94e7f0', 'width': 216}, {'height': 587, 'url': 'https://external-preview.redd.it/BGh4LaT0BuWwX97OoPVukkla19q-geyauVQ0D5Jw4RM.jpg?width=320&crop=smart&auto=webp&s=160b0955d0f4a9d58368d218fd8f5e82df089e15', 'width': 320}], 'source': {'height': 610, 'url': 'https://external-preview.redd.it/BGh4LaT0BuWwX97OoPVukkla19q-geyauVQ0D5Jw4RM.jpg?auto=webp&s=272519387e41a4e726ab9e02d5cc76435814af7c', 'width': 332}, 'variants': {}}]} |
Introduce the newest WizardMath models (70B/13B/7B) ! | 1 | WizardMath 70B achieves:
1. Surpasses **ChatGPT-3.5**, **Claude Instant-1**, **PaLM-2** and **Chinchilla** on GSM8k with ***81.6 Pass@1***
2. Surpasses **Text-davinci-002**, **GAL**, **PaLM**, **GPT-3** on MATH with ***22.7 Pass@1***
3. Surpasses **all other open-source LLMs** on both GSM8k and MATH by a substantial margin
4. Paper is coming next week, with brand-new ***Reinforced Evol-Instruct*** method for math LLMs
Twitter: [(5) WizardLM on Twitter: "🔥🔥🔥 Introduce the newest WizardMath models (70B/13B/7B) ! WizardMath 70B achieves: 1. Surpasses ChatGPT-3.5, Claude Instant-1, PaLM-2 and Chinchilla on GSM8k with 81.6 Pass@1 2. Surpasses Text-davinci-002, GAL, PaLM, GPT-3 on MATH with 22.7 Pass@1 3. Surpasses all other… https://t.co/XOSggcOePP" / X](https://twitter.com/WizardLM_AI/status/1689998428200112128)
HF Weights: [https://huggingface.co/WizardLM/WizardMath-70B-V1.0](https://huggingface.co/WizardLM/WizardMath-70B-V1.0)
Github: [WizardLM/WizardMath at main · nlpxucan/WizardLM (github.com)](https://github.com/nlpxucan/WizardLM/tree/main/WizardMath)
​
​
​
​
https://preview.redd.it/tteefzagvhhb1.png?width=1321&format=png&auto=webp&s=96945fa1ce01bc09f4ac318c63a7865e749d2c6b
https://preview.redd.it/uum9t0bgvhhb1.png?width=1780&format=png&auto=webp&s=ae75cb3094b043f25ddfd9e039d00875894e7aa2
​ | 2023-08-11T14:53:12 | https://www.reddit.com/r/LocalLLaMA/comments/15ob1zh/introduce_the_newest_wizardmath_models_70b13b7b/ | ApprehensiveLunch453 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ob1zh | false | {'oembed': {'author_name': 'WizardLM', 'author_url': 'https://twitter.com/WizardLM_AI', 'cache_age': 3153600000, 'height': None, 'html': '<blockquote class="twitter-video"><p lang="en" dir="ltr">🔥🔥🔥<br>Introduce the newest WizardMath models (70B/13B/7B) ! <br><br>WizardMath 70B achieves:<br>1. Surpasses ChatGPT-3.5, Claude Instant-1, PaLM-2 and Chinchilla on GSM8k with 81.6 Pass@1<br><br>2. Surpasses Text-davinci-002, GAL, PaLM, GPT-3 on MATH with 22.7 Pass@1<br><br>3. Surpasses all other… <a href="https://t.co/XOSggcOePP">pic.twitter.com/XOSggcOePP</a></p>— WizardLM (@WizardLM_AI) <a href="https://twitter.com/WizardLM_AI/status/1689998428200112128?ref_src=twsrc%5Etfw">August 11, 2023</a></blockquote>\n<script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>\n', 'provider_name': 'Twitter', 'provider_url': 'https://twitter.com', 'type': 'rich', 'url': 'https://twitter.com/WizardLM_AI/status/1689998428200112128', 'version': '1.0', 'width': 350}, 'type': 'twitter.com'} | t3_15ob1zh | /r/LocalLLaMA/comments/15ob1zh/introduce_the_newest_wizardmath_models_70b13b7b/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'eyFFaewGfuH1qS_paH6sm4q831ml5iEwBALV-n_H0g8', 'resolutions': [{'height': 158, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=108&crop=smart&auto=webp&s=84b21dc2e0326c7eb1a291ab12af09713b665f2d', 'width': 108}, {'height': 317, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=216&crop=smart&auto=webp&s=6458ba3e240f85390c260ae4f57c454bf468c0b3', 'width': 216}, {'height': 469, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=320&crop=smart&auto=webp&s=43034fdbaa7ee734be37a0e36789ea12ca302dff', 'width': 320}, {'height': 939, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=640&crop=smart&auto=webp&s=f33c26e051a1061ddf0ea84b792262d4536c05c3', 'width': 640}, {'height': 1408, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=960&crop=smart&auto=webp&s=cb09e3c5b0837185aff6a7be1a55e6f1498862c3', 'width': 960}, {'height': 1585, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?width=1080&crop=smart&auto=webp&s=1739083b21869b9ac23b1e51ec2ac774b9c75266', 'width': 1080}], 'source': {'height': 1930, 'url': 'https://external-preview.redd.it/dyM5Kd44kaad8i6-WBvsbK92hwNouijcVmEuyloHf00.jpg?auto=webp&s=8feb886890364055c561abed7643a71f85d3ac3b', 'width': 1315}, 'variants': {}}]} |
|
