title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| 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
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
What is with all of the different versions and their suffixes?
| 1 |
llama-2-7b-chat.ggmlv3.q3_K_L.bin
llama-2-7b-chat.ggmlv3.q3_K_M.bin
llama-2-7b-chat.ggmlv3.q3_K_S.bin
I don't understand what the difference is between the following three files. They are all quantized to the 3bit yeh? What does the suffixes mean?
| 2023-08-29T01:35:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1644pl9/what_is_with_all_of_the_different_versions_and/
|
wh33t
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1644pl9
| false | null |
t3_1644pl9
|
/r/LocalLLaMA/comments/1644pl9/what_is_with_all_of_the_different_versions_and/
| false | false |
self
| 1 | null |
Running Phind's Fine-Tuned Model
| 1 |
Hey all,
Has anyone managed to get their fine-tune up and running? I've done my best to replicate their specs but I'm seeing a lot of garbage in my output with my implementation [here](https://github.com/emrgnt-cmplxty/zero-shot-replication/blob/main/zero_shot_replication/model/hugging_face_model/phind_model.py).
​
Does anyone see an obvious bug here?
| 2023-08-29T03:19:22 |
https://www.reddit.com/r/LocalLLaMA/comments/164707l/running_phinds_finetuned_model/
|
docsoc1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164707l
| false | null |
t3_164707l
|
/r/LocalLLaMA/comments/164707l/running_phinds_finetuned_model/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': '6mRmcVqCA-G-KDpMtQGD_ej498isxXGHFZNs5VWdah4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=108&crop=smart&auto=webp&s=5fe51a7bfbf8bc6942bd8a6af56ffa8196e05b23', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=216&crop=smart&auto=webp&s=dc8319befdce089aec521b7b2fb5d26ee692a3ed', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=320&crop=smart&auto=webp&s=3e73cef98555928cf472ecfc7e08b51a77ce4565', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=640&crop=smart&auto=webp&s=995e0fe4fb638a1fe3032d6b39a2cd120ad65cd1', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=960&crop=smart&auto=webp&s=c548385fa0e37b99416dcedc12005e5c481d29db', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=1080&crop=smart&auto=webp&s=2c234292f8990dec23e845cfb1dd7835c906a797', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?auto=webp&s=d75c64774fcd977688786a1776c42286a40b06f9', 'width': 1200}, 'variants': {}}]}
|
WizardCoder Eval Results (vs. ChatGPT and Claude on external dataset)
| 143 |
The recent Code-Llama has allowed for a number of new exciting open-source AI models, but I'm finding they still fall far short of GPT-4!.
​
After reproducing their HumanEval and assessing on \~400 OOS LeetCode problem, I see that it is more on par w/ Claude-2 or GPT-3.5. This is still a good result, but we are far from matching GPT-4 in the open-source sphere.
​
You can see [the results here](https://github.com/emrgnt-cmplxty/zero-shot-replication), and if you are interested in contributing or getting your model added, please reach out!
​
https://preview.redd.it/5a3h35jfxykb1.png?width=1976&format=png&auto=webp&s=9a007d0689c2f1802ef72dffd5f6d85798f5e318
| 2023-08-29T03:25:56 |
https://www.reddit.com/r/LocalLLaMA/comments/164754t/wizardcoder_eval_results_vs_chatgpt_and_claude_on/
|
docsoc1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164754t
| false | null |
t3_164754t
|
/r/LocalLLaMA/comments/164754t/wizardcoder_eval_results_vs_chatgpt_and_claude_on/
| false | false | 143 |
{'enabled': False, 'images': [{'id': '6mRmcVqCA-G-KDpMtQGD_ej498isxXGHFZNs5VWdah4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=108&crop=smart&auto=webp&s=5fe51a7bfbf8bc6942bd8a6af56ffa8196e05b23', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=216&crop=smart&auto=webp&s=dc8319befdce089aec521b7b2fb5d26ee692a3ed', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=320&crop=smart&auto=webp&s=3e73cef98555928cf472ecfc7e08b51a77ce4565', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=640&crop=smart&auto=webp&s=995e0fe4fb638a1fe3032d6b39a2cd120ad65cd1', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=960&crop=smart&auto=webp&s=c548385fa0e37b99416dcedc12005e5c481d29db', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?width=1080&crop=smart&auto=webp&s=2c234292f8990dec23e845cfb1dd7835c906a797', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/WRYHCQ5G7M4nCSzdHvOBssOGJRneFieU3AZMXneAXsU.jpg?auto=webp&s=d75c64774fcd977688786a1776c42286a40b06f9', 'width': 1200}, 'variants': {}}]}
|
|
Open-Source CodeLlama Server: Streaming, Caching, Model Fallbacks (OpenAI + Anthropic), Prompt-tracking
| 1 |
**TLDR;** We're open-sourcing our CodeLlama server. It handles streaming, caching, model fallbacks, and tracks prompts + token usage - [https://github.com/BerriAI/litellm-CodeGen-server](https://github.com/BerriAI/litellm-CodeGen-server)
\~\~
Hello r/LocalLLaMA,
I’m the maintainer of liteLLM() - package to simplify input/output to OpenAI, Azure, TogetherAI, Cohere, Anthropic, Baseten, Hugging face API Endpoints: [https://github.com/BerriAI/litellm/](https://github.com/BerriAI/litellm/)
We're open sourcing our CodeLlama server:
What can our server do? - It uses Together AI's CodeLlama to answer coding questions, with GPT-4 + Claude-2 as backups (you can easily switch this to any model from Huggingface, Replicate, Cohere, AI21, Azure, OpenAI, etc.)
Consistent Input/Output Format - Call all models using the OpenAI format: completion(model, messages) - Text responses will always be available at \['choices'\]\[0\]\['message'\]\['content'\]
* Streaming & Async Support - Return generators to stream text responses
* Error Handling Using Model Fallbacks (if Phind-CodeLlama fails, use Claude-2, fine-tuned GPT-3.5 etc.)
* Logging - It's integrated with promptlayer, so you can automatically track your prompt + model changes there.
* Token Usage & Spend - Track Input + Completion tokens used + Spend/model
* Caching - In-memory + Redis Cache solutions provided (works for streaming too!).
You can deploy liteLLM to your own infrastructure using Railway, GCP, AWS, Azure
Happy completion() !
| 2023-08-29T03:44:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1647isl/opensource_codellama_server_streaming_caching/
|
VideoTo
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1647isl
| false | null |
t3_1647isl
|
/r/LocalLLaMA/comments/1647isl/opensource_codellama_server_streaming_caching/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'lEgyXS6yum0oelpR9QnEuifvKtMmqw1g_wg5HVu-gxU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=108&crop=smart&auto=webp&s=5a7f799302b878ce65030864920ace4ade22cb8d', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=216&crop=smart&auto=webp&s=72193855a223da10afbbc9985af3222f2fc1ad8a', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=320&crop=smart&auto=webp&s=4bea1529dfd5ffeaef3924ca25866290dee7306c', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=640&crop=smart&auto=webp&s=b3a1b368b48c393051e33e4d0c223d707a7781c9', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=960&crop=smart&auto=webp&s=70a9068a272222835969c801bea176fad1735ce0', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?width=1080&crop=smart&auto=webp&s=1897f4aa72887ea06ef9d56d46a1366b21f45167', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/xLzfZxI5dBQxvPFprmoeelEWFp5fj-cw5VjwRAZNJN0.jpg?auto=webp&s=fe5e6f1be8e27d02f91618fe5f80e694d22b4fb5', 'width': 1200}, 'variants': {}}]}
|
Thoughts on Autonomous LLM Agents
| 1 |
Has anyone successfully deployed them in production? How reliable are they? What was your experience like?
| 2023-08-29T04:45:11 |
https://www.reddit.com/r/LocalLLaMA/comments/1648pav/thoughts_on_autonomous_llm_agents/
|
XhoniShollaj
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1648pav
| false | null |
t3_1648pav
|
/r/LocalLLaMA/comments/1648pav/thoughts_on_autonomous_llm_agents/
| false | false |
self
| 1 | null |
What is the relationship between context size, tokens and models?
| 1 |
So is every word generated or parsed a "token"? And is contextsize the number of tokens an LLM can "remember" and is context size somehow limited by the models construction?
Also curious if RoPE somehow plays into all of this.
| 2023-08-29T05:39:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1649oyd/what_is_the_relationship_between_context_size/
|
wh33t
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1649oyd
| false | null |
t3_1649oyd
|
/r/LocalLLaMA/comments/1649oyd/what_is_the_relationship_between_context_size/
| false | false |
self
| 1 | null |
Quantize model to 4-bit GPTQ with 6GB VRAM?
| 1 |
Is it possible to quantanize a model into 4bit GPTQ?
I've tried GPTQ-for-LLaMa and pyllama, but both gives me CUDA out of memory error. Even pyllama, which says it works to quantianize with 4GB GPU, it does not work for me.
This is the model I'm trying to make into 4-bit GPTQ: mplug-owl-llama-7b-ft
| 2023-08-29T06:11:25 |
https://www.reddit.com/r/LocalLLaMA/comments/164a97a/quantize_model_to_4bit_gptq_with_6gb_vram/
|
fetballe
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164a97a
| false | null |
t3_164a97a
|
/r/LocalLLaMA/comments/164a97a/quantize_model_to_4bit_gptq_with_6gb_vram/
| false | false |
self
| 1 | null |
Fine-tune and then Quantize vs. Quantize and fine-tune
| 1 |
We are trying to understand whether it is advisable -
to take Meta's llama2, fine-tune them using custom datasets, and then 4-bit quantize them
OR
to take an already 4-bit quantized llama2 model (e.g, TheBloke) and then fine-tune them. What approach would you recommend and why?
| 2023-08-29T06:38:47 |
https://www.reddit.com/r/LocalLLaMA/comments/164aqbu/finetune_and_then_quantize_vs_quantize_and/
|
sbs1799
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164aqbu
| false | null |
t3_164aqbu
|
/r/LocalLLaMA/comments/164aqbu/finetune_and_then_quantize_vs_quantize_and/
| false | false |
self
| 1 | null |
Why did the AI just say "Yes"?
| 1 |
Using Kobold and codellama-13b.ggmlv3.Q2\_K.bin
​
https://preview.redd.it/7phldknf20lb1.png?width=1144&format=png&auto=webp&s=8b6268becaaca2ab2dae6f83af4bf40a55b08e40
| 2023-08-29T07:17:42 |
https://www.reddit.com/r/LocalLLaMA/comments/164bekn/why_did_the_ai_just_say_yes/
|
McpeIsSoBuggy
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164bekn
| false | null |
t3_164bekn
|
/r/LocalLLaMA/comments/164bekn/why_did_the_ai_just_say_yes/
| false | false | 1 | null |
|
Alignment kills performance
| 1 | 2023-08-29T07:30:15 |
https://arxiv.org/pdf/2308.13449.pdf
|
DotaMate
|
arxiv.org
| 1970-01-01T00:00:00 | 0 |
{}
|
164bm39
| false | null |
t3_164bm39
|
/r/LocalLLaMA/comments/164bm39/alignment_kills_performance/
| false | false |
default
| 1 | null |
|
Can someone recommend an uncensored model which can be trained on raw text?
| 1 |
I have used quite a few models such as:
Chronos-hermes-13b
Nous-hermes-Llama2
Wizard-Vicuna-13
etc...
I'm trying to train off a small selection of raw text files, but all of the models which I have used cough up some error or another. Typical errors with Chronos-hermes for example would be:
TypeError: a bytes-like object is required, not 'str'
| 2023-08-29T08:16:03 |
https://www.reddit.com/r/LocalLLaMA/comments/164cdy7/can_someone_recommend_an_uncensored_model_which/
|
jasondray_1998
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164cdy7
| false | null |
t3_164cdy7
|
/r/LocalLLaMA/comments/164cdy7/can_someone_recommend_an_uncensored_model_which/
| false | false |
self
| 1 | null |
Is there already a way to use Llama 2 with a very big system prompt?
| 1 |
[removed]
| 2023-08-29T08:18:02 |
https://www.reddit.com/r/LocalLLaMA/comments/164cf6h/is_there_already_a_way_to_use_llama_2_with_a_very/
|
Botanical0149
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164cf6h
| false | null |
t3_164cf6h
|
/r/LocalLLaMA/comments/164cf6h/is_there_already_a_way_to_use_llama_2_with_a_very/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'x9KSkL1w8g7ONAVbHh791t7xxoK_zsW3FznzlcaLsbg', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=108&crop=smart&auto=webp&s=ae4f2942588f8bfcfdff7621b1a9913f98cef85e', 'width': 108}, {'height': 144, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=216&crop=smart&auto=webp&s=9fbfc55078b859d084a0ae94e6400521c0bfd6d6', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=320&crop=smart&auto=webp&s=52bc824ec72db45f66bbf7025782409fbdfe9eb4', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=640&crop=smart&auto=webp&s=81c9e1edbb5fcbb3bbbeeb4db2e76ff94bb993c8', 'width': 640}, {'height': 640, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=960&crop=smart&auto=webp&s=6bbd1d8d1308507800a2e52a311a682a2bf84880', 'width': 960}, {'height': 720, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?width=1080&crop=smart&auto=webp&s=7332202b27e1037f2fbf271c4b1e5da6c0c58216', 'width': 1080}], 'source': {'height': 1000, 'url': 'https://external-preview.redd.it/volVXV5NmcZHt6VD_ATnoeGpqc6lJT_rqQFMmb_Jk-k.jpg?auto=webp&s=eea146fa0abe7adcad3f3635d3d222ad5b7dfd01', 'width': 1500}, 'variants': {}}]}
|
How to run (or convert) PyTorch model with llama.cpp?
| 1 |
I recently can get a model running in the local (Wizard-Vicuna-7B-Uncensored.ggmlv3.q4\_0) with llama.cpp and rustformers/llm.
Is it possible to run pytorch model (e.g [Wizard-Vicuna-7B-Uncensored](https://huggingface.co/meta-llama/Llama-2-7b)) with llama.cpp? Or does anyone know how to convert pytorch model to ggml format?
https://preview.redd.it/r2c7srxfl0lb1.png?width=366&format=png&auto=webp&s=680ec2229fc4721fcd00da2b5ebefb526504b1ad
| 2023-08-29T09:10:41 |
https://www.reddit.com/r/LocalLLaMA/comments/164dbip/how_to_run_or_convert_pytorch_model_with_llamacpp/
|
ougizee
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164dbip
| false | null |
t3_164dbip
|
/r/LocalLLaMA/comments/164dbip/how_to_run_or_convert_pytorch_model_with_llamacpp/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'cQf8VWaf-xGehCpVJfC5PBPqy_6SPP52BQSegGE5rTU', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=108&crop=smart&auto=webp&s=fc98606e6096c95830db3cddf20b412f75e8c8b9', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=216&crop=smart&auto=webp&s=f618bcdc45db1d2146c57e46e551a0e1f1a8b7e0', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=320&crop=smart&auto=webp&s=3ffcb9e5f9327da68c48bf7048d717edd1f08e28', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=640&crop=smart&auto=webp&s=55699385eee145c4aaf85f58252b658dbede0a3b', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=960&crop=smart&auto=webp&s=09253789f00933d811bc0d3b7df9c87638761006', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?width=1080&crop=smart&auto=webp&s=5a4633945f621e336e2730202509ad79f08ea604', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/XCakRieLhCzqSiOj_2cszqVz27MK2dN8NVJRVQ5Zq6s.jpg?auto=webp&s=64e03f113314b24a094b0a9c903c7b77252ac7f8', 'width': 1200}, 'variants': {}}]}
|
|
Metric to compare performance of inference optimization methods
| 1 |
Hello community! I'm taking lots of information and help from LocalLLaMA!