Grammar correction and sentence rephrase | 1 | Can I use Llama to build a local model that can rephrase sentences and correct grammar? GenAI tools like ChatGpt aren’t allowed at workplace, so I want to have a local model to help me to write better. | 2023-08-11T15:35:57 | https://www.reddit.com/r/LocalLLaMA/comments/15oc5uu/grammar_correction_and_sentence_rephrase/ | PsychologicalSky4377 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oc5uu | false | null | t3_15oc5uu | /r/LocalLLaMA/comments/15oc5uu/grammar_correction_and_sentence_rephrase/ | false | false | self | 1 | null |
ChatGPT and its Doppelgangers: A Study on the Limits of Model Imitation | 7 | I found an [interesting study](https://arxiv.org/abs/2305.15717) discussing ChatGPT "imitation models" like Alpaca and Vicuna. Here are the bullet points:
* Emerging method involves finetuning weaker language models on outputs from stronger models, like ChatGPT, to imitate their capabilities using open-source models.
* Research involved finetuning various LMs to mimic ChatGPT using different model sizes, data sources, and imitation data amounts.
* Initial findings showed the imitation models were good at following instructions and were rated similarly to ChatGPT by crowd workers.
* Targeted automatic evaluations revealed imitation models failed to bridge the capability gap between the base LM and ChatGPT, especially in tasks not prevalent in imitation data.
* Imitation models effectively mimic ChatGPT's style but fall short in factuality.
* Conclusion: Model imitation is not the best approach due to the capabilities gap. Emphasis should be on improving base LMs instead of trying to imitate proprietary systems.
What are your thoughts on this? Do you agree with their conclusion? | 2023-08-11T15:38:59 | https://www.reddit.com/r/LocalLLaMA/comments/15oc8ji/chatgpt_and_its_doppelgangers_a_study_on_the/ | DecipheringAI | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oc8ji | false | null | t3_15oc8ji | /r/LocalLLaMA/comments/15oc8ji/chatgpt_and_its_doppelgangers_a_study_on_the/ | false | false | self | 7 | {'enabled': False, 'images': [{'id': 'q3evP6JeDpAC2MdSQHWYxnCYTqbJkElIQsLFqVSdkss', 'resolutions': [{'height': 63, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=108&crop=smart&auto=webp&s=2711d572cfc6c713893cf24e8c4a7344d5ad8a4c', 'width': 108}, {'height': 126, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=216&crop=smart&auto=webp&s=b6624f0c1eedc14997e7f1780efbe6e5cb50c1e2', 'width': 216}, {'height': 186, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=320&crop=smart&auto=webp&s=9db38144ef3065833b9ba158c764f7be47de3016', 'width': 320}, {'height': 373, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=640&crop=smart&auto=webp&s=72b056142e7533b5628a2a34f37f7e5415727075', 'width': 640}, {'height': 560, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=960&crop=smart&auto=webp&s=2637f961ee21190172b9ca6c8adf3ac9612db083', 'width': 960}, {'height': 630, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?width=1080&crop=smart&auto=webp&s=782eead871df2939a587ee3beae442cc59282f64', 'width': 1080}], 'source': {'height': 700, 'url': 'https://external-preview.redd.it/0HhwdU6MKIAKjL9Y8-B_iH374a3NiPTy0ib8lmloRzA.jpg?auto=webp&s=f1cd025aeb52ffa82fc9e5a4a2f157da0d919147', 'width': 1200}, 'variants': {}}]} |
New model and new app - Layla | 1 | [removed] | 2023-08-11T15:39:15 | https://www.reddit.com/r/LocalLLaMA/comments/15oc8rp/new_model_and_new_app_layla/ | Tasty-Lobster-8915 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oc8rp | false | null | t3_15oc8rp | /r/LocalLLaMA/comments/15oc8rp/new_model_and_new_app_layla/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'WcDPzOnQZ3t8b1fwQPJ1k01l878a2HIs1GCu8CJR5Wc', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=108&crop=smart&auto=webp&s=94f499cdd8453f6de73be6128e8745af9395e73a', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=216&crop=smart&auto=webp&s=fc6609edde1d19bad64a317c91df2995357948cc', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=320&crop=smart&auto=webp&s=fa369a7fe8db0d33cdeacdf39a268a07a308c1c1', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=640&crop=smart&auto=webp&s=b5ecbc1e760ea8a7663379961b869a67fa6d308d', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=960&crop=smart&auto=webp&s=2c75459f65f93f055b4ad48195bd3825f3a32c66', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?width=1080&crop=smart&auto=webp&s=8aff96d05385988fabfdbdecfa12e2cac1a3f579', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/okf3dEs9GIyG6sSxiAzvSv19-OLWzycgPVIcM18CaSQ.jpg?auto=webp&s=fbf3159bd48475d763aeb36af21027a0c93108c3', 'width': 1200}, 'variants': {}}]} |
PrivateGPT example with Llama 2 Uncensored | 1 | 2023-08-11T16:30:51 | https://github.com/jmorganca/ollama/tree/main/examples/privategpt | helloPenguin006 | github.com | 1970-01-01T00:00:00 | 0 | {} | 15odjmy | false | null | t3_15odjmy | /r/LocalLLaMA/comments/15odjmy/privategpt_example_with_llama_2_uncensored/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'qWYf_hGwsFfjEOzHhraYQjkUJJlsotgW5CofgR3t1f4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=108&crop=smart&auto=webp&s=cc5a7d81b1db7f17d71cab5a1a022c483ba8d216', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=216&crop=smart&auto=webp&s=d28ec690d6c6c33b274bf2dde3c7f27ead4be5bb', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=320&crop=smart&auto=webp&s=30fa83699356cf1925298ab5259a705b8f24ccee', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=640&crop=smart&auto=webp&s=294fd6b1499ee75bc2808308b8a406a1361b2611', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=960&crop=smart&auto=webp&s=21dae982cf5c0af79b36998e542094273e9f52b4', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?width=1080&crop=smart&auto=webp&s=89265375793472b434f3b2c71bf5f029ca9c1d5b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/YHtfBPBmwuqL4EW9yt8j6W7UxDd5zR0pFNarkurUzC0.jpg?auto=webp&s=2f0d610fee1af34ec04e0abed720ec9c5180e0c8', 'width': 1200}, 'variants': {}}]} |
||
Can a team of 10-20 people access a Llama 2 model deployed in a local server with medium requirements? | 31 | I'm planning on spending $3-5k on a local server with Llama v2 deployed on it, such as a team of 10-20 people can each access the inference from their own computers whenever they please. Since I'm not really an infra guy, I have questions on how to approach this. I guess that while someone is running a query nobody else can run theirs until the first one is complete, correct? Is there any easy way to run Llama locally in a way where multiple people can access it synchronously? | 2023-08-11T16:41:57 | https://www.reddit.com/r/LocalLLaMA/comments/15odtsn/can_a_team_of_1020_people_access_a_llama_2_model/ | Heco1331 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15odtsn | false | null | t3_15odtsn | /r/LocalLLaMA/comments/15odtsn/can_a_team_of_1020_people_access_a_llama_2_model/ | false | false | self | 31 | null |
Does HF inference endpoint work? | 1 | I’ve been trying to deploy LLMs using HF inference endpoint (e.g. stablecode-instruct-alpha-3b, llama2, etc) but the deployment fails all the time, yet I’ve been charged for an hour while it remains in “installing” state. It never becomes “ready for inference” with the form to query model, I’ve tried them with various CPU/GPU configs.
Is there any rocket science behind the deployment?
And the support via email — it’s so weird. | 2023-08-11T16:42:15 | https://www.reddit.com/r/LocalLLaMA/comments/15odu23/does_hf_inference_endpoint_work/ | Greg_Z_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15odu23 | false | null | t3_15odu23 | /r/LocalLLaMA/comments/15odu23/does_hf_inference_endpoint_work/ | false | false | self | 1 | null |
llama | 1 | 2023-08-11T17:16:23 | https://teesdesk-us.shop/limited-edition-273 | AnneCampbell54 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 15oepre | false | null | t3_15oepre | /r/LocalLLaMA/comments/15oepre/llama/ | false | false | 1 | null |
||
Anyone got TextGen/LlamaCPP working with Metal for new GGML models and Llama2? | 1 | I'm getting constant errors/crashes, even though I updated Torch to nightly and rebuilt the LlamaCPP wheel with Metal. | 2023-08-11T17:16:57 | https://www.reddit.com/r/LocalLLaMA/comments/15oeqaf/anyone_got_textgenllamacpp_working_with_metal_for/ | -becausereasons- | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oeqaf | false | null | t3_15oeqaf | /r/LocalLLaMA/comments/15oeqaf/anyone_got_textgenllamacpp_working_with_metal_for/ | false | false | self | 1 | null |
How to get the answer from local llama2 and send it to my app? | 1 | I'm building a software that needs to use information from Llama2 query.
I'm using oobabooga. Is there a way to do this with it?
Do you need to code your own local API? How do you guys retrieve the information from the chat? | 2023-08-11T17:48:25 | https://www.reddit.com/r/LocalLLaMA/comments/15ofjzj/how_to_get_the_answer_from_local_llama2_and_send/ | ppcfbadsfree | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ofjzj | false | null | t3_15ofjzj | /r/LocalLLaMA/comments/15ofjzj/how_to_get_the_answer_from_local_llama2_and_send/ | false | false | self | 1 | null |
LLM, Semantic search and large volume of documents | 1 | Hello,
I know that this question was probably asked a few times but I really cannot decide on the best approach and could use your help.