Always appreciate your knowledge :)
I have one question, what is the best suitable (or most used metric) when comparing inference speed of optimization methods?
I'm doing research which optimization method to use, for example vLLM, flash attention, etc.. In doing this, I should select some metrics by which methods are compared.
After searching, I found that 1) transactions per second 2) tokens per second 3) throughput are used commonly. At first glance, I think that transactions per second is not a suitable metric because llm generates different lenghts according to different prompt, hence different output time is measured.
Among them, which metric would you think that best suitable for describing performance of inference method? Or is there any other metrics that I do not know?
Thank you very much! :)
| 2023-08-29T09:58:08 |
https://www.reddit.com/r/LocalLLaMA/comments/164e4vj/metric_to_compare_performance_of_inference/
|
Fluffy_Pianist_3135
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164e4vj
| false | null |
t3_164e4vj
|
/r/LocalLLaMA/comments/164e4vj/metric_to_compare_performance_of_inference/
| false | false |
self
| 1 | null |
Prompt Engineering Seems Like Guesswork - How To Evaluate LLM Application Properly?
| 1 |
How are folks evaluating the quality of your LLM applications? I'm running a mental health chatbot in production (small scale - 10's of active users) and I've spent a lot of time finetuning prompts but it's all just guesswork.
I'll make a tweak to the prompt and run a few test conversations and just kinda get the vibes of whether it's better or worse than before the tweak. Is this what y'all are doing too or am I missing something???
| 2023-08-29T10:48:49 |
https://www.reddit.com/r/LocalLLaMA/comments/164f2sq/prompt_engineering_seems_like_guesswork_how_to/
|
foazi
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164f2sq
| false | null |
t3_164f2sq
|
/r/LocalLLaMA/comments/164f2sq/prompt_engineering_seems_like_guesswork_how_to/
| false | false |
self
| 1 | null |
LLM benchmarks are broken, what can we do to fix them?
| 1 |
Like many others here, I've found that the benchmarks that people are using are not representative to real world usability. Models that are at the top of the leaderboard fail to do simple tasks and have very high error rates when tested at scale. I've tested about 10 highly ranked models using 10k news articles and find in QA testing that they have as high as 60% error rates.. there is plenty of speculation why and I'm sure there is various causes.
Would love to hear what ideas this community can come up with to replace these academic tests which clearly aren't good for any of us (buildersor users).
Before you try to nitpick my methodology or give your anecdotal defense of your favorite model.. I'm a senior SME who works for a tech giant and I work with many of the leading startups in this space. So no need for debate, I know how to evaluate at scale.
| 2023-08-29T11:11:36 |
https://www.reddit.com/r/LocalLLaMA/comments/164fivc/llm_benchmarks_are_broken_what_can_we_do_to_fix/
|
Tiny_Arugula_5648
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164fivc
| false | null |
t3_164fivc
|
/r/LocalLLaMA/comments/164fivc/llm_benchmarks_are_broken_what_can_we_do_to_fix/
| false | false |
self
| 1 | null |
Domain expert llama
| 1 |
i am trying to use llama 2 as a chatbot alongside a vector store in a way that processes pdf and web pages and responds according to the similarity of the prompt with stored chunks. is anyone doing the same?
i have some questions:
1) which model to use? at the moment i am using the Meta chat version model, are there better fine tuned templates?
2) i'm using llama-cpp-python and it seems that the embedder has poor performance, am i wrong about chuck size or is it better to use 3rd party embedders? What I see is basically a low score with the chunks, I would like to have scores >= 75%.
| 2023-08-29T11:11:53 |
https://www.reddit.com/r/LocalLLaMA/comments/164fj25/domain_expert_llama/
|
SK33LA
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164fj25
| false | null |
t3_164fj25
|
/r/LocalLLaMA/comments/164fj25/domain_expert_llama/
| false | false |
self
| 1 | null |
Is parallel inference possible for llama2 with a single Nvidia GPU?
| 1 |
We are using llama2 on a single Nvidia GPU for inferring information based on input texts which are stored in a data frame. Currently we are going sequentially through the data frame.
We were wondering if there is some way we can send multiple prompts to the model and store the responses. Is this possible at all?
| 2023-08-29T11:25:12 |
https://www.reddit.com/r/LocalLLaMA/comments/164fskh/is_parallel_inference_possible_for_llama2_with_a/
|
sbs1799
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164fskh
| false | null |
t3_164fskh
|
/r/LocalLLaMA/comments/164fskh/is_parallel_inference_possible_for_llama2_with_a/
| false | false |
self
| 1 | null |
Generate Q&A from text
| 1 |
[removed]
| 2023-08-29T13:19:10 |
https://www.reddit.com/r/LocalLLaMA/comments/164ib09/generate_qa_from_text/
|
Enkay55
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164ib09
| false | null |
t3_164ib09
|
/r/LocalLLaMA/comments/164ib09/generate_qa_from_text/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'UaW090vbWKsuGEBeEQ5zksrl2gy7ygk3iN3k7NcjNRo', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/KTL8DUGaG6D46g-6-o-rbbw8kPKBCjsurSWrWW114ys.jpg?width=108&crop=smart&auto=webp&s=612630af980cc63ccd1035f378b4766ec5f51f7e', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/KTL8DUGaG6D46g-6-o-rbbw8kPKBCjsurSWrWW114ys.jpg?width=216&crop=smart&auto=webp&s=ecf9b8cd0b7a33f32ae2ed01e129e39c6743a0c9', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/KTL8DUGaG6D46g-6-o-rbbw8kPKBCjsurSWrWW114ys.jpg?width=320&crop=smart&auto=webp&s=ef3ef716cf4cccccb43c100051a55d65c282524e', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/KTL8DUGaG6D46g-6-o-rbbw8kPKBCjsurSWrWW114ys.jpg?auto=webp&s=c93894bbb40de18d8af6e5f4833f06ff8a4cba41', 'width': 480}, 'variants': {}}]}
|
Airoboros-l2-70b-2.1's uniquely high TruthfulQA causes it to become 1st on leaderboard
| 1 |
airoboros-l2-70b-2.1, an experimental version of airoboros, has just been released and evaluated. It has good performance in ARC, HellaSwag, and MMLU, but its very high TruthfulQA has pushed it to 1st on the leaderboard, and with a decent distance between it and 2nd. The other overall top ranked models have TruthfulQA scores of around 61-63 while airoboros-l2-70b-2.1 gets a 70!
Its huggingface says it uses an "RP style instruction set" and supports "longer, more detailed writing prompts" as well as having multiple other experimental changes.
For people that can test the model, I am curious if the large increase in TruthfulQA has any noticeable change in feel when comparing it to other top models.
Note that there are two airoboros-l2-70b-2.1 models on the leaderboard linking to the same model. The first place one has the model commit sha 30605bd7679a74898c238f9b011125a205079cd1
[https://huggingface.co/spaces/HuggingFaceH4/open\_llm\_leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
[https://huggingface.co/jondurbin/airoboros-l2-70b-2.1](https://huggingface.co/jondurbin/airoboros-l2-70b-2.1)
[https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GGML](https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GGML)
[https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GGUF](https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GGUF)
[https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GPTQ](https://huggingface.co/TheBloke/Airoboros-L2-70B-2.1-GPTQ)
| 2023-08-29T13:40:06 |
https://www.reddit.com/r/LocalLLaMA/comments/164itew/airoborosl270b21s_uniquely_high_truthfulqa_causes/
|
DontPlanToEnd
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164itew
| false | null |
t3_164itew
|
/r/LocalLLaMA/comments/164itew/airoborosl270b21s_uniquely_high_truthfulqa_causes/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'EN0-abblERL52DxeoNzcxdkhvXEwLdZMJTS58Umjs64', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=108&crop=smart&auto=webp&s=6fbb309f983333cbaf528bd40f8d6ffb39877704', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=216&crop=smart&auto=webp&s=1ae10c5a53638209dee07b017628d2a1fadc8d05', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=320&crop=smart&auto=webp&s=cf36565d3bac3086aaea4458c31609ff1b2c00b3', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=640&crop=smart&auto=webp&s=8e182cefcf8da97d7b4369734149986feca334e5', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=960&crop=smart&auto=webp&s=7699d0ad09185e2f560115cae5cb71e907073327', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?width=1080&crop=smart&auto=webp&s=7b11f6f2294899964ec8ed081777f4b6e19723b6', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/pnEIDVgN3O3UZSEZF8G101Cpm5FLu9i3k_abBlep_2c.jpg?auto=webp&s=81db4d9e1dd01a76f499e499f78aed3478ae6658', 'width': 1200}, 'variants': {}}]}
|
PyCharm Plugin
| 1 |
For various reasons (company policy), I can't use external code assistants but I could use a locally running Code Llama. I can't find a plugin that supports it though (I did see the one for VS Code, though). Does one exist? If I need to write one, what should I use as an example?
Code Llama: [https://about.fb.com/news/2023/08/code-llama-ai-for-coding/](https://about.fb.com/news/2023/08/code-llama-ai-for-coding/)
| 2023-08-29T13:46:34 |
https://www.reddit.com/r/LocalLLaMA/comments/164iz75/pycharm_plugin/
|
beezlebub33
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164iz75
| false | null |
t3_164iz75
|
/r/LocalLLaMA/comments/164iz75/pycharm_plugin/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': '_-Buf_9EsYsfZBPyO83DU0c6n6Bfz_7qOxm-a3CNSig', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=108&crop=smart&auto=webp&s=88ce6cfa97c258adf200b30857dd13240d3faefb', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=216&crop=smart&auto=webp&s=2ddb5b4929135e581c26fef18023563547bc62c4', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=320&crop=smart&auto=webp&s=a63f0886b29c9c8c3bee0f678ed2c800c314bc79', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=640&crop=smart&auto=webp&s=7b121767859e42fc99088f9fe0158b1c34ceac17', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=960&crop=smart&auto=webp&s=21e5047692ba6e2c4a0f40b32501fc3717ac237d', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?width=1080&crop=smart&auto=webp&s=ac0421070084b45c0832da3221a946b8f6340a68', 'width': 1080}], 'source': {'height': 675, 'url': 'https://external-preview.redd.it/ktm8L0FwgIrST_6p16OmYr1-tQ_Dsaawj3YmvUb3SaE.jpg?auto=webp&s=0d2a5dc16f22d5bf4a071c8ded73c71a4b594887', 'width': 1200}, 'variants': {}}]}
|
Examples of actual fine tunings with data?
| 1 |
Are there any example where the full data and training process for fine tuning (specifically parameter efficient fine tuning ideally, but any sort would be ok) in addition to having the model itself after fine tuning available to use and chat with?
I really don't care so much what sort of task the fine tuning was about, I'm more interested in seeing what their goal was, and how they structured the training data and other inputs to reach that goal, and how it turned out.
I've seen plenty of "how tos" about tuning models, where they let you see the data, but those often seem "thrown together"- not much effort put into the fine tuning itself because their goal was to teach people the basics of how fine tuning works rather than actually create a highly effective model.
In other words, I'm looking for cases where it's well beyond a "Hello World" approach to fine tuning, rather, someone actually put effort into making the thing turn out good (whatever "good" meant for them) and they shared the actual training examples as well.
| 2023-08-29T13:54:01 |
https://www.reddit.com/r/LocalLLaMA/comments/164j5cp/examples_of_actual_fine_tunings_with_data/
|
ascendant23
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164j5cp
| false | null |
t3_164j5cp
|
/r/LocalLLaMA/comments/164j5cp/examples_of_actual_fine_tunings_with_data/
| false | false |
self
| 1 | null |
CodeLlama-7b-Instruct-hf via TGI vs meta's Codellama-7b-Instruct
| 1 |
Ahoi,
I'm very new to the local LLM thing but got quite excited when hearing about CodeLlama and (at least the 7b version) being able to run on a single GPU. So I got codellama-7b-instruct from meta [here](https://ai.meta.com/resources/models-and-libraries/llama-downloads/). To me it seems to work \_really\_ well with the default parameters they set.
Example:
I ask as role "user":
"write a python function to integrate x**2 from x_min to x_max"
and it answers:
"To integrate x**2 from x_min to x_max in Python, you can use the `scipy.integrate` module. Here's an example of how to do this:
```
from scipy.integrate import quad
def integrate_x2(x_min, x_max):
return quad(lambda x: x**2, x_min, x_max)[0]
```
This function uses the `quad` function from the `scipy.integrate` module to integrate the function x**2 from x_min to x_max. The `quad` function returns a tuple containing the integral and an error estimate. We only need the integral, so we take the first element of the tuple using square brackets.
You can then call this function with the desired limits, like this:
```
print(integrate_x2(0, 1))
```
This will print the value of the integral of x**2 from 0 to 1.
Note that this function assumes that the function being integrated is continuous and smooth over the entire interval. If the function has discontinuities or sharp gradients, you may need to use a different integration method or modify the function to make it more integrable."
Now I read about quantization, sharding and stuff like that, that might (tbh I'm not sure yet) make it possible to run this on older GPUs with less VRAM (currently using a 4090, but have older ones lying around). Read about hugging face and that the community gathers all the finetuned and in other ways customized models there. Read about [text-generation-inference](https://github.com/huggingface/text-generation-inference) (TGI) which comes as a docker container that makes it easy to run all kinds of models from huggingface as a server with a rest endpoint. It also seems to be able to use quantization and whatnot simply via starting with corresponding cmdline arguments.
So I tried out TGI with codellama-7b-instruct-hf
docker run --gpus all --shm-size 1g -p 8080:80 -v $(pwd -P):/data ghcr.io/huggingface/text-generation-inference:1.0.3 --model-id codellama/CodeLlama-7b-Instruct-hf
and ask it the same
curl 127.0.0.1:8080/generate -X POST -H 'Content-Type: application/json' -d '{"inputs":"write a python function to integrate x**2 from x_min to x_max", "parameters": {"max_new_tokens": 200}'
(or with
"parameters": {"max_new_tokens": 200, "temperature": 0.2, "top_p": 0.95}
to reproduce some of meta's default parameters) and it gives me:
{"generated_text":"\n\n\\begin{code}\ndef integrate(x_min, x_max):\n return (x_max**2 - x_min**2)/2\n\\end{code}\n\nComment: What is the question?\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: What is the problem with your code?\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment: I want to integrate x**2 from x_min to x_max\n\nComment:"}
which is subpar...
So to my questions:
* Am I missing something obvious to you?
* What do I have to do to make the huggingface model in the TGI container return results similar to the "vanilla" codellama?
* Is codellama-hf retrained somehow or in other ways substantially different from meta's codellama?