We have hundreds of thousands of documents and we want to create a "chatbot" that could possibly answer questions that can only be found in those documents. Now, the documents can be very very similar but contain data for different dates (textual data).
Would a vector database using semantic search work? And then passing the result to an LLM? (Llama 2) Or is there a better approach these days? Currently thinking of running milvus as a vector db and connecting that to an LLM via langchain. Any guidance, recommendations, suggestions are highly appreciated!
We do have the resources to host an LLM and a vector db. | 2023-08-11T17:58:26 | https://www.reddit.com/r/LocalLLaMA/comments/15oftk7/llm_semantic_search_and_large_volume_of_documents/ | Mayloudin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oftk7 | false | null | t3_15oftk7 | /r/LocalLLaMA/comments/15oftk7/llm_semantic_search_and_large_volume_of_documents/ | false | false | self | 1 | null |
Access to my server with a httpRequest or other | 1 | My model is runing on localhost:7860
I want to access it, I have tryed with python
import requests
request = {'prompt': 'hi', 'max\_new\_tokens': 4096}
r = requests.post(url='[http://localhost:7860/api/v1/generate](http://localhost:7860/api/v1/generate)', json=request)
print(r.json())
I have on request reply : detail:not found or detail:method not allowed
What's wrong?
CG. | 2023-08-11T18:01:38 | https://www.reddit.com/r/LocalLLaMA/comments/15ofwpo/access_to_my_server_with_a_httprequest_or_other/ | ppcfbadsfree | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ofwpo | false | null | t3_15ofwpo | /r/LocalLLaMA/comments/15ofwpo/access_to_my_server_with_a_httprequest_or_other/ | false | false | self | 1 | null |
New Model RP Comparison/Test (7 models tested) | 1 | This is a follow-up to my previous post here: [Big Model Comparison/Test (13 models tested) : LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/15lihmq/big_model_comparisontest_13_models_tested/)
Here's how I evaluated these (same methodology as before) for their role-playing (RP) performance:
- Same (complicated and limit-testing) long-form conversation with all models, [SillyTavern](https://github.com/SillyTavern/SillyTavern) frontend, [KoboldCpp](https://github.com/LostRuins/koboldcpp) backend, GGML q5\_K\_M, Deterministic generation settings preset, [Roleplay instruct mode preset](https://www.reddit.com/r/LocalLLaMA/comments/15mu7um/sillytaverns_roleplay_preset_vs_modelspecific/), > 22 messages, going to full 4K context, noting especially good or bad responses.
So here's the list of models and my notes plus my very personal rating (➕ = worth a try, - ➖ disappointing, ❌ = unusable):
- ➕ **[huginnv1.2](https://huggingface.co/TheBloke/huginnv1.2-GGML)**: Much better than the previous version (Huginn-13B), very creative and elaborate, focused one self-made plot point early on, nice writing and actions/emotes, repetitive emoting later, redundant speech/actions (says what she's going to do and then emotes doing it), missed important detail later and became nonsensical because of that. More creative but less smart than other models.
- ➖ **[MythoMix-L2-13B](https://huggingface.co/TheBloke/MythoMix-L2-13B-GGML)**: While other models often went too fast, this one needed a bit of coaxing to proceed, got confused about who's who and anatomy, mixing up people and instructions, wrote what User does, actions switched between second and third person. But good actions and descriptions, and believable and lively characters, and no repetition/looping all the way to full 4K context and beyond! **Only gets a ➖ instead of a ➕ because there's already a successor, [MythoMax-L2-13B-GGML](https://huggingface.co/TheBloke/MythoMax-L2-13B-GGML), which I like even more!**
- ➕ **[MythoMax-L2-13B](https://huggingface.co/TheBloke/MythoMax-L2-13B-GGML)**: Started talking/acting as User (had to use non-deterministic preset and enable "Include Names" for the first message)! While other models often went too fast, this one needed a bit of coaxing to proceed, got confused about who's who and anatomy, mixing up people and instructions, mentioned scenario being a simulation. But nice prose and excellent writing, and no repetition/looping all the way to full 4K context and beyond! **This is my favorite of this batch! I'll use this a lot more from now on, right now it's my second favorite model next to my old favorite [Nous-Hermes-Llama2](https://huggingface.co/TheBloke/Nous-Hermes-Llama2-GGML)!**
- ➖ **[orca_mini_v3_13B](https://huggingface.co/TheBloke/orca_mini_v3_13B-GGML)**: Repeated greeting message verbatim (but not the emotes), talked without emoting, spoke of agreed upon parameters regarding limits/boundaries, terse/boring prose, had to ask for detailed descriptions, description was in past tense, speech within speech, wrote what User does, got confused about who's who and anatomy, became nonsensical later. **May be a generally smart model, but apparently not a good fit for roleplay!**
- ➖ **[Stable-Platypus2-13B](https://huggingface.co/TheBloke/Stable-Platypus2-13B-GGML)**: Extremely short and terse responses (despite Roleplay preset!), had to ask for detailed descriptions, got confused about who's who and anatomy, repetitive later. But good and long descriptions when specifically asked for! **May be a generally smart model, but apparently not a good fit for roleplay!**
- ❌ **[vicuna-13B-v1.5-16K](https://huggingface.co/TheBloke/vicuna-13B-v1.5-16K-GGML)**: Confused about who's who from the start, acted and talked as User, repeated greeting message verbatim (but not the very first emote), normal afterwards (talks and emotes and uses emoticons normally), but mentioned boundaries/safety multiple times, described actions without doing them, needed specific instructions to act, switched back from action to description in the middle of acting, repetitive later, some confusion. Seemed less smart (grammar errors, mix-ups), but great descriptions and sense of humor, but broke down completely within 20 messages (> 4K tokens)! **SCALING ISSUE (despite using `--contextsize 16384 --ropeconfig 0.25 10000`)?**
- ❌ **[WizardMath-13B-V1.0](https://huggingface.co/TheBloke/WizardMath-13B-V1.0-GGML)**: Ends every message with "The answer is: ", making it unsuitable for RP! So I instead did some logic tests - unfortunately it failed them all ("Sally has 3 brothers...", "What weighs more, two pounds of feathers or one pound of bricks?", and "If I have 3 apples and I give two oranges...") even with "Let's think step by step." added.
Looking forward to your comments, especially if you have widely different experiences, so I may go back to retest some models with different settings... | 2023-08-11T18:17:40 | https://www.reddit.com/r/LocalLLaMA/comments/15ogc60/new_model_rp_comparisontest_7_models_tested/ | WolframRavenwolf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ogc60 | false | null | t3_15ogc60 | /r/LocalLLaMA/comments/15ogc60/new_model_rp_comparisontest_7_models_tested/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'bDW7jyCB5L7RKBwRUqrzWSn3bIb_Szu_GogYRebiCjw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=108&crop=smart&auto=webp&s=22d2e1896c94ecebda58fed69478453d4b16fd4f', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=216&crop=smart&auto=webp&s=019bd779b582098d4b9aa01b87ee530132195fa6', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=320&crop=smart&auto=webp&s=55daeabbed00d9b3c1e7f3207edea4d0a265db39', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=640&crop=smart&auto=webp&s=47d7877d194270162d75f4922c4ecb60b17c101d', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=960&crop=smart&auto=webp&s=004f5643d41eee63624b163efc53427073882f4f', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?width=1080&crop=smart&auto=webp&s=e6ee7ad7840a9a71890c76db5e4df6a3f669e762', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/NFgRhAZ_cgs4xao5V1cOWNjptqU5JwIiGBtsvOdhlPU.jpg?auto=webp&s=44d160d8b5087122f25fba2443dc2c5a77adf472', 'width': 1280}, 'variants': {}}]} |
Out of memory using multiple GPUs | 1 | I have an EC2 p2.8xlarge instance running on AWS with 8x Nvidia K80 GPUs, each with 12 GB VRAM for a total of 96 GB. I am trying to run LLaMA 2, and have tried both 7B and 70B. If I run it with 7B, I get the error `loading checkpoint for MP=1 but world size is 8`, and with 70B, `torch.cudatorch.OutOfMemoryError: Tried to allocate 448.00 MiB (GPU 7; 11.17 GiB total capacity, 10.21 GiB already allocated; 324.19 MiB free; 10.62 GiB reserved in total by PyTorch)`. How can I spread the memory across all GPUs? | 2023-08-11T18:28:31 | https://www.reddit.com/r/LocalLLaMA/comments/15ogmc9/out_of_memory_using_multiple_gpus/ | EffectiveFood4933 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ogmc9 | false | null | t3_15ogmc9 | /r/LocalLLaMA/comments/15ogmc9/out_of_memory_using_multiple_gpus/ | false | false | self | 1 | null |
Thinking about purchasing a 4090 for KoboldCPP... Got some questions. | 1 | So currently I'm using a 5600g with 32GB of ram and a 12GB 3060 on Linux.
What I would like to do is try and find a 24GB-ish LLM model that excels at collaborative story writing that I can run on my present hardware (doesn't matter how slow it is) just to get an idea of what improvements the 4090 would give me. My main goal is to have coherent generation and have the model stay on track and produce few anomalies.
It would be amazing if you could make two suggestions for me:
1. Which model I should use (something that a 4090 can fully utilize but will also be "usable" on my present hardware).
2. What command launch options I should use. For example: I don't particularly understand what ropeconfig is and am confused why some of us set context size as a launch option when you can set it in the interface.
Thanks for reading and any replies.
What an awesome community this is turning out to be. Very happy to be here with y'all.