* I only found the parameters temperature and top\_p in both variants. Do they even mean the same?
​
| 2023-08-29T13:59:38 |
https://www.reddit.com/r/LocalLLaMA/comments/164ja5p/codellama7binstructhf_via_tgi_vs_metas/
|
zernichtet
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164ja5p
| false | null |
t3_164ja5p
|
/r/LocalLLaMA/comments/164ja5p/codellama7binstructhf_via_tgi_vs_metas/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'ilC2qprzEOhvondbER2GPm9DXBMFQhdj6lShAI3fqUQ', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=108&crop=smart&auto=webp&s=b96f0fb64d0fd3022dd85d7522591d32ffa3e30e', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=216&crop=smart&auto=webp&s=9912a2752494571ed70d5a86ac12b82605c4f45c', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=320&crop=smart&auto=webp&s=56ed0063c62caf22cd7da6c252e1217e3110c1b7', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=640&crop=smart&auto=webp&s=de6bc123c3d7a92ad1b5d7d6155a79bbbf60123f', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=960&crop=smart&auto=webp&s=e0c2d0341b3c852b53903f8db3781047c285ed18', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?width=1080&crop=smart&auto=webp&s=7aa7b2985c05b52eff9a4cdcefefafca8c3ba9c7', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/-lx7IoVnPKtS1s2Rq8IcxH6q6WBMlXXfBQF43Q3okcU.jpg?auto=webp&s=188e3053d99818d509c6f9549c04cc4f13e6981a', 'width': 1920}, 'variants': {}}]}
|
Looking to sell a couple ML servers EU
| 6 |
So I have been tasked with finding a buyer for a couple high end machine learning servers. They were owned by my wife’s father who passed recently.
The servers are powered by a couple Epyc 7003s and have A series gpus. We have invoices for them and VAT has been paid on everything.
Hopefully it’s ok to post this here. Also feel free to PM .
| 2023-08-29T14:05:57 |
https://www.reddit.com/r/LocalLLaMA/comments/164jftz/looking_to_sell_a_couple_ml_servers_eu/
|
Obnomad
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164jftz
| true | null |
t3_164jftz
|
/r/LocalLLaMA/comments/164jftz/looking_to_sell_a_couple_ml_servers_eu/
| false | false |
default
| 6 | null |
Is there a way to weight tokens in context?
| 1 |
In stable diffusion, you can weight up or down specific tokens - is there a way to do something similar with Llama models using oobabooga? Or is there a way to add probabilities, like be "happy" 70% of the time but "sad" 30% of the time? Thanks.
| 2023-08-29T14:26:21 |
https://www.reddit.com/r/LocalLLaMA/comments/164jxk3/is_there_a_way_to_weight_tokens_in_context/
|
emptyplate
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164jxk3
| false | null |
t3_164jxk3
|
/r/LocalLLaMA/comments/164jxk3/is_there_a_way_to_weight_tokens_in_context/
| false | false |
self
| 1 | null |
Local Models missing/skipping words in response.
| 1 |
[removed]
| 2023-08-29T15:44:24 |
https://www.reddit.com/r/LocalLLaMA/comments/164lyi1/local_models_missingskipping_words_in_response/
|
halphrican
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164lyi1
| false | null |
t3_164lyi1
|
/r/LocalLLaMA/comments/164lyi1/local_models_missingskipping_words_in_response/
| false | false |
self
| 1 | null |
Small model for text classification
| 1 |
Hi everyone,
I need a small model which is specialized to classify text. It takes the text and map it to an intent from my intent list.
For example, here is the list of my intent:
- Register: User wants to register.
- Order: User wants to purchase.
My expected result is when the model receives a text such as: “Hey, I want to register.” ; The model will response “The intent is Register.”
My intent list is having only 10 items. If the message is unable to map to an appropriate intent, just response “I can not support.”.
I am thinking of using Llama might be a bit overkill for this task. Hence, I am seeking for a lightweight model to do so.
Any recommendation is much appreciated. Thank a lot.
| 2023-08-29T16:02:07 |
https://www.reddit.com/r/LocalLLaMA/comments/164mevq/small_model_for_text_classification/
|
jThaiLB
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164mevq
| false | null |
t3_164mevq
|
/r/LocalLLaMA/comments/164mevq/small_model_for_text_classification/
| false | false |
self
| 1 | null |
TelegramBot <-> LLM-API like KoboldCpp
| 1 |
Hi!
Anybody know of a TelegramBot that I can hook up to the KoboldCpp-API and chat with my local running LLM? I heard there aren't any yet.
| 2023-08-29T17:15:45 |
https://www.reddit.com/r/LocalLLaMA/comments/164ocic/telegrambot_llmapi_like_koboldcpp/
|
honuvo
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164ocic
| false | null |
t3_164ocic
|
/r/LocalLLaMA/comments/164ocic/telegrambot_llmapi_like_koboldcpp/
| false | false |
self
| 1 | null |
How would you fine-tune for my use case?
| 1 |
Hi all. I have a scoring system against a set of environmental criteria and after I apply this to a company, by looking at their annual report/sustainability disclosure. Normally, this is a manual effort which involves a lot of ctrl + f and going through PDFs to source information. So, I have implemented RAG against GPT-4 and using Retrieval QA to ask basic questions which works and definitely helps speed up the process of finding information, but I'd love to be able to make the whole end-to-end process automated so an LLM can score a client.
I outline an example of the scoring criteria below:
​
Category: Green spending
Score 1 - The client has quantified plans to spend money on reducing their emissions in the next two years.
Score 2 - The client has plans to spend money on reducing emissions but has not quantified this.
Score 3 - The client has no plans to spend money to reduce emissions.
​
So how would I put this in an Instruct Tune format I use for a Llama model?
​
Instruction:
Score the client from 1-3 based on the following criteria.
Score 1 - The client has quantified plans to spend money on reducing their emissions in the next two years.
Score 2 - The client has plans to spend money on reducing emissions but has not quantified this.
Score 3 - The client has no plans to spend money to reduce emissions.
​
Output:
Insert model answer here for one of the scores.
​
Repeat this so each score has a model answer across the categories.
​
There are about 15 categories, and I could generate 1,000-2,000 examples across the various combinations of score and category.
​
My plan would then be to use this fine-tuned model in RAG, and ask it to score British Petroleum on Green Spending by accessing the documents.
​
1. Does this sound a sound strategy?
2. Is there a limit to the tokens that can be in an instruction?
​
Thanks!
| 2023-08-29T17:15:58 |
https://www.reddit.com/r/LocalLLaMA/comments/164ocpj/how_would_you_finetune_for_my_use_case/
|
99OG121314
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164ocpj
| false | null |
t3_164ocpj
|
/r/LocalLLaMA/comments/164ocpj/how_would_you_finetune_for_my_use_case/
| false | false |
self
| 1 | null |
Meta Code Llama – Python, further fine-tuned on 100B tokens of Python code
| 1 |
[removed]
| 2023-08-29T17:31:39 |
https://www.reddit.com/r/LocalLLaMA/comments/164orap/meta_code_llama_python_further_finetuned_on_100b/
|
w7gg33h
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164orap
| false | null |
t3_164orap
|
/r/LocalLLaMA/comments/164orap/meta_code_llama_python_further_finetuned_on_100b/
| false | false |
self
| 1 | null |
Lenovo T480s
| 1 |
Good day fellow Redditors,
I am in need for **organic** feedback.
I am starting next month a CS degree and have decided to buy a Lenovo T480s with these Specs
​
​
|Hard Disk |1TB|
|:-|:-|
|OS |Win 11 pro|
|Processor |i7 8th gen |
|Processing Speed |1,8 GHZ|
|RAM |16GB|
​
My train of thought buying this machine :
​
* Open-architecture so I can learn something about Hardware in general
* Cheap
* I can learn linux with it
​
So my question is how well will it perform using LLM ? (I am going to try myself , but it will arrive later this week...)
​
So any input is appreciated.
​
​
| 2023-08-29T17:36:59 |
https://www.reddit.com/r/LocalLLaMA/comments/164owcd/lenovo_t480s/
|
hustla17
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164owcd
| false | null |
t3_164owcd
|
/r/LocalLLaMA/comments/164owcd/lenovo_t480s/
| false | false |
self
| 1 | null |
SillyTavern 1.10.0 Release
| 1 |
## SillyTavern 1.10.0 has been released.
---
Due to the scope of changes, it is highly recommended to **backup your files before updating**.
---
**Highlights**
- Prompt Manager for Chat Completions
- Advanced Formatting for Text APIs
- Dynamic Audio extension
- RVC and Coqui TTS support
- Simplified UI mode
**Other Improvements**
- Preset management for Context templates and Instruct templates
- OpenRouter prompt cost calculations
- Support for Markdown tables
- Renamed Live2D extension to TalkingHead
- Proxy passwords hidden by default
- More NovelAI settings
- Chat Lazy Loading
- AI21 API support
- Per-chat CFG support
- HotKey: Escape key to close panels and popups.
- API Icons next to Timestamp
- Performance improvements and pagination for character list, groups, and world info entries
- Fuzzy search for characters and groups
- Improvements to NovelAI API: logit bias, samplers order, banned tokens, etc.
- Manual UI language selector and new UI languages: Dutch, Italian, and Russian
- Chat Completion source is shown on timestamp hover
- More stable file saving to prevent accidental chat deletion during a PC crash
- New StableDiffusion option to render a background based on chat
- Add a button to hide the upper portion of the Character panel
- Console window output coloring
- Search for past chat via content keywords
- Auto-clean the Uploads folder
- Individual Swipes can now be deleted
- Dialogue examples can be removed from the prompt entirely via toggle
- Favorited characters stand out more in the character list
- Token counter for each box in Character Panel, and Persona Description
- Alternative 'Cookie method' for Scale API
- Bottom and top bars now resize based on the Main Font Size
- Fix for accidental slider adjustment on touch devices (300ms delay before activating)
- Quick 'Continue' button in the chat bar
- Add support for OpenRouter fallback models
- Fix bug to preserve Swipes that were Continued upon
- LibreTranslate added as an auto-translate source
- Improvements for Instruct mode handling and panel UI
---
https://github.com/SillyTavern/SillyTavern/releases/tag/1.10.0
How to update: https://docs.sillytavern.app/usage/update/
| 2023-08-29T18:24:03 |
https://www.reddit.com/r/LocalLLaMA/comments/164q3xq/sillytavern_1100_release/
|
RossAscends
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164q3xq
| false | null |
t3_164q3xq
|
/r/LocalLLaMA/comments/164q3xq/sillytavern_1100_release/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'OjR083SAmXjUzkWDy5HXkqnu4zCUisIFX8zzBCyxUdY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=108&crop=smart&auto=webp&s=0c6041247442046e9790c4d48902c06f46c2a33c', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=216&crop=smart&auto=webp&s=1d301dece2d34ba9073777f7bab391d69a84b0c3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=320&crop=smart&auto=webp&s=4c209ab5724f8774a2502bc0506bf22cc9dd6301', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=640&crop=smart&auto=webp&s=30c41f364d896866a3501fd24e526e3fb5c359ce', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=960&crop=smart&auto=webp&s=6cc1860d371d6d15b373d4584859b251a17b3cc9', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?width=1080&crop=smart&auto=webp&s=63eedb629223e57b53edc7e781f2fefbe2578306', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/8kajQwDFCrWGBERO6m8PDuAdYCunxlahQM42cLwkeaQ.jpg?auto=webp&s=2d78102e5b34468d994e3902e02c014b7d4078dc', 'width': 1200}, 'variants': {}}]}
|
Codelamma 7b code completion giving multiple responses and i only want one?
| 1 |
I am using Codelamma 7b code completion model for the purposes of generating sql statements. I get a return of a sequence with multiple answers. the first answer in the sequence seems to be the answer i am looking for, however, the model then continues to ramble on.
My prompt is formatted so that there is a schema context as create table statements, the user question, then the beginning of a select statement(SELECT).
Any thoughts on how to limit the response's content to the first answer in the sequence? the prompt seems to follow the instructions from \[codelammas github page\]([https://github.com/facebookresearch/codellama/tree/main](https://github.com/facebookresearch/codellama/tree/main))
My thoughts are either the prompt format is off or im doing something foolish. Likely the latter.
Thanks everyone!!
from transformers import AutoTokenizer
import transformers
import torch
model = "codellama/CodeLlama-34b-hf"
tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
"text-generation",
model=model,
torch_dtype=torch.float16,
device_map="auto",
)
sequences = pipeline(
'import socket\n\ndef ping_exponential_backoff(host: str):',
do_sample=True,
top_k=10,
temperature=0.1,
top_p=0.95,
num_return_sequences=1,
eos_token_id=tokenizer.eos_token_id,
max_length=200,
)
for seq in sequences:
print(f"Result: {seq['generated_text']}")
​
| 2023-08-29T19:20:23 |
https://www.reddit.com/r/LocalLLaMA/comments/164rl7k/codelamma_7b_code_completion_giving_multiple/
|
llamasaresavager
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164rl7k
| false | null |
t3_164rl7k
|
/r/LocalLLaMA/comments/164rl7k/codelamma_7b_code_completion_giving_multiple/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'XmAy706JfCO0EDPPUMDIyLi-CUvkdIrG6M6l2hynU68', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=108&crop=smart&auto=webp&s=0c593be0ec2fcc008030e6e5cd47da44f5d392d0', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=216&crop=smart&auto=webp&s=8d906fd0fe8ec8c0f976bf237e892fef0df617c7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=320&crop=smart&auto=webp&s=d3e18cc89c07ba3b2d073f10caa91091d1720a50', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=640&crop=smart&auto=webp&s=692eedbc71a97c21010f40362274b92781052a10', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=960&crop=smart&auto=webp&s=8e00c337a467d17da781b8d135ec64c38fc5b864', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?width=1080&crop=smart&auto=webp&s=f7a4458dbde53ad876e8de9a2aba493573dba453', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/5pHknhugjJRv7NRqjx8tgTnJUIcNmknWGT_51USkvX4.jpg?auto=webp&s=2c9aa7b6b2db90679b7294cc01827d5a5637e169', 'width': 1200}, 'variants': {}}]}
|
What could you say about Qwen-VL by Alibaba?
| 1 |
[removed]
| 2023-08-29T19:47:42 |
https://www.reddit.com/r/LocalLLaMA/comments/164sbtg/what_could_you_say_about_qwenvl_by_alibaba/
|
Ronrel
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164sbtg
| false | null |
t3_164sbtg
|
/r/LocalLLaMA/comments/164sbtg/what_could_you_say_about_qwenvl_by_alibaba/
| false | false | 1 |
{'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': {}}]}
|
|
How to fine tune an LLM locally using C#
| 1 |
Hi, I'm new to LLMs and trying to understand how to fine-tune an LLM locally.
However, it seems that almost all tutorials are about fine-tuning using Python as the primary language.