Cheers. | 2023-08-11T18:47:00 | https://www.reddit.com/r/LocalLLaMA/comments/15oh3ie/thinking_about_purchasing_a_4090_for_koboldcpp/ | wh33t | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oh3ie | false | null | t3_15oh3ie | /r/LocalLLaMA/comments/15oh3ie/thinking_about_purchasing_a_4090_for_koboldcpp/ | false | false | self | 1 | null |
What the fuck is wrong with WizardMath??? | 222 | 2023-08-11T18:48:55 | bot-333 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 15oh576 | false | null | t3_15oh576 | /r/LocalLLaMA/comments/15oh576/what_the_fuck_is_wrong_with_wizardmath/ | false | false | 222 | {'enabled': True, 'images': [{'id': 'xbug615PVicVm3MHsJ-wILaxpgeEIJx5h-4v1PZoxAQ', 'resolutions': [{'height': 44, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=108&crop=smart&auto=webp&s=a81fb1bfbf27bf6a8d9bd4458ae4eb8578d42dbf', 'width': 108}, {'height': 89, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=216&crop=smart&auto=webp&s=ebf65c1ae8ef5de63d96926611fb6a2eb07813b4', 'width': 216}, {'height': 131, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=320&crop=smart&auto=webp&s=1ace9e8d2f4bbd82cf42b2bd4eb879fc51075174', 'width': 320}, {'height': 263, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=640&crop=smart&auto=webp&s=ecf1d96609e661cc91206f3bdb7659b8ef039c1d', 'width': 640}, {'height': 395, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=960&crop=smart&auto=webp&s=d725b6df4e5816a9e32e914c41255f2d48c1dcf2', 'width': 960}, {'height': 445, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?width=1080&crop=smart&auto=webp&s=bed9bb6b349616ac22f527e979dc29b936b51d68', 'width': 1080}], 'source': {'height': 716, 'url': 'https://preview.redd.it/47823tkm1jhb1.png?auto=webp&s=12513bec3bda3b09f781799a2ea95a1454181bc4', 'width': 1736}, 'variants': {}}]} |
|||
Is it possible to use multiple GPUs of different generations on a single PC? | 1 | I recently swapped out my 1070 with a 3060 to load 13b 4-bit models, and it's working like a charm, but I would like to try for more. I am on a budget, so I am hesitant to buy another 3060 if it's not necessary. Would it be possible to put the 1070 back in my machine, and use both GPUs at once?
I currently have the drivers for the 3060 installed, and I heard that it's not possible to install two different sets of Geforce drivers on one machine. Would the 1070 function with the 3060 drivers for the purpose of loading bigger models through exllama? Would my machine recognize it, and would I benefit from it? | 2023-08-11T19:17:34 | https://www.reddit.com/r/LocalLLaMA/comments/15ohw0g/is_it_possible_to_use_multiple_gpus_of_different/ | Zugzwang_CYOA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ohw0g | false | null | t3_15ohw0g | /r/LocalLLaMA/comments/15ohw0g/is_it_possible_to_use_multiple_gpus_of_different/ | false | false | self | 1 | null |
Platypus models | 1 | 2023-08-11T19:19:39 | https://twitter.com/natanielruizg/status/1690048207030493189 | ninjasaid13 | twitter.com | 1970-01-01T00:00:00 | 0 | {} | 15ohxvx | false | {'oembed': {'author_name': 'Nataniel Ruiz', 'author_url': 'https://twitter.com/natanielruizg', 'cache_age': 3153600000, 'height': None, 'html': '<blockquote class="twitter-video"><p lang="en" dir="ltr">We are 🔥super excited🔥 to release the Platypus family of finetuned LLMs 🥳🥳. Platypus achieves the top score in the Hugging Face Open LLM Leaderboard 🏆! The main focus of our work is to achieve cheap, fast and powerful refinement of base LLMs.<br>page: <a href="https://t.co/QHJ6kDoCYa">https://t.co/QHJ6kDoCYa</a> <a href="https://t.co/MOSiflQLDU">pic.twitter.com/MOSiflQLDU</a></p>— Nataniel Ruiz (@natanielruizg) <a href="https://twitter.com/natanielruizg/status/1690048207030493189?ref_src=twsrc%5Etfw">August 11, 2023</a></blockquote>\n<script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>\n', 'provider_name': 'Twitter', 'provider_url': 'https://twitter.com', 'type': 'rich', 'url': 'https://twitter.com/natanielruizg/status/1690048207030493189', 'version': '1.0', 'width': 350}, 'type': 'twitter.com'} | t3_15ohxvx | /r/LocalLLaMA/comments/15ohxvx/platypus_models/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'vVURZSvN8RjsPJbzUNoyI4xYt1V2yGYOSNsvuKbFmaQ', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/RgulP3L45yR0BcX2I4VThtZt02S3L7XjFZ8D3rEGKo4.jpg?width=108&crop=smart&auto=webp&s=62a0e833e88283fd675c0789e2ffad916cb1b1f3', 'width': 108}], 'source': {'height': 140, 'url': 'https://external-preview.redd.it/RgulP3L45yR0BcX2I4VThtZt02S3L7XjFZ8D3rEGKo4.jpg?auto=webp&s=cb48fa81b98753463cf4c36e24d2b3eda10da9a1', 'width': 140}, 'variants': {}}]} |
||
Preventing LLAMA from hallucinating responses. | 1 | So we are using LLAMA is a typical RAG scenario, give it some context and ask it a question. What i have found is, no matter how much i yell at it in the prompt, for certain questions, it always gives the wrong, hallucinated answer, even if the right answer is in the document inside.
For example, the document would be like:
Student A has score 100
Student B has score 95
Student C has score 99
(very very simplified, in reality these are all chunks - about 200 tokens - and there are a dozen chunks)
LLAMA will always answer Student B wrong. All others are right - just one is wrong. GPT3/4 do not have this problem, although GPT3 did once in a while, and on the same student. Which is bizarre.
GPT 4 however, 100% of time is correct.
​
Second category - if you ask it score history for example - when did student x score decrease - it will bring up a place where the score increased and write about it as if it decreased. Which is non sensical. GPT3 does the same thing. GPT4 will tell you - It does not appear the score for this student decreased.
What is GPT 4 doing? How can I help llama or even gpt3 do something like this? Is this just RLHF? Am i stuck using gpt4 for this?
​
​ | 2023-08-11T20:09:17 | https://www.reddit.com/r/LocalLLaMA/comments/15oj83h/preventing_llama_from_hallucinating_responses/ | Alert_Record5063 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15oj83h | false | null | t3_15oj83h | /r/LocalLLaMA/comments/15oj83h/preventing_llama_from_hallucinating_responses/ | false | false | self | 1 | null |
Encourage Your Workplace to Host its Own LLMs | 1 | [ChatGPT fever spreads to US workplace, sounding alarm for some](https://www.reuters.com/technology/chatgpt-fever-spreads-us-workplace-sounding-alarm-some-2023-08-11/)
The biggest issue with the ubiquitous use of ChatGPT in the workplace is that all of the information gets leaked. Most corporations who are interested in making money off their ideas should be interested in keeping those ideas largely quiet. This is not widely appreciated by employers today.
You think you may care about NSFW work or censorship. Think about how much Siemens will care once they realize their workers are divulging their trade secrets.
If we can a reasonable number of corporations to start running internal LLMs, the size of the addressable market for LLMs will grow exponentially. It will be fantastic for LLaMA and home LLM use orthogonally.
Please advocate for local LLM use at work. Thank you for your time. | 2023-08-11T20:18:04 | https://www.reddit.com/r/LocalLLaMA/comments/15ojg5c/encourage_your_workplace_to_host_its_own_llms/ | friedrichvonschiller | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ojg5c | false | null | t3_15ojg5c | /r/LocalLLaMA/comments/15ojg5c/encourage_your_workplace_to_host_its_own_llms/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '67c-IjzWz8qmTmo-aRRdu58s6Tmf9wruf18jvgnZz0w', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=108&crop=smart&auto=webp&s=c0022678e0ff8b9660760f7d8383f89338d69f0f', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=216&crop=smart&auto=webp&s=ae94d7e0f34a909ea0a594bae054b56a5802a21f', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=320&crop=smart&auto=webp&s=71f85e485328bf95a39bf1b2d2413550faf66815', 'width': 320}, {'height': 334, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=640&crop=smart&auto=webp&s=d8ba89c96f3a4c336b059eed7cffc634b0eca737', 'width': 640}, {'height': 502, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=960&crop=smart&auto=webp&s=727363fa992caac483cbaa04f7b22b95469f969b', 'width': 960}, {'height': 565, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?width=1080&crop=smart&auto=webp&s=45f2e263d547168fdc730e1f99703c9b28972403', 'width': 1080}], 'source': {'height': 628, 'url': 'https://external-preview.redd.it/2kMixBo1N7fReiDsbtxGR48VdJ3f9OeQ0gdMGmMxz7c.jpg?auto=webp&s=b6e5440e0dca913a3dff038c1edfc3e4384fac67', 'width': 1200}, 'variants': {}}]} |
How to use multiple GPUs on different systems? | 1 | I want to use my Gaming Laptop with an 8GB 3080, and one other system with an 8GB RX580. I don't know what the performance hit would be. TBH I just want to do it for the sake of it! I'm pretty new to running LLMs, so some explanation would be really helpful!
Thankyou! | 2023-08-11T20:28:25 | https://www.reddit.com/r/LocalLLaMA/comments/15ojptj/how_to_use_multiple_gpus_on_different_systems/ | KvAk_AKPlaysYT | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15ojptj | false | null | t3_15ojptj | /r/LocalLLaMA/comments/15ojptj/how_to_use_multiple_gpus_on_different_systems/ | false | false | self | 1 | null |
Trouble Running Llama-2 70B on HPC with Limited GPUs - Need Help! | 1 | I'm utilizing Llama-2 on a high-performance computing (HPC) setup and dispatching tasks through SLURM. I managed to run the Llama 7B model, but I ran into problems with the 70B variant.