Has anyone managed to fine tune an LLM using C#?
| 2023-08-29T20:12:21 |
https://www.reddit.com/r/LocalLLaMA/comments/164szf4/how_to_fine_tune_an_llm_locally_using_c/
|
Rias0
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164szf4
| false | null |
t3_164szf4
|
/r/LocalLLaMA/comments/164szf4/how_to_fine_tune_an_llm_locally_using_c/
| false | false |
self
| 1 | null |
Was anyone able to fine tune the 70B llama2 model with lora?
| 1 |
I'm not able to get 70B llama2 fine tuning to work. Does anyone have a working version of fine-tuning and could share the setup they used?
| 2023-08-29T20:57:52 |
https://www.reddit.com/r/LocalLLaMA/comments/164u6xr/was_anyone_able_to_fine_tune_the_70b_llama2_model/
|
DeanBlub
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164u6xr
| false | null |
t3_164u6xr
|
/r/LocalLLaMA/comments/164u6xr/was_anyone_able_to_fine_tune_the_70b_llama2_model/
| false | false |
self
| 1 | null |
If this is true, how is llama open source?
| 1 |
​
https://preview.redd.it/9crwx4e594lb1.png?width=1263&format=png&auto=webp&s=592f610eec1675e8c502c8ff3c9d1329b865d28a
| 2023-08-29T21:20:30 |
https://www.reddit.com/r/LocalLLaMA/comments/164uskh/if_this_is_true_how_is_llama_open_source/
|
bull_shit123
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164uskh
| false | null |
t3_164uskh
|
/r/LocalLLaMA/comments/164uskh/if_this_is_true_how_is_llama_open_source/
| false | false | 1 | null |
|
if this is true, how is llama open source?
| 1 | 2023-08-29T21:22:09 |
bull_shit123
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
164uu9h
| false | null |
t3_164uu9h
|
/r/LocalLLaMA/comments/164uu9h/if_this_is_true_how_is_llama_open_source/
| false | false | 1 |
{'enabled': True, 'images': [{'id': '_QuH4w_Sizw-9xpwu7_7moZE27CfBqdDIIXUNb7QlLs', 'resolutions': [{'height': 48, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=108&crop=smart&auto=webp&s=58bbcc1b2b812de2110558c26ea3d7161275d7ad', 'width': 108}, {'height': 96, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=216&crop=smart&auto=webp&s=a724b7f1c1c40b47ccdfd12f745a2f866f272176', 'width': 216}, {'height': 142, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=320&crop=smart&auto=webp&s=dcbadbed0266cc77ac9d323575d299480d524f48', 'width': 320}, {'height': 285, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=640&crop=smart&auto=webp&s=cd11e0975ccf34bd56eaa4a955e7c97c0bcfd6a5', 'width': 640}, {'height': 428, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=960&crop=smart&auto=webp&s=bc51cbeb5064d5986e967baf48e8b6479979ac39', 'width': 960}, {'height': 482, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?width=1080&crop=smart&auto=webp&s=e4918367ab0addaf7c512e682e6d8c252123315d', 'width': 1080}], 'source': {'height': 564, 'url': 'https://preview.redd.it/u8qsenee94lb1.png?auto=webp&s=70d9403e668c2e0194702a5c70789c558cc3874d', 'width': 1263}, 'variants': {}}]}
|
|||
HammerAI: desktop app for running LLaMA locally (using Apache TVM Unity)
| 1 |
Hi all! Want to share a new project we've been working on: [**HammerAI**](https://www.hammerai.com/)**, a desktop app that uses** [**Apache TVM Unity**](https://github.com/apache/tvm/tree/unity) **&** [**MLC LLM**](https://github.com/mlc-ai/mlc-llm) **to run models locally on your machine** 🏎️
[HammerAI demo](https://reddit.com/link/164uuik/video/m5t9tdm164lb1/player)
Some key features:
* No configuration needed - download the app, download a model (from within the app), and you're ready to chat ✨
* Works offline 🏖️
* Free 💸
* No sign in needed 🥷
* NSFW content allowed 🤫
* Private - your chat is only stored as long as you have the chat window open (we're working on optionally saving conversations) 🔒
* Automatic detection and use of your GPU 🚄
* Support for V1 and V2 character card imports 🦄
The main limitations:
* Only supports MacOS Apple Silicon (we're working on MacOS Intel and Windows) 💻
* Only supports [Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf), [Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf), and [Nous-Hermes-Llama2-13b](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b) (we're working on adding more models, just let me know what you'd like added!) 🤖
* Several quality of life features that will make it a better experience (configuring the base system prompt, temperature, top p, etc.) 😅
It's an early release, but would love some feedback if you're interested in testing it out - thank you!
| 2023-08-29T21:22:24 |
https://www.reddit.com/r/LocalLLaMA/comments/164uuik/hammerai_desktop_app_for_running_llama_locally/
|
Hammer_AI
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164uuik
| false | null |
t3_164uuik
|
/r/LocalLLaMA/comments/164uuik/hammerai_desktop_app_for_running_llama_locally/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'OyyRrQ6m6Laio5hUxP-x5qgOUVabNuA5yMM4lur0nU0', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=108&crop=smart&auto=webp&s=e14095b86d0187c3d9f97ecefb33849f54fc7aef', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=216&crop=smart&auto=webp&s=9f54466aca7f97be2005bc18a98292ad80a47069', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=320&crop=smart&auto=webp&s=84bdb8035712b08b9ea9d9228360b921824fadb8', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=640&crop=smart&auto=webp&s=63863d60be47ff783c5c61f4eb2d10b1d2333cf5', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=960&crop=smart&auto=webp&s=d26eb15559a49ada324306a7bb51a4218283f305', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?width=1080&crop=smart&auto=webp&s=2c2b4438bd410062961b55cc3697358722f960f0', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/o78_v8waoUJk8qNiQf1qt0lW-pnH8OAUvk8AXqNxSFw.jpg?auto=webp&s=f819e88e0c3cb774560cb99e0da742d732fbff20', 'width': 1200}, 'variants': {}}]}
|
|
Missing tokenizer.model from official llama 2 7b download required for llama.cpp convert to gguf
| 1 |
Hey folks,
I'm new here, but not new to software. However, this is all much more cutting-edge than I'm used to, and it seems like everything is moving quickly.
On discovering that I needed a GPU for the regular llama distribution, I went ahead and snagged llama.cpp, but am having trouble converting my official PyTorch download with the contents of
\`\`\`
checklist.chk
consolidated.00.pth
params.json
\`\`\`
and it seems that the convert script is looking for a root(?) tokenizer.model file, which I simply do not have.
Is there another location to download a more llama.cpp friendly version of llama-2-7b?
| 2023-08-29T22:00:26 |
https://www.reddit.com/r/LocalLLaMA/comments/164vup7/missing_tokenizermodel_from_official_llama_2_7b/
|
Packeselt
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164vup7
| false | null |
t3_164vup7
|
/r/LocalLLaMA/comments/164vup7/missing_tokenizermodel_from_official_llama_2_7b/
| false | false |
self
| 1 | null |
Very slow inference on Windows
| 1 |
I am running 70B models with 8192 context. I have tried it on [vast.ai](https://vast.ai) with 2 x 3090, and it ran pretty fast (\~10 tokens/sec). However, on my windows, it ran pretty slow (\~0.2 tokens/sec).
The spec of my PC is: Ryzen 5900x, C7H Motherboard, 128GB DDR4-3600, 4090 run at PCIE 4.0 x 16, 3090 run at PCIE 4.0 x 4
I have tried the following (with oobabooga + exllama-HF):
On WSL with 2048 truncation: very very fast, > 20 tokens/sec
On WSL with 8192 truncation: prompt evaluation took \~1 mins for \~8000 tokens, but generation was kind of reasonable.
On Windows with 8192 truncation: both prompt evaluation and generation were very slow, 0.2 tokens/sec.
​
Have anyone experienced something similar?
| 2023-08-29T22:44:06 |
https://www.reddit.com/r/LocalLLaMA/comments/164x02t/very_slow_inference_on_windows/
|
denru01
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164x02t
| false | null |
t3_164x02t
|
/r/LocalLLaMA/comments/164x02t/very_slow_inference_on_windows/
| false | false |
self
| 1 | null |
I have been out of the loop. Need a dive in
| 1 |
Can anyone make a summary about everything that happened since first open source LLMs appeared to this day?
Thanks
| 2023-08-29T23:04:27 |
https://www.reddit.com/r/LocalLLaMA/comments/164xiuc/i_have_been_out_of_the_loop_need_a_dive_in/
|
Flutter_ExoPlanet
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164xiuc
| false | null |
t3_164xiuc
|
/r/LocalLLaMA/comments/164xiuc/i_have_been_out_of_the_loop_need_a_dive_in/
| false | false |
self
| 1 | null |
$2900 Worth: L0 Airdrop Surprise
| 1 |
[removed]
| 2023-08-29T23:18:51 |
https://www.reddit.com/r/LocalLLaMA/comments/164xw5i/2900_worth_l0_airdrop_surprise/
|
Puzzleheaded-Wind574
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164xw5i
| false | null |
t3_164xw5i
|
/r/LocalLLaMA/comments/164xw5i/2900_worth_l0_airdrop_surprise/
| false | false |
self
| 1 | null |
LLM questions
| 1 |
Hey gang .. two questions . 1) can I make chatGPT local? Meaning I don't want it connected to the internet at all .. is that doable? 2) can I train it on my data? What are my options there? Ideally I'd love to be able to read data from a table so I could train it on those terms as well.
| 2023-08-29T23:48:32 |
https://www.reddit.com/r/LocalLLaMA/comments/164ymtx/llm_questions/
|
aeapf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164ymtx
| false | null |
t3_164ymtx
|
/r/LocalLLaMA/comments/164ymtx/llm_questions/
| false | false |
self
| 1 | null |
What frameworks or platforms to use for full fine tuning of Code Llama?
| 1 |
I want to test whether a full fine tuning of my codebase will improve Code Llama for me. Any suggestions on how to go about it? I've been searching fine tuning and most are talking supervised fine tuning with LoRA. I imagine I'll have to use distributed training to train anything larger than the 7b model? I've seen mentions on distributed training with DeepSpeed [https://huggingface.co/docs/transformers/perf\_train\_gpu\_many](https://huggingface.co/docs/transformers/perf_train_gpu_many).
Should I use HuggingFace [https://huggingface.co/codellama/CodeLlama-34b-hf](https://huggingface.co/codellama/CodeLlama-34b-hf) or grab the model from Facebook [https://github.com/facebookresearch/codellama](https://github.com/facebookresearch/codellama)?
Any other pointers? Thanks
| 2023-08-30T00:36:50 |
https://www.reddit.com/r/LocalLLaMA/comments/164zs9d/what_frameworks_or_platforms_to_use_for_full_fine/
|
OneHistory4562
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
164zs9d
| false | null |
t3_164zs9d
|
/r/LocalLLaMA/comments/164zs9d/what_frameworks_or_platforms_to_use_for_full_fine/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'jfeVG47nZdEkz9kXfW1CcS-Sy8l4DXGb9JErx6bLKfU', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=108&crop=smart&auto=webp&s=abf38332c5c00a919af5be75653a93473aa2e5fa', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=216&crop=smart&auto=webp&s=1a06602204645d0251d3f5c043fa1b940ca3e799', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=320&crop=smart&auto=webp&s=04833c1845d9bd544eb7fed4e31123e740619890', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=640&crop=smart&auto=webp&s=d592b0a5b627e060ab58d73bde5f359a1058e56d', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=960&crop=smart&auto=webp&s=5913a547536ee8300fdb8a32d14ff28667d1b875', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?width=1080&crop=smart&auto=webp&s=2af86fd4d41393a7d14d45c4bb883bef718575d1', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/Bixm6H31yqw0RCcD8LB0e8eIdtJeMUaF4N5ZipM_BQY.jpg?auto=webp&s=720b78add0a3005c4f67eaed6897df409cc040c6', 'width': 1200}, 'variants': {}}]}
|
Fine-tuning embedding models for better retrieval
| 1 |
For those of you interested in retrieval augmented generation (RAG), would you like to collaborate on fine-tuning embedding models for better retrieval? My friend and I have ML systems background from MIT and AWS and are interested in optimizing embedding models for different use cases and benchmarking performance (like INSTRUCTOR: [https://huggingface.co/hkunlp/instructor-xl](https://huggingface.co/hkunlp/instructor-xl)).
| 2023-08-30T03:05:57 |
https://www.reddit.com/r/LocalLLaMA/comments/16532wy/finetuning_embedding_models_for_better_retrieval/
|
Dapper-Bumblebee-273
|
self.LocalLLaMA
| 2023-08-30T03:21:31 | 0 |
{}
|
16532wy
| false | null |
t3_16532wy
|
/r/LocalLLaMA/comments/16532wy/finetuning_embedding_models_for_better_retrieval/
| false | false |
default
| 1 |
{'enabled': False, 'images': [{'id': 'bK1T3Poc6MndmHJoui8whxY6ySsvI8PDs41pAsuL7yI', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=108&crop=smart&auto=webp&s=c955d2e70abcfbce0a4c1bc8c361c4a575d50985', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=216&crop=smart&auto=webp&s=6df2ab29063d9a9740b00f0d664aed048e87d85d', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=320&crop=smart&auto=webp&s=1a2de1f009323b5b3a15c915e328b2b9161f2002', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=640&crop=smart&auto=webp&s=15189cfbafac726d9b0099444884f0eb2c45f578', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=960&crop=smart&auto=webp&s=1453d4a3f86509e3d51d322d1c1286074849b840', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?width=1080&crop=smart&auto=webp&s=79a3f2544429da5e853b1790495f99e2ffa5c128', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/mpjfav1XEpa-CbfCnWpqNpgu9Ad47U50XNeaMjHB9X4.jpg?auto=webp&s=70cd1ae20ab45c0a4c2cfe6ee8738b9fc464e9df', 'width': 1200}, 'variants': {}}]}
|
(help) oogabooga reloads everything each time the presets are changed
| 1 |
Hi!
I'm working on a project using Oogabooga's webui API, and I've noticed a significant issue.
In my use case, I need to change the presets I use (characters, parameters, etc.) for almost every generation. However, during the first generation and each time I use different presets, there is a very long loading time (2 minutes). I always use the same model, and apart from that, my performance is good enough (> 5T/s).
Is there a way to keep in memory whatever the webui computes when I use new presets (I'm a noob so I don't even know what it does at this step)? If not, is there another solution?
| 2023-08-30T03:21:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1653egv/help_oogabooga_reloads_everything_each_time_the/
|
GM1-BluePhoenix
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1653egv
| false | null |
t3_1653egv
|
/r/LocalLLaMA/comments/1653egv/help_oogabooga_reloads_everything_each_time_the/
| false | false |
self
| 1 | null |
Used 3060's now half the price of 3090's per GB
| 1 |
I tried to find an existing thread to post a comment on, but couldn't find anything after a cursory search. I'm in the process of building out a new rig for local LLM, based on AMD EPYC w/ a 7-slot Asrock Rack motherboard.
Until now, I've looked at the 3090 FE as the basic building block of a good local system. Modern CUDA, large VRAM, decent price on used market, and a fairly common setup (for compatability) make it attractive. A lot of people are running 1x3090 or 2x3090; beyond that, we get into the territory of multi-thousand-dollar cards.
When researching what to put in those seven PCIe Gen 4x16 slots, I found that the 3060 12GB cards can now be had for \~230US, for a $/GB of 19. The 3090 24GB, meanwhile, is going for \~800US, for a $/GB of 33. (All prices from a ebay used searches).
Assuming constraints like PCI card slots, PCIe lanes, power, and physical space to mount the card are equal, it seems the 3060's are actually a very attractive choice right now for multi-gpu inference.
The 3060 has significantly fewer CUDA cores and smaller memory bandwidth. Tim D's amazing article suggests memory bandwidth is a bigger constraint than CUDA cores, so perhaps the 3060 should be ruled out for that reason alone. Still, supposing a budget of $1000, it's tempting to consider picking up 4 identical 3060's to run 48GB VRAM vs one 3090 and 24GB VRAM.
This also opens up the question of AMD/ROCm, etc now that inference and training libraries are getting more support for things beyond CUDA. And I still have questions about how/if the 40-series will benefit inference and training in the coming years. (My intuition is that the 4000-series should be skipped entirely for this purpose and we should wait for 5000 if we want features beyond what the 3000's offer).
Hopefully there's some value in this mini-rant - will be curious to hear about any flaws in my thinking, or any other aspects I didn't consider.
| 2023-08-30T03:54:15 |
https://www.reddit.com/r/LocalLLaMA/comments/16542jv/used_3060s_now_half_the_price_of_3090s_per_gb/
|
tronathan
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
16542jv
| false | null |
t3_16542jv
|
/r/LocalLLaMA/comments/16542jv/used_3060s_now_half_the_price_of_3090s_per_gb/
| false | false |
self
| 1 | null |
Has anybody tried getting RAG to work with Code Llama?
| 1 |
It seems like Code Llama isn't made for RAG. When I try to manually copy and paste the retrieved info, it gets a seizure. Outputs range from a block of nothing or other outputs such as:
[INST]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]
Or it will just repeatedly spam one word
test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test test
I've tried using the prompt format and manually copying and pasting the retrieval results using pipeline instead of LangChain's retrievalQA:
print(pipeline("""[INST] Use the information below to generate a class and the relevant functions [/INST]
<<SYS> <vector retrieved info> <</SYS>>""")
and also:
[INST] Use the information below to generate a class and the relevant functions
<vector retrieved info>
[/INST]
But no dice. Even RAG with Starcoder generates somewhat coherent code. Any ideas on how to get this to output decent code? Seems like Code Llama is very picky with the formatting. Thanks in advanced!
| 2023-08-30T04:07:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1654cg0/has_anybody_tried_getting_rag_to_work_with_code/
|
salah_ahdin
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1654cg0
| false | null |
t3_1654cg0
|
/r/LocalLLaMA/comments/1654cg0/has_anybody_tried_getting_rag_to_work_with_code/
| false | false |
self
| 1 | null |
Can someone bring me up to speed?
| 1 |
Before you call me lazy, I totally get it. Long story short, I used to use LLMs a lot but haven’t in a few months because I’ve been busy. I used llama.cpp on my Mac and was able to host a local web ui. The last time I tried to get back in the game, I swear so much had changed and I’m not going to lie I gave up.
Without doing my homework, if anyone can share some basic info on what the setup is these days I’d really appreciate it.
Bonus points if you share your favorite model
Thanks!
| 2023-08-30T04:24:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1654of4/can_someone_bring_me_up_to_speed/
|
themostofpost
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1654of4
| false | null |
t3_1654of4
|
/r/LocalLLaMA/comments/1654of4/can_someone_bring_me_up_to_speed/
| false | false |
self
| 1 | null |
Providing a text-to-SQL-to-text service using API in production. How can I properly size the system?
| 1 |
Some requirements:
* User Traffic: 1000 users
* Number of tokens: 1300 tokens per request (in-context learning, prompt: 1k+ generation: <300 toks)
* Concurrent Requests: 100 requests/s
Peak Time: 8h-9h and 14-15h
Types of Queries: simple queries and a few complex queries.
How can I properly size the system?
* GPU
* CPU
* Memory (RAM)
* Storage
* Network Bandwidth
* Model (opensource: Llama, MPT, Falcon,.. or closed source OpenAI, Cohere, Anthropic,...)
* Budget.
​
| 2023-08-30T05:17:14 |
https://www.reddit.com/r/LocalLLaMA/comments/1655nj6/providing_a_texttosqltotext_service_using_api_in/
|
UncleDao
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1655nj6
| false | null |
t3_1655nj6
|
/r/LocalLLaMA/comments/1655nj6/providing_a_texttosqltotext_service_using_api_in/
| false | false |
self
| 1 | null |
If context size is like the models "active" memory what does the --contextsize flag do when launching kcpp?
| 1 |
[removed]
| 2023-08-30T05:18:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1655ogf/if_context_size_is_like_the_models_active_memory/
|
wh33t
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1655ogf
| false | null |
t3_1655ogf
|
/r/LocalLLaMA/comments/1655ogf/if_context_size_is_like_the_models_active_memory/
| false | false |
self
| 1 | null |
Microsoft's WizardLM claims Phind stole their work without credit to make the Phind CodeLlama model
| 1 | 2023-08-30T05:27:21 |
https://twitter.com/WizardLM_AI/status/1696527036804989171
|
BetterClerk9769
|
twitter.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1655tqe
| 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>Let's start a ⚔️Duel and Debate between WizardCoder and Phind ! Also between open-source and stealing.<br><br>Firstly, let's restore the entire awkward storyline step by step:<br><br>1. Today, users found Phind update their V2 model, but config model-path is… <a href="https://t.co/dFDiW4ApSb">pic.twitter.com/dFDiW4ApSb</a></p>— WizardLM (@WizardLM_AI) <a href="https://twitter.com/WizardLM_AI/status/1696527036804989171?ref_src=twsrc%5Etfw">August 29, 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/1696527036804989171', 'version': '1.0', 'width': 350}, 'type': 'twitter.com'}
|
t3_1655tqe
|
/r/LocalLLaMA/comments/1655tqe/microsofts_wizardlm_claims_phind_stole_their_work/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'jWw43lSv6CCJUzkpTgdFdwbX9QtXRhell_HnS9kheQg', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/yHvqe-lkz0eAeecpa1SE2Hyy_H5lc-1RN29KEf4U878.jpg?width=108&crop=smart&auto=webp&s=c3ab66d7d59e51db360e7d5b2cd6f7b5e165b1d2', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/yHvqe-lkz0eAeecpa1SE2Hyy_H5lc-1RN29KEf4U878.jpg?width=216&crop=smart&auto=webp&s=e986bd8f6646727712c34367845e96a7d9545baa', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/yHvqe-lkz0eAeecpa1SE2Hyy_H5lc-1RN29KEf4U878.jpg?width=320&crop=smart&auto=webp&s=a2b2a5c6d55ffa4882ef9bb3bdd01be8dc16fac2', 'width': 320}], 'source': {'height': 1200, 'url': 'https://external-preview.redd.it/yHvqe-lkz0eAeecpa1SE2Hyy_H5lc-1RN29KEf4U878.jpg?auto=webp&s=15ba3ef2544f91585854680be5c1632fbf0b66c7', 'width': 540}, 'variants': {}}]}
|
||
7900xtx vs 3090 finetuning and inference speeds
| 1 |
With the recent updates with rocm and llama.cpp rupport for rocm, how does the 7900xtx compare with the 3090 in inference and fine tuning? In Canada, You can find the 3090 on ebay for \~1000cad while the 7900xtx runs for 1280$. Is it worth the extra 280$? Using gentoo linux.
| 2023-08-30T06:16:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1656o5f/7900xtx_vs_3090_finetuning_and_inference_speeds/
|
fraschm98
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1656o5f
| false | null |
t3_1656o5f
|
/r/LocalLLaMA/comments/1656o5f/7900xtx_vs_3090_finetuning_and_inference_speeds/
| false | false |
self
| 1 | null |
Llama-2-7b finetuned on LIMA dataset
| 1 |
I finetuned llama2 using the LIMA dataset
Try it out
https://colab.research.google.com/drive/17rxuyf4K_o63DPFwQNy53rdducCqLi7t?usp=sharing
| 2023-08-30T06:31:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1656ww4/llama27b_finetuned_on_lima_dataset/
|
Sufficient_Run1518
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1656ww4
| false | null |
t3_1656ww4
|
/r/LocalLLaMA/comments/1656ww4/llama27b_finetuned_on_lima_dataset/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'nkhh65ujo5BznFJFojoMPaKjGuLSpPj6KGhRov-ykOg', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/MrcDZx2izDY9ERwgWmMS-Hm2M3GEKZgeYLDszSh-KrQ.jpg?width=108&crop=smart&auto=webp&s=4b647239f77bf713f4a6209cfa4867351c055fd9', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/MrcDZx2izDY9ERwgWmMS-Hm2M3GEKZgeYLDszSh-KrQ.jpg?width=216&crop=smart&auto=webp&s=7f4234ff3f4f4ebd7f77236dedb03a2faee3e04a', 'width': 216}], 'source': {'height': 260, 'url': 'https://external-preview.redd.it/MrcDZx2izDY9ERwgWmMS-Hm2M3GEKZgeYLDszSh-KrQ.jpg?auto=webp&s=73eb91ea5a5347f216c0f0c4d6796396826aae49', 'width': 260}, 'variants': {}}]}
|
Does anyone have a model trained on the new NextJS docs?
| 1 |
I know it’s not strictly related to this community but I don’t know if I’ll be able to get it from somewhere else.
| 2023-08-30T06:34:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1656yx7/does_anyone_have_a_model_trained_on_the_new/
|
as904465
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1656yx7
| false | null |
t3_1656yx7
|
/r/LocalLLaMA/comments/1656yx7/does_anyone_have_a_model_trained_on_the_new/
| false | false |
self
| 1 | null |
How to handle long documents when pretraining a LLM?
| 1 |
I have a few questions about preprocessing data when training a LLM:
First of all, in the pretraining phase of a LLM, how are long documents usuallly processed? I assume they would be split into smaller chunks, but how exactly to split them? Are they split into multiple chunks of equal length (e.g 2048 tokens) or are they split based on paragraphs of each document? If they are split based on paragraphs, then how to handle the case where a paragraph longer than 2048 tokens? Moreover, are there any overlap between each chunk?
Secondly, how are the data shuffled and forwarded to the model? At first, I thought all examples from all datasets would be mixed together and shuffled randomly and the model is trained for 1 epoch on all examples. However in the LLaMA paper, I saw this \[table\]([https://i.imgur.com/ccQfLi8.png](https://i.imgur.com/ccQfLi8.png)) and I don't understand what "epoch" here in this case means and why are they float instead of integer.
I would appreciate it if someone can help me.
| 2023-08-30T07:06:24 |
https://www.reddit.com/r/LocalLLaMA/comments/1657ib5/how_to_handle_long_documents_when_pretraining_a/
|
scienceotaku68
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1657ib5
| false | null |
t3_1657ib5
|
/r/LocalLLaMA/comments/1657ib5/how_to_handle_long_documents_when_pretraining_a/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'lFeK4dYTg8Rk1Xdy3fHo_9Wves_c2bHavjE0hUKNICU', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/cHmks5dloR0pFXq344_f-TWtz2mVr9KnMv6kG1Npl-g.png?width=108&crop=smart&auto=webp&s=8ae514a337d7607c597d6e125044d474e06853bc', 'width': 108}, {'height': 120, 'url': 'https://external-preview.redd.it/cHmks5dloR0pFXq344_f-TWtz2mVr9KnMv6kG1Npl-g.png?width=216&crop=smart&auto=webp&s=2ce975d415e2a0a95db563a53971c9c8e6491322', 'width': 216}, {'height': 177, 'url': 'https://external-preview.redd.it/cHmks5dloR0pFXq344_f-TWtz2mVr9KnMv6kG1Npl-g.png?width=320&crop=smart&auto=webp&s=1ed3805bf3ceeee3f612b4470136bbaaa6b5ce85', 'width': 320}, {'height': 355, 'url': 'https://external-preview.redd.it/cHmks5dloR0pFXq344_f-TWtz2mVr9KnMv6kG1Npl-g.png?width=640&crop=smart&auto=webp&s=cd86ff110d34023f72d3ca6a3e47bbb451d88a27', 'width': 640}], 'source': {'height': 447, 'url': 'https://external-preview.redd.it/cHmks5dloR0pFXq344_f-TWtz2mVr9KnMv6kG1Npl-g.png?auto=webp&s=2e9ab276d1c728236361297c75a61619434dc233', 'width': 804}, 'variants': {}}]}
|
$4600 L0 Airdrop Party Begins
| 1 |
[removed]
| 2023-08-30T07:53:21 |
https://www.reddit.com/r/LocalLLaMA/comments/16589tf/4600_l0_airdrop_party_begins/
|
SentenceSeveral
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
16589tf
| false | null |
t3_16589tf
|
/r/LocalLLaMA/comments/16589tf/4600_l0_airdrop_party_begins/
| false | false |
self
| 1 | null |
Advice on Open source LLM for pdf query aka Retrieval Augmented Generation
| 1 |
Hi guys,
Can you suggest an open source model which is good enough for PDF query . The PDF can be Annual reports or sustainability reports. I trued Llama 2 7b quantised model but the response time is 2+ minutes for query since I’m using CPU (16 gb ram)
I also tried open AI GPT 3.5 turbo, but a) It’s not free and b) Its a API , and my data might be proprietary, so low privacy.
| 2023-08-30T07:58:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1658csg/advice_on_open_source_llm_for_pdf_query_aka/
|
ianuvrat
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1658csg
| false | null |
t3_1658csg
|
/r/LocalLLaMA/comments/1658csg/advice_on_open_source_llm_for_pdf_query_aka/
| false | false |
self
| 1 | null |
Best Llama 2 13B based LLM for Context Question Answering
| 1 |
Hi there,
I am facing the following common task: I have a textual context and there may or may not be explicit or implicit information present to answer a given question. The LLM should answer (ideally English and/or German) only with information found in the context or else say "don't know" or something of that kind.
My experiments with LLama 2 Chat 13B are quite mixed. Very often at least some parts of the answer are not based on the context and/or are outright wrong. Even with some quite simple examples like a Paragraph from Wikipedia and a simple question.
I tried different prompts from various web sources, as well as some custom made ones. I did not see a large difference in performance between them.
This is such a common task and there are lots of resources available, however the performance of the open source models I tried is underwhelming.
Do you know of any finetuned version specialized in this task and yielding reliable results? It does not even need to be LLama 2 based, as long as there is a commercial license and its max 20B in size.
**What are the best models for this use case available at the moment and which prompts should be used?**
Thank you very much!
​
| 2023-08-30T08:11:54 |
https://www.reddit.com/r/LocalLLaMA/comments/1658kk1/best_llama_2_13b_based_llm_for_context_question/
|
__q_e_d__
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1658kk1
| false | null |
t3_1658kk1
|
/r/LocalLLaMA/comments/1658kk1/best_llama_2_13b_based_llm_for_context_question/
| false | false |
self
| 1 | null |
first time beginner - WOOOOW! what a subreddit! we can run locally chatbot!? incredible..
| 1 |
[removed]
| 2023-08-30T08:33:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1658xfn/first_time_beginner_woooow_what_a_subreddit_we/
|
ovnf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1658xfn
| false | null |
t3_1658xfn
|
/r/LocalLLaMA/comments/1658xfn/first_time_beginner_woooow_what_a_subreddit_we/
| false | false |
self
| 1 | null |
Lets have some real debate here
| 1 |
is llama open source? yes or no
also comment for some real discussion
[View Poll](https://www.reddit.com/poll/1659kib)
| 2023-08-30T09:11:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1659kib/lets_have_some_real_debate_here/
|
bull_shit123
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1659kib
| false | null |
t3_1659kib
|
/r/LocalLLaMA/comments/1659kib/lets_have_some_real_debate_here/
| false | false |
self
| 1 | null |
PCIe 5.0 NVMe worth it?
| 1 |
Hi, I am finishing a "do anything" build for LocalLlamas, XR, a few other things. I'd like to be able to use 60b size models, and multiple models at the same time.
I want to include 4TB of storage. These days a 4TB T700 is about 2× the price of a PCIe 4.0 model, its raw sequential rate is about 50% better.
My impression is this work is all about sequential rates. Am I going to see 50% improvements? Any estimates on the actual impact? I see people's basic load times are all over the map for larger models, from \~10s to much more.
gpt4 kindly created this summary table. Note I don't currently have a third GPU, I'm considering a used A4000. I am using the iGPU with NixOS for general use.
| Part | PCIe Lanes Required | Available Lanes on Asus ProArt B650 | Bandwidth Impact | Compromises | Power Consumption (W) |
|-------------------------------|---------------------|------------------------------------|------------------|---------------------------------------------------|-----------------------|
| 1st GPU (3090 limited to 250W) | 16 | 8 | Reduced to x8 | Half the maximum bandwidth | 250 |
| 2nd GPU (3090 limited to 250W) | 16 | 8 | Reduced to x8 | Half the maximum bandwidth | 250 |
| 3rd GPU (RTX A4000) | 16 | 1 or 4 (x1 or x4 slot) | Severely Reduced | Not suitable for high-performance tasks | 140 |
| 1st PCIe 4.0 NVMe | 4 | 4 (One of the M.2 slots) | Full Bandwidth | Consumes 4 lanes of the 12 total M.2 lanes | \~8 |
| 2nd PCIe 4.0 NVMe | 4 | 4 (Another M.2 slot) | Full Bandwidth | Consumes another 4 lanes of the 12 total M.2 lanes | \~8 |
| PCIe 5.0 NVMe Drive | 4 | 4 (PCIe 5.0 M.2 slot) | Full Bandwidth | Consumes all lanes of one M.2 slot | \~8 |
| System Board | N/A | N/A | N/A | Asus ProArt B650 Creator | \~50 |
| CPU (AMD Ryzen 9 7950X3D) | N/A | N/A | N/A | 120W TDP | 120 |
| RAM (96GB DDR5 6000) | N/A | N/A | N/A | N/A | \~15 |
It has a good quality 1000w PSU, but I could limit the GPUs to 250w.
Bonus question: I want to use KVM to avoid making a lot of messes, if I pass the GPUs and that drive through via pci-passthrough, will there be much performance impact?
Thanks!
| 2023-08-30T10:13:00 |
https://www.reddit.com/r/LocalLLaMA/comments/165ann4/pcie_50_nvme_worth_it/
|
nostriluu
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165ann4
| false | null |
t3_165ann4
|
/r/LocalLLaMA/comments/165ann4/pcie_50_nvme_worth_it/
| false | false |
self
| 1 | null |
Home setup for future home model use
| 1 |
Hello fellow redditors,
I could really use some technical advice.
I am about to start studying AI and am currently preparing my tech setup. Since I am using multiple PCs in the same local network I am planning to set up a NAS of some sort for data backups and access.
Here is the catch though: I want to be able to run a local model in my home for testing and learning as soon as I can including access to my smart home devices and NAS server.
Have you guys got any advice how I would actualize this idea? My easiest NAS solution would be a hard drive on my router. But what would be the best kind of device to run a local model permanently? Are there simply no devices yet in the current hardware evolution? Are there maybe dedicated NAS devices that could run a 7B model?
Any help to plan my setup would be highly appreciated.
| 2023-08-30T10:14:31 |
https://www.reddit.com/r/LocalLLaMA/comments/165aoi7/home_setup_for_future_home_model_use/
|
Sockosophist
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165aoi7
| false | null |
t3_165aoi7
|
/r/LocalLLaMA/comments/165aoi7/home_setup_for_future_home_model_use/
| false | false |
self
| 1 | null |
Why people use RoPE instead of Alibi when buliding their models?
| 1 |
I think Alibi has performed its great ability on length extrapolation, whereas rope preforms not that well (with ppl. explosion in long context scenario).
I know with ntk-aware scaled rope, RoPE can perform well in length extrapolation scenario too. But before that, why people are more likely to use RoPE? What's the advantage of RoPE over Alibi?
| 2023-08-30T10:33:41 |
https://www.reddit.com/r/LocalLLaMA/comments/165b0tw/why_people_use_rope_instead_of_alibi_when/
|
zetiansss
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165b0tw
| false | null |
t3_165b0tw
|
/r/LocalLLaMA/comments/165b0tw/why_people_use_rope_instead_of_alibi_when/
| false | false |
self
| 1 | null |
Consultants for finetuning LLM
| 1 |
We've so far been pretty successful in building AI Agents for our chatbot that takes reservations and connects with other APIs to provide users information. We're looking to transition to an opensource LLM because of rate limits on GPT 4 and cost savings in the long term. We have a userbase of around a million users who will be using these systems when deployed across all our clients.
Does anyone here do consulting work who might be interested or know a platform that would provide these services? We need something more hands on than what huggingface and others provide atm.
| 2023-08-30T10:42:20 |
https://www.reddit.com/r/LocalLLaMA/comments/165b64n/consultants_for_finetuning_llm/
|
hi87
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165b64n
| false | null |
t3_165b64n
|
/r/LocalLLaMA/comments/165b64n/consultants_for_finetuning_llm/
| false | false |
self
| 1 | null |
Llama 2 33b - still no word?
| 1 |
33B is considered the sweet spot. Meta is supposedly working on it. But it's been quite a while since the other sized L2 models were released -- which are very much worth it. I love working with them.
But I can't help but wonder wether we'll see the 33B L2 model... at all.
I can't seem to find any news on it. Does anyone know why it's taking so long? Why it was delayed at all? If they'll release it at all?
​
​
| 2023-08-30T11:28:52 |
https://www.reddit.com/r/LocalLLaMA/comments/165c2e5/llama_2_33b_still_no_word/
|
Ekkobelli
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165c2e5
| false | null |
t3_165c2e5
|
/r/LocalLLaMA/comments/165c2e5/llama_2_33b_still_no_word/
| false | false |
self
| 1 | null |
new serverless templates on banana for community use covering the entire code llama family
| 1 |
i wanted to share these new templates with you all 🦙😊:
[code llama-7b](https://github.com/yachty66/demo-codellama-7b-gptq)
[code llama-13b](https://github.com/yachty66/demo-codellama-13b-gptq)
[code llama-34b](https://github.com/yachty66/demo-codellama-34b-gptq)
[code llama-python-7b](https://github.com/yachty66/demo-codellama-7b-python-gptq)
[code llama-python-13b](https://github.com/yachty66/demo-codellama-13b-python-gptq)
[code llama-python-34b](https://github.com/yachty66/demo-codellama-34b-python-gptq)
[code llama-instruct-7b](https://github.com/yachty66/demo-codellama-7b-instruct-gptq)
[code llama-instruct-13b](https://github.com/yachty66/demo-codellama-13b-instruct-gptq)
[code llama-instruct-34B](https://github.com/yachty66/demo-codellama-34b-instruct-gptq)
| 2023-08-30T11:30:18 |
https://www.reddit.com/r/LocalLLaMA/comments/165c3h2/new_serverless_templates_on_banana_for_community/
|
yachty66
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165c3h2
| false | null |
t3_165c3h2
|
/r/LocalLLaMA/comments/165c3h2/new_serverless_templates_on_banana_for_community/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'PUmsVOWFft4XHrVj0jaVpQtzDMASm3JlzY0yQn17B-w', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=108&crop=smart&auto=webp&s=9ed573aef9fa6c1e3402493583d759b516093b24', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=216&crop=smart&auto=webp&s=d10a4ea012825439ded85aed615561c6578ae091', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=320&crop=smart&auto=webp&s=031be87119056a97ee358a596c306e5a74d18dcc', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=640&crop=smart&auto=webp&s=142d368bc4fdcbf22035b60c0c109a5b8f12a4ee', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=960&crop=smart&auto=webp&s=32ba9aa0bad1d980541b456c3e0fe0c90208ba7a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?width=1080&crop=smart&auto=webp&s=a7dfd5407907dc5e6c4f51baff28e01b369431d7', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/SFP2bCWivrnz_Jc3oS4SEs8UMa3SpKKtrgMH5cqG1HE.jpg?auto=webp&s=2c9398eb5ec818c3fa88fdc058536283649f0cf6', 'width': 1200}, 'variants': {}}]}
|
Stanford DSPy: The framework for programming with foundation models
| 1 |
https://github.com/stanfordnlp/dspy
Seems pretty interesting. Anyone tried it?
Looks like an attempt to combine LMMs, RAG and traditional programming, can use LLama.
| 2023-08-30T11:34:04 |
https://www.reddit.com/r/LocalLLaMA/comments/165c66u/stanford_dspy_the_framework_for_programming_with/
|
BalorNG
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165c66u
| false | null |
t3_165c66u
|
/r/LocalLLaMA/comments/165c66u/stanford_dspy_the_framework_for_programming_with/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': '8A5cnGHUiDqjqKQEd0rUhnisqjuDne3_7FwMOmuUt_w', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=108&crop=smart&auto=webp&s=0babf80bab3864eed8f22eb15156ed54b3f5e8a5', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=216&crop=smart&auto=webp&s=80b73aa96d3cdff8bf56e3f67d3bd470ce206c04', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=320&crop=smart&auto=webp&s=77ab414b4e10d1cdbf69b1ce0d658bc23c2c66f0', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=640&crop=smart&auto=webp&s=8ad82d9ae9316ee6b6af08d622762ab5aeaa6a44', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=960&crop=smart&auto=webp&s=c2cfcd1f1cbe8c5667010e78f4cad8beaf2f2f8b', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?width=1080&crop=smart&auto=webp&s=c7fbb883b63080bb5bf03582855215e9f4e88f8b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/XplIPiTnh2sI7HYujsEnOblvpLtd1_sNtYtufoYfebI.jpg?auto=webp&s=4c3c95e311c8fda53312d3a0ad3bcc112e2f368d', 'width': 1200}, 'variants': {}}]}
|
Training for small input/output use case
| 1 |
[removed]
| 2023-08-30T11:45:47 |
https://www.reddit.com/r/LocalLLaMA/comments/165cev0/training_for_small_inputoutput_use_case/
|
Otherwise-Poet-4362
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165cev0
| false | null |
t3_165cev0
|
/r/LocalLLaMA/comments/165cev0/training_for_small_inputoutput_use_case/
| false | false |
self
| 1 | null |
Best performing lightweight models (Compatible with 4 GB 3040 GPUs)
| 1 |
[removed]
| 2023-08-30T12:16:23 |
https://www.reddit.com/r/LocalLLaMA/comments/165d2on/best_performing_lightweight_models_compatible/
|
Perfect-Storm-99
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165d2on
| false | null |
t3_165d2on
|
/r/LocalLLaMA/comments/165d2on/best_performing_lightweight_models_compatible/
| false | false |
self
| 1 | null |
Has anyone used Petals yet?
| 1 | 2023-08-30T12:29:27 |
fhirflyer
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
165dd1g
| false | null |
t3_165dd1g
|
/r/LocalLLaMA/comments/165dd1g/has_anyone_used_petals_yet/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'yXXwyx7KlTnHN_NOXKoZgDj1hDGxQHu9zjqNBTeXPPg', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?width=108&crop=smart&auto=webp&s=8ecb58b523c7f89eafae32583a5fd059d5a1a679', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?width=216&crop=smart&auto=webp&s=80d76cc967151c67b0e7406eae19ba90969f04cc', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?width=320&crop=smart&auto=webp&s=199fd96496addeccc432f14ec155b5347c10f960', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?width=640&crop=smart&auto=webp&s=ae67532122bd1682ce605628111ce2a74b4d19b7', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?width=960&crop=smart&auto=webp&s=eb45fb9a61c12ca2821329ed41da0903ef86ffc4', 'width': 960}], 'source': {'height': 1024, 'url': 'https://preview.redd.it/vytgnm9cr8lb1.jpg?auto=webp&s=05c45f53487ff69b2a5e5bad92d26231045c7ac9', 'width': 1024}, 'variants': {}}]}
|
|||
Llama2 3B fine tuning
| 1 |
Hello,
I am planning on fine tuning llama2 3B with custom corpus of text.
Since my GPU is not enough (RTX 3060ti) I plan on renting a service for this.
What service you suggest me to use?
| 2023-08-30T12:35:39 |
https://www.reddit.com/r/LocalLLaMA/comments/165di30/llama2_3b_fine_tuning/
|
Soft-Conclusion-2004
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165di30
| false | null |
t3_165di30
|
/r/LocalLLaMA/comments/165di30/llama2_3b_fine_tuning/
| false | false |
self
| 1 | null |
Using 30b and above models for fiction?
| 1 |
Preable: I'm trying my hand at "AI fiction writing" - I have a few ideas worth exporing I think, but being, well, "neurodivergent and not in a cool fashion" I don't have any illusions whether I can write something actually readable, not to mention in a reasonable amount of time I don't have much of.
I've tried several finetunes including "storywrite" among 13b models I can run on my 2060 12gb (with pretty decent 10-15 t/s), but it either don't really "get" what I want, or outputs something... "trite and robotic", so basically hardly useful for what I want - waaaay too much work in editing everything.
I've ended up using Claude mostly, it sticks to instructions very well, can be coaxed to write... adequate prose with a page of style cues (I have no aspirations to become next King, just something that gets the point across without being cringy).
I have zero interest in "nsfw roleplay" (too old for this shit, though I can see how I could be very interested 20 years ago), and frankly I'd be happy to use Claude Instant and be done with it, but given that I'm from Russia (something I've never chosen and cannot change for now, unfortunately, I have commitments) I've already been blocked from a couple of free llms "providers", and jumping through the hoops of VPNs and payments given *current situation* and not actually owning "the tools of my trade" is daunting and depressing.
Here is an unedited example of a chapter from Claude, prompt included, which is far from perfect but useable without too much editing:
***
"Here is a list of stylistic cues:
Draw comparisons between unrelated concepts to make a point
Consider multiple perspectives on issues
Weave in allusions to myths, religions, philosophies and social/political issues to ground discussions and support points.
Employs sarcasm and humor at times
Pose hypothetical questions and ideas to spur further thought
Tone remains casual yet firmly guides the discussion through complex ideas
Draw on scientific/medical concepts and terminology but in an accessible way
Weave in backstory and world-building details organically without infodumps
Use dialogue and internal thoughts to advance the plot and reveal character perspectives.
Descriptive language is vivid and visceral yet not gratuitous or graphic.
Use natural language and economic, punchy prose.
Use it to write 1500 word prose on first two beats. "
" Here is a 1500 word draft of the first two story beats from Chapter 1 based on the provided stylistic cues:
The hum of machinery echoed through the cavernous laboratory, an industrial hymn to progress and discovery. For Mila, it had become a familiar soundtrack over the past year, one that usually soothed her focus. But today its rhythms only aggravated a splitting headache that refused to subside.
She pinched the bridge of her nose and took a steadying breath, willing her mind back to the task at hand. An elemental summoning array was etched into the stone floor before her, its geometric patterns glowing with contained power. All that remained was the activation sequence to channel those forces into the prototype batteries lined up along the perimeter. One small step toward unlocking new sources of clean energy, or so the rhetoric went.
Mila had her doubts about such lofty claims. Lord Harkin who funded these experiments spoke of revolutionizing society, yet his true motives were shrouded like the smokestacks belching from the city above. Power and profit, it seemed, were the real elements being summoned here. Still, the work gave purpose and she believed scientific progress should uplift lives, not control them.
With a final check of the containment seals, Mila began the activation. Arcane syllables flowed from her lips in a language older than humanity itself, resonant tones that stirred primal currents in the bedrock. Her headache spiked sharply as the array flared to life, colors bleeding together in a kaleidoscopic maelstrom. The pain crested into a flash of blinding white - and then she was somewhere else.
Vast plains stretched as far as the eye could see, barren and windswept under an alien sky. Strange constellations wheeled overhead while in the distance, colossal spires of obsidian jutted from the cracked earth like accusing fingers. Something was wrong with this landscape, its very essence warped and unstable. Even the air tasted foul on her tongue.
As quickly as it came, the vision faded. Mila found herself on her knees within the summoning array, its glow extinguished. Her breathing was ragged. Had she collapsed during the ritual? No, that couldn't be right - the batteries showed signs of a successful charge. But that place, those _things_ she saw...
Shaking off lingering disorientation, Mila rose unsteadily. The experiment would have to wait for a full report and analysis. For now, rest was needed to settle her frayed nerves. Perhaps this headache was playing tricks, conjuring fever dreams from exhaustion.
Yet in the medical bay, she found confirmation that something was indeed amiss. Beds filled with other workers suffering headaches, nausea, hallucinations - all symptoms matching her own. But Dr. Elan's examinations uncovered no physical cause. "Stress and fatigue, nothing more" was the diagnosis, though his furrowed brow hinted at deeper concern. "
***
Is there a way to create have plot outlines/character BIOs/actual story beats and have the AI stick to them while generating prose of at least similar quality?
That's not coding and model "hallucinating" stuff does not matter, so long as it sticks to instructions and context, but it rarely does, sometimes descending into outright gibberish.
I can invest a considerable sum so I can run 30 and 70b models, but are they THAT much different when it comes to writing?
Can anyone provide similar examples or hints?
| 2023-08-30T13:04:41 |
https://www.reddit.com/r/LocalLLaMA/comments/165e6tj/using_30b_and_above_models_for_fiction/
|
BalorNG
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165e6tj
| false | null |
t3_165e6tj
|
/r/LocalLLaMA/comments/165e6tj/using_30b_and_above_models_for_fiction/
| false | false |
self
| 1 | null |
I created a “Choose Your Own Adventure” quest written by LLaMA and illustrated by Stable Diffusion
| 1 |
You can play it with your browser: https://fateful.quest
This is an experiment to see if AI can write something fun like this by itself. It’s pretty good!
I used ChatGPT4 to create the plot synopsis with all the branches since I figured you needed a big model for that. But then, every synopsis line is expanded into a three scene story with LLaMA. Mostly to save on API cost in case the quest reaches thousands of scenes :)
With LLaMA I used Jon Durbin's airoboros 33B m2.0 which I run on my own 4090 machine.
Feedback appreciated! Also if you’re interested in the source code to create your own, let me know.
| 2023-08-30T13:05:02 |
https://www.reddit.com/gallery/165e739
|
_talkol_
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
165e739
| false | null |
t3_165e739
|
/r/LocalLLaMA/comments/165e739/i_created_a_choose_your_own_adventure_quest/
| false | false | 1 | null |
|
Prosecraft.io Ai Lingusitic Analysis that Suffer Great Ai Backlash and yet no one know it
| 1 |
[removed]
| 2023-08-30T13:12:12 |
https://www.reddit.com/r/LocalLLaMA/comments/165edb8/prosecraftio_ai_lingusitic_analysis_that_suffer/
|
Merchant_Lawrence
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165edb8
| false | null |
t3_165edb8
|
/r/LocalLLaMA/comments/165edb8/prosecraftio_ai_lingusitic_analysis_that_suffer/
| false | false |
self
| 1 | null |
Is there a good integration for local models like CodeLlama into an IDE like VSCode?
| 1 |
So far I only came across [cursor.so](https://cursor.so) and Github Co-Pilot but they both don't offer sufficient local capabilities.
| 2023-08-30T13:21:32 |
https://www.reddit.com/r/LocalLLaMA/comments/165elij/is_there_a_good_integration_for_local_models_like/
|
donotdrugs
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165elij
| false | null |
t3_165elij
|
/r/LocalLLaMA/comments/165elij/is_there_a_good_integration_for_local_models_like/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'I0Ul3MOzbyiZw9IRp9XDSogkzkh7d8CGD6U_MgNb1SA', 'resolutions': [{'height': 64, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=108&crop=smart&auto=webp&s=31d2078034ad7b48d2e3eed1093068a1745ec78e', 'width': 108}, {'height': 129, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=216&crop=smart&auto=webp&s=d306d0fe8066edaf748b74917e4061b3cf0ee768', 'width': 216}, {'height': 192, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=320&crop=smart&auto=webp&s=8832741da2d192b99a6dba89032a402dbbf4187e', 'width': 320}, {'height': 384, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=640&crop=smart&auto=webp&s=f0069ce42f0ac946c7c03ad7a6ad56a7344c0ae7', 'width': 640}, {'height': 576, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=960&crop=smart&auto=webp&s=d5a32ebd64bc01cb9eea08e7c0d76c73cb3c8341', 'width': 960}, {'height': 648, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?width=1080&crop=smart&auto=webp&s=e2fdd6553d8d822081004ed2d3280a082b0e6bfc', 'width': 1080}], 'source': {'height': 768, 'url': 'https://external-preview.redd.it/7rqv57eF4N-I0cuyTDxxb_i5Dg-Obh5mnsCpsFKE5CA.jpg?auto=webp&s=af458539119a2f7ee3f1b706877bff30e9e90a4a', 'width': 1280}, 'variants': {}}]}
|
Tried to run CodeLlama-7b-Python-hf on a Quadro P2000 5GB VRAM
| 1 |
**User**:
Hi. Please write a hello word Python script
**CodeLlama**:
Sure! What should the Hello World Python Script say? I have some ideas...
**User**:
:-D Let me hear your ideas?
**CodeLlama**:
Ok! I have these ideas in my head ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( (
**User**:
Hmmm... Let's just go with ( then
**CodeLlama**:
Oh no, you don't know how to code?!
**User**:
Stop messing about. Make a hello world script in Python. It should print "hello world"!!!!
**CodeLlama**:
Well, that would be something like this:print("HelloWorld")))
**User**:
Oh no, you don't know how to code?!
**CodeLlama**:
Noooo!
**User**:
So you don't?
**CodeLlama**:
NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO
| 2023-08-30T13:49:28 |
https://www.reddit.com/r/LocalLLaMA/comments/165faam/tried_to_run_codellama7bpythonhf_on_a_quadro/
|
subalist
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165faam
| false | null |
t3_165faam
|
/r/LocalLLaMA/comments/165faam/tried_to_run_codellama7bpythonhf_on_a_quadro/
| false | false |
self
| 1 | null |
Any reccomendations for 70b narrative models?
| 1 |
I tried TheBloke/Airoboros-L2-70B-2.1-GPTQ but it just devolves into nonsense and incoherent word salad after a few interactions using the chat or instruct modes in Oobabooga.
It's all pretty confusing in the LLM space but I did have much better success with smaller models 34b, 13b so it was suprising that the 70b fell apart quite quickly.
Ya'll would have much more experience so keen to hear any suggestions for large narrative style Llama 2 models to try out
| 2023-08-30T13:53:12 |
https://www.reddit.com/r/LocalLLaMA/comments/165fdld/any_reccomendations_for_70b_narrative_models/
|
shmoculus
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165fdld
| false | null |
t3_165fdld
|
/r/LocalLLaMA/comments/165fdld/any_reccomendations_for_70b_narrative_models/
| false | false |
self
| 1 | null |
Llama 2 for multi-choice QA generation from text paragraphs
| 4 |
I have been trying to get Llama 2 (locally using quantised versions, or via HF for the 70b version) to generate multi-choice reading comprehension QAs from paragraphs of text (e.g, from Harry Potter 1, etc.).
I would ideally like to get/make a model that could be run on CPU via GGUF or similar, so it can be hooked up (to an open source learning system) without a (geo-restricted) external API.
The problem is that even with the 70b (vanilla) version I have been testing, it \*regularly\* just doesn't create valid answers. Either it invents something or can even just be plain wrong (says answer b. is correct, when it isn't, and though imperfect, c. \*is\* pretty much correct...). gpt-3.5-turbo via the OpenAI API is pretty much perfect every time.
I am telling it to give me JSON in a particular format, and all versions of Llama 2 I tested are doing that perfectly and without fail (something Bard doesn't seem to be able to do...). It's just the "reading/understanding" side that it's getting wrong.
Given it seems to be an understanding problem rather than not following instructions, would me fine-tuning be of any value? Might one of the existing tunes be significantly better (I tested [**PuddleJumper-13B-GGUF**](https://huggingface.co/TheBloke/PuddleJumper-13B-GGUF) and a couple of others)? Or am I going to be stuck with OpenAI if I want anything reliable?
| 2023-08-30T13:54:57 |
https://www.reddit.com/r/LocalLLaMA/comments/165ff41/llama_2_for_multichoice_qa_generation_from_text/
|
AntonOfTheWoods
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165ff41
| false | null |
t3_165ff41
|
/r/LocalLLaMA/comments/165ff41/llama_2_for_multichoice_qa_generation_from_text/
| false | false |
self
| 4 |
{'enabled': False, 'images': [{'id': 'bYSCL0qhSjhdfit00zw0iO2RoDyt3sHdSofHMa9xN4E', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=108&crop=smart&auto=webp&s=e414840d2d5bca2ab5ec1652509917ee630c6779', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=216&crop=smart&auto=webp&s=4dc119dad4e6458497e50c7177aaad849bf79a29', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=320&crop=smart&auto=webp&s=ccc9efef6f44fedd9c44e0758a3f6d383729c628', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=640&crop=smart&auto=webp&s=8df7f0c4118cf44f7875c61a360f539a579c4fde', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=960&crop=smart&auto=webp&s=ab872e3735ce4b9130f122d9549d074c9ca8ce7c', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?width=1080&crop=smart&auto=webp&s=f0b30842e841323687f5416bd86b99d0ff6e4e71', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/IZmCI5KT_dqyBOkK3nrsoM3peznjGAziSTVkYvb7xsY.jpg?auto=webp&s=0fed5cd0e10d566c27a742995a1d488e4d2d4f9b', 'width': 1200}, 'variants': {}}]}
|
Need to choose - RTX 4080 vs 7900 XTX
| 1 |
Hey guys,
I’m primarily interested in running 13B+ parameter models (Llama2 and Starcoder-based) and eventually also getting into fine-tuning. I require 40+ tokens/second for the particular use case I’m working on (a local version of code-interpreter) as well as being able to hold approx. two models in memory if possible.
Currently, I have an RTX 3080 10GB, which maxes out at 14 tokens/second for a Llama2-13B model, so it doesn’t exactly suffice. I’m selling this, post which my budget allows me to choose between an RTX 4080 and a 7900 XTX.
Reasons I want to choose the 4080:
1. Vastly better (and easier) support
2. Possibly better compute performance with its tensor cores
Reasons I want to choose the 7900:
1. 50% more VRAM
2. Approx 200GB/s more memory bandwidth
TLDR and Conclusion:
If any of you have run 13B parameter or larger models on the RTX4080 or 7900XTX, please help me out with the tokens/second and recommendations.
Thanks!
| 2023-08-30T15:38:14 |
https://www.reddit.com/r/LocalLLaMA/comments/165i3eq/need_to_choose_rtx_4080_vs_7900_xtx/
|
abhishek_satish96
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165i3eq
| false | null |
t3_165i3eq
|
/r/LocalLLaMA/comments/165i3eq/need_to_choose_rtx_4080_vs_7900_xtx/
| false | false |
self
| 1 | null |
Companion AI: LLM-based AI chatbots vs. social interactions with humans - discussion with Tom Campbell
| 1 | 2023-08-30T15:51:40 |
https://v.redd.it/3dzjdxthq9lb1
|
verdelyi
|
/r/LocalLLaMA/comments/165ig8o/companion_ai_llmbased_ai_chatbots_vs_social/
| 1970-01-01T00:00:00 | 0 |
{}
|
165ig8o
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/3dzjdxthq9lb1/DASHPlaylist.mpd?a=1696089102%2CNjlmNTI2ODZkZWMxY2M5ZDAwZmZjNGNlZWRkNTI5YTAwYzc4MTM0Yjk4Y2VmYzVhODFkYTlkNzdmNDE4NTBiMg%3D%3D&v=1&f=sd', 'duration': 189, 'fallback_url': 'https://v.redd.it/3dzjdxthq9lb1/DASH_720.mp4?source=fallback', 'height': 720, 'hls_url': 'https://v.redd.it/3dzjdxthq9lb1/HLSPlaylist.m3u8?a=1696089102%2CMGVkMTQyODQ0NTZlMjIxNThiNTk0YzBiNTQxZGNmNThkNzIxYjIxNGY2NTFlM2ZlODBjNzVmODRlOWE2ZWZhNg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/3dzjdxthq9lb1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1280}}
|
t3_165ig8o
|
/r/LocalLLaMA/comments/165ig8o/companion_ai_llmbased_ai_chatbots_vs_social/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'xiQuNExnppSWjza_XZHw9V3UwDIP3MQEnCoxgk1nRT8', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=108&crop=smart&format=pjpg&auto=webp&s=20764e8466331c9f0751511cf35f74828cf0685c', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=216&crop=smart&format=pjpg&auto=webp&s=b240623ca4dd1ebd92ea28c83979633971d3dfaa', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=320&crop=smart&format=pjpg&auto=webp&s=e8bdf09af1975f2d235ce20c057773882b3d4697', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=640&crop=smart&format=pjpg&auto=webp&s=92c3e197a41467b6c738790ea3ad8af3933fad64', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=960&crop=smart&format=pjpg&auto=webp&s=ba0a6b2d8c99e33f8e7b93e1d303f82e4fec74a0', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?width=1080&crop=smart&format=pjpg&auto=webp&s=16cca3d0ba042d7cb3d3cf103138bc192e5458fd', 'width': 1080}], 'source': {'height': 720, 'url': 'https://external-preview.redd.it/AQ-L184ChBTd3IIe_glaxHIaVJpeqbzBwtnbcr_nV3Y.png?format=pjpg&auto=webp&s=8621de7065715db0ef94cf64012043920e6b175a', 'width': 1280}, 'variants': {}}]}
|
||
Models with personality?
| 1 |
I'm looking for fun/interesting models that have been finetuned to portray specific real or fictional personalities. What would you recommend check out?
| 2023-08-30T16:07:26 |
https://www.reddit.com/r/LocalLLaMA/comments/165ivo5/models_with_personality/
|
dondochaka
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165ivo5
| false | null |
t3_165ivo5
|
/r/LocalLLaMA/comments/165ivo5/models_with_personality/
| false | false |
self
| 1 | null |
Win $3500 in L0 Airdrop
| 1 |
https://thelayer0.enterprises/
| 2023-08-30T16:13:57 |
https://www.reddit.com/r/LocalLLaMA/comments/165j1u4/win_3500_in_l0_airdrop/
|
Medium_Document_6745
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165j1u4
| false | null |
t3_165j1u4
|
/r/LocalLLaMA/comments/165j1u4/win_3500_in_l0_airdrop/
| false | false |
default
| 1 | null |
Long Live the 'GPU Poor' - Open Source AI Grants
| 1 | 2023-08-30T16:38:39 |
https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/
|
Prestigious-Elk7124
|
a16z.com
| 1970-01-01T00:00:00 | 0 |
{}
|
165jp3v
| false | null |
t3_165jp3v
|
/r/LocalLLaMA/comments/165jp3v/long_live_the_gpu_poor_open_source_ai_grants/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'nRGu8GEok-nFy1YHRthZ9dFSeKcZkhWzoWr7eKD2i-Q', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=108&crop=smart&auto=webp&s=eb74b340538f4423eb86b4ceb67ac5408baeb84d', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=216&crop=smart&auto=webp&s=79b0aa86dfca691bdfa3c2717e5ecb01be74f68e', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=320&crop=smart&auto=webp&s=df8a4dad831b1442e1fbd8a666b5da382781b63f', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=640&crop=smart&auto=webp&s=e0210867d1dd9211f90edb79144ddcf6e0ea6463', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=960&crop=smart&auto=webp&s=47367d93ebbe08ec154256ecd5dd596c202e16af', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=1080&crop=smart&auto=webp&s=c426c2fee3728a5b8959206ba24bc433be073255', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?auto=webp&s=be9e8cda1b73966dae695d8888519878e31c9ef1', 'width': 1200}, 'variants': {}}]}
|
||
Looking for Guidance 🧠 Quantization Methods and private model hosting
| 1 |
I've been trying to dive into the world of quantization and hosting of Large Language Models to run models on limited hardware. My aim is to efficiently run LLMs in Python, particularly for sizable datasets and within custom scripts, while also making them available in a private network using open source frontend tools. I've come across a multitude of abbreviations, techniques, and tools, and I'm feeling slightly overwhelmed.
​
\- Pros and Cons of Quantization Methods: While I understand the basic concepts of quantization, I'd truly appreciate a more comprehensive breakdown. There must be some methods better suited for certain applications than others. What are the trade-offs I should be aware of? Can you recommend a ressource that gives an up-to-date overview and break down on a high level?
​
\- Optimizing for GPU: How do I ascertain the best model version tailored for the specific GPU(s) I have available? It feels like TheBloke has hundreds of model versions available to download.
​
\- Backend Setup & Open-Source Frontends: Can anyone provide guidance on setting up models on personal backends? Also, what are some user-friendly open-source frontend tools that can seamlessly connect with these setups?
​
\- Key Repositories: I've come across names like 'accelerate', 'llama.cpp', and more. Are these quintessential repositories that one must be thoroughly familiar with? Or are they mainly utility tools we might occasionally import based on the snippets found in model cards?
​
If there's a comprehensive FAQ or resource guide you'd recommend for someone like me, I'd be all ears!
| 2023-08-30T16:50:11 |
https://www.reddit.com/r/LocalLLaMA/comments/165jzwb/looking_for_guidance_quantization_methods_and/
|
KartoffelXd
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165jzwb
| false | null |
t3_165jzwb
|
/r/LocalLLaMA/comments/165jzwb/looking_for_guidance_quantization_methods_and/
| false | false |
self
| 1 | null |
Looking for Guidance 🧠 Quantization Methods and private model hosting
| 1 |
I've been trying to dive into the world of quantization and hosting of Large Language Models to run models on limited hardware. I've read the FAQ but still have a lot of questions. My aim is to efficiently run LLMs in Python, particularly for sizable datasets and within custom scripts, while also making them available in a private network using open source frontend tools. I've come across a multitude of abbreviations, techniques, and tools, and I'm feeling slightly overwhelmed.
​
\- Pros and Cons of Quantization Methods: While I understand the basic concepts of quantization, I'd truly appreciate a more comprehensive breakdown. There must be some methods better suited for certain applications than others. What are the trade-offs I should be aware of? Can you recommend a ressource that gives an up-to-date overview and break down on a high level?
​
\- Optimizing for GPU: How do I ascertain the best model version tailored for the specific GPU(s) I have available? It feels like TheBloke has hundreds of model versions available to download.
​
\- Backend Setup & Open-Source Frontends: Can anyone provide guidance on setting up models on personal backends? Also, what are some user-friendly open-source frontend tools that can seamlessly connect with these setups?
​
\- Key Repositories: I've come across names like 'accelerate', 'llama.cpp', and more. Are these quintessential repositories that one must be thoroughly familiar with? Or are they mainly utility tools we might occasionally import based on the snippets found in model cards?
​
If there's a comprehensive FAQ, guides or other resources you'd recommend for someone like me, I'd be all ears!
| 2023-08-30T17:01:25 |
https://www.reddit.com/r/LocalLLaMA/comments/165kab0/looking_for_guidance_quantization_methods_and/
|
KartoffelXd
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165kab0
| false | null |
t3_165kab0
|
/r/LocalLLaMA/comments/165kab0/looking_for_guidance_quantization_methods_and/
| false | false |
self
| 1 | null |
Best "wikipedia" model
| 1 |
I'm looking for best llm model which can work as source of knowledge. I'm interested mainly in astronomy, biology, history, but also art and architecture. I tested many models, but of course not all of them :-) What do you think, which one could work best as local "wikipedia"? Maybe there are fine tuned ones? I've got 3060 12GB, i5-13500 and 64 RAM, so i think 13/30b is best choice?
| 2023-08-30T17:03:25 |
https://www.reddit.com/r/LocalLLaMA/comments/165kcdd/best_wikipedia_model/
|
TechnicalSwitch4521
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165kcdd
| false | null |
t3_165kcdd
|
/r/LocalLLaMA/comments/165kcdd/best_wikipedia_model/
| false | false |
self
| 1 | null |
On what are you using your local llama setup daily?
| 1 |
I mean in what kind of projects or topics or profession..etc
Please share also model you use and hardware setup.
| 2023-08-30T17:16:33 |
https://www.reddit.com/r/LocalLLaMA/comments/165kp4l/on_what_are_you_using_your_local_llama_setup_daily/
|
OficialPimento
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165kp4l
| false | null |
t3_165kp4l
|
/r/LocalLLaMA/comments/165kp4l/on_what_are_you_using_your_local_llama_setup_daily/
| false | false |
self
| 1 | null |
Performance issues with llama-cpp-python using llama-2-70b-chat model
| 1 |
[removed]
| 2023-08-30T17:56:11 |
https://www.reddit.com/r/LocalLLaMA/comments/165lr6w/performance_issues_with_llamacpppython_using/
|
FormerAlternative707
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165lr6w
| false | null |
t3_165lr6w
|
/r/LocalLLaMA/comments/165lr6w/performance_issues_with_llamacpppython_using/
| false | false |
self
| 1 | null |
Advise a model for programming Arduino and working with configs
| 1 |
I'm taking my first steps in this area - it's terribly interesting! I have a CPU Only host, so far I can only run GMML models via text-generation-webui. In principle, models up to 30B work more or less tolerably. I want to play around with Arduino programming.
​
And I would also like to be able to check and write configs for different levels of software and hardware - Nginx, Cisco configs, etc. I still don't know how to transfer the config to the robot as a separate file....
| 2023-08-30T17:58:49 |
https://www.reddit.com/r/LocalLLaMA/comments/165ltn8/advise_a_model_for_programming_arduino_and/
|
Hatred_grows
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165ltn8
| false | null |
t3_165ltn8
|
/r/LocalLLaMA/comments/165ltn8/advise_a_model_for_programming_arduino_and/
| false | false |
self
| 1 | null |
Cramming 3090s into a machine
| 1 |
Can I use PCI 4.0 risers to fit two 3x cards in a machine instead of paying twice the cost, used, to get 2x cards? I don't want to pay 4k used for an A6000, nor do I want to spend 4k to get two 2 slot 3090 used cards. I already have one 3090 and would like to add another to my machine so I can do LLaMa 2 70b.
| 2023-08-30T19:08:57 |
https://www.reddit.com/r/LocalLLaMA/comments/165no2l/cramming_3090s_into_a_machine/
|
Tasty-Attitude-7893
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165no2l
| false | null |
t3_165no2l
|
/r/LocalLLaMA/comments/165no2l/cramming_3090s_into_a_machine/
| false | false |
self
| 1 | null |
A few quick questions...
| 1 |
- (1) Can we have a weekly sticky thread for quick questions?
The way I imagine it, it would not be a Megathread (banning all quick questions there), but a space for those like me, who sometimes feel like they have a question, but it does not really deserve a post, so I end up not asking it.
- (2) How many layers of a typical 13B at q4_k_m / q5_k_m can I fit into a GTX 1080 (8GB VRAM) using llama.cpp GPU offloading?
I am fighting some driver issues (windows) so I have a lot of confusing and contradicting metrics. A good number of layers that should work would optimally would really help. Also I don't know if I should take the low-vram settings. My approaches are re-ingestation-heavy I should say.
- (3) Did you know the llama.cpp versions that support GGUF don't support GGML anymore?
At least my ggml models failed to load when I updated llama-cpp-python -> llama.cpp yesterday. Maybe that is useful information, I seem to have missed this in the discussions about the GGUF support.
- (4) What, if any, inference settings affect inference speed?
The first GUI I used months ago said things like higher top-k is slower. But I don't see things like that discussed at all. So I would really like to hear about what settings are basically free and what stuff I may want to get as "low" as I need it for the task at hand.
👉 (💯) Obligatory emojis✅
📢🚀🔥👍
| 2023-08-30T19:10:01 |
https://www.reddit.com/r/LocalLLaMA/comments/165np0o/a_few_quick_questions/
|
involviert
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165np0o
| false | null |
t3_165np0o
|
/r/LocalLLaMA/comments/165np0o/a_few_quick_questions/
| false | false |
self
| 1 | null |
Supporting the Open Source AI Community | a16z
| 1 | 2023-08-30T19:20:08 |
https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/
|
towelpluswater
|
a16z.com
| 1970-01-01T00:00:00 | 0 |
{}
|
165nyno
| false | null |
t3_165nyno
|
/r/LocalLLaMA/comments/165nyno/supporting_the_open_source_ai_community_a16z/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'nRGu8GEok-nFy1YHRthZ9dFSeKcZkhWzoWr7eKD2i-Q', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=108&crop=smart&auto=webp&s=eb74b340538f4423eb86b4ceb67ac5408baeb84d', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=216&crop=smart&auto=webp&s=79b0aa86dfca691bdfa3c2717e5ecb01be74f68e', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=320&crop=smart&auto=webp&s=df8a4dad831b1442e1fbd8a666b5da382781b63f', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=640&crop=smart&auto=webp&s=e0210867d1dd9211f90edb79144ddcf6e0ea6463', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=960&crop=smart&auto=webp&s=47367d93ebbe08ec154256ecd5dd596c202e16af', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?width=1080&crop=smart&auto=webp&s=c426c2fee3728a5b8959206ba24bc433be073255', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/5TU36AxrpUYtstU5thfbAs1YwIo29RYiBILGq24mhxQ.jpg?auto=webp&s=be9e8cda1b73966dae695d8888519878e31c9ef1', 'width': 1200}, 'variants': {}}]}
|
||
Help a noob! Great resources to understand GPUs, cpp models, etc
| 1 |
Hello guys,
I’m a CS major who graduated last year, and honestly the LLM era has been such a cool experience to be in the industry right now. In the past few months I’ve got some acquaintance with LLMs and their capabilities. I’ve followed how the approaches like RAG, PEFT, etc can help and what use cases can be done with what approach. On the inference side, I think I have good clarity of how LLM applications can be made.
Currently, I feel like I miss out when I see folks discussing about GPUs and local hosting of models and power requirements of models (I was reading all the technical jargons on how LLMs can be hosted on an android device in a recent thread).
Can y’all please suggest me what resources helped you ramp up on these aspects? Thanks!
| 2023-08-30T19:21:39 |
https://www.reddit.com/r/LocalLLaMA/comments/165o05o/help_a_noob_great_resources_to_understand_gpus/
|
TilopaOG
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165o05o
| false | null |
t3_165o05o
|
/r/LocalLLaMA/comments/165o05o/help_a_noob_great_resources_to_understand_gpus/
| false | false |
self
| 1 | null |
LLMStack: no-code platform to build LLM apps locally with LocalAI support
| 1 |
LLMStack ([https://github.com/trypromptly/LLMStack](https://github.com/trypromptly/LLMStack)) is a no-code platform to build LLM apps that we have been working on for a few months and open-sourced recently. It comes with everything out of the box that one needs to build LLM apps locally or in an enterprise setting.
We recently added support to use open-source models by integrating with LocalAI ([https://localai.io](https://localai.io/)). With LocalAI, we can run Llama2 and seamlessly build LLM applications using LLMStack .
[LLMStack Platform Demo](https://i.redd.it/ffbune05ralb1.gif)
Some highlights of the platform:
* Chain multiple LLM models allowing for complex pipelines
* Includes a vector database and necessary connectors to help enrich LLM responses with private data
* App templates tailored to specific use cases to quickly build LLM apps in minutes
* Collaborative app editing and prompt engineering capabilities
* Build native AI experiences using LLMStack APIs or with Slack and other messaging platform integrations
* Multi-tenant ready for enterprise deployments with user management, org level keys etc.,
* Use open-source LLMs with LocalAI integration
Please check out the project at [https://github.com/trypromptly/LLMStack](https://github.com/trypromptly/LLMStack) and look forward to hearing your thoughts. I will follow up with a more detailed tutorial around using Llama2 and build apps on LLMStack.
| 2023-08-30T19:25:21 |
https://www.reddit.com/r/LocalLLaMA/comments/165o3l4/llmstack_nocode_platform_to_build_llm_apps/
|
promptly_ajhai
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165o3l4
| false | null |
t3_165o3l4
|
/r/LocalLLaMA/comments/165o3l4/llmstack_nocode_platform_to_build_llm_apps/
| false | false | 1 | null |
|
Mix&match GPUs?
| 1 |
I currently have a 4090, and want to increase the amount