It seems that the Llama-2 70B model anticipates 8 distinct GPUs, given its MP configuration of 8, implying nproc\_per\_node = 8 in torchrun settings. However, my HPC only allows 4 GPUs per node.
Does anyone know if running the 70B model is feasible under this constraint? I think I might be missing a workaround, especially since the HPC boasts high-end GPUs like the A100. For clarity, here's the SLURM configuration I'm deploying.
​
\#!/bin/bash
\#SBATCH --job-name=llama\_chat\_run # Change the job name to something more descriptive
\#SBATCH --nodes=1
\#SBATCH --cpus-per-task=4
\#SBATCH --ntasks-per-node=1
\#SBATCH --mem=16GB
\#SBATCH --gres=gpu:1
\#SBATCH --time=2:00:00 # Extend runtime based on your expectations
\#SBATCH --output=llama\_chat\_run.%j.out # Optional: name the output file to reflect the job
\#SBATCH --error=llama\_chat\_run.%j.err # Optional: name the error file to reflect the job
​
module purge;
module load anaconda3/2020.07;
export OMP\_NUM\_THREADS=$SLURM\_CPUS\_PER\_TASK;
source /share/apps/anaconda3/2020.07/etc/profile.d/conda.sh;
conda activate ./penv;
export PATH=./penv/bin:$PATH;
​
\# Use torchrun command instead of the python command
torchrun --nproc\_per\_node 1 example\_chat\_completion.py \\
\--ckpt\_dir llama-2-70b-chat/ \\
\--tokenizer\_path tokenizer.model \\
\--max\_seq\_len 512 --max\_batch\_size 6
​
​
​ | 2023-08-11T22:12:55 | https://www.reddit.com/r/LocalLLaMA/comments/15omgnq/trouble_running_llama2_70b_on_hpc_with_limited/ | MasterJaguar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15omgnq | false | null | t3_15omgnq | /r/LocalLLaMA/comments/15omgnq/trouble_running_llama2_70b_on_hpc_with_limited/ | false | false | self | 1 | null |
strange behavior with newhope.ggmlv3.q4_K_S | 1 | I've been testing the [newhope.ggmlv3.q4\_K\_S](https://huggingface.co/TheBloke/NewHope-GGML) by theBloke and it's been acting super weird and im not sure if its just poor parameters, system prompt, the 4-bit quant or the model just sucks in general?
https://preview.redd.it/u02b8s4r1khb1.png?width=1544&format=png&auto=webp&s=ab570dc78f71f75015d8be09437a3c00ba436acf
\`sys prompt\`: " You are a gifted python developer. Provide ALL your scripts in within single python markdown block. Ensure they are executable. Be efficient with compute. Maintain clear communication and a friendly demeanor. Use emojis occasionally." lol
core parameters: \` {"-c", "2048", "-ngl", "200"}
\`inference params\`:
\`\`\`python
const params = signal({
temperature: 0.7,
repeat\_last\_n: 256, *// 0 = disable penalty, -1 = context size*
repeat\_penalty: 1.18, *// 1.0 = disabled*
top\_k: 40, *// <= 0 to use vocab size*
top\_p: 0.5, *// 1.0 = disabled*
tfs\_z: 1.0, *// 1.0 = disabled*
typical\_p: 1.0, *// 1.0 = disabled*
presence\_penalty: 0.0, *// 0.0 = disabled*
frequency\_penalty: 0.0, *// 0.0 = disabled*
mirostat: 0, *// 0/1/2*
mirostat\_tau: 5, *// target entropy*
mirostat\_eta: 0.1, *// learning rate*
})
\`\`\`
​ | 2023-08-11T22:27:18 | https://www.reddit.com/r/LocalLLaMA/comments/15omu0u/strange_behavior_with_newhopeggmlv3q4_k_s/ | LyPreto | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 15omu0u | false | null | t3_15omu0u | /r/LocalLLaMA/comments/15omu0u/strange_behavior_with_newhopeggmlv3q4_k_s/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'q933B8y48VjFiOf9DmJnoMHpcG_sNy-2VRxGOgaJblE', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=108&crop=smart&auto=webp&s=1e1b58069998283803dc36c46425e88e56cf1aad', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=216&crop=smart&auto=webp&s=4a7fd6d93595a17e59a3dd39c70220d208194ab2', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=320&crop=smart&auto=webp&s=5c92b1f14dc65191a09dee27407ceb263fc26d9d', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=640&crop=smart&auto=webp&s=c7825f39480c3777d49bfafcc1cd46037abf08d9', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=960&crop=smart&auto=webp&s=0643c3d28f88e1385ce223fc02c9be1231643d39', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?width=1080&crop=smart&auto=webp&s=574b545858aa29fabbf723dff3f1b20629d71794', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/NbPMAtdGD0wzNzeW6bfXQxUamK3UR4zTtCBokSuyWyY.jpg?auto=webp&s=8155d622d2491334a597038a984aa27b45c66213', 'width': 1200}, 'variants': {}}]} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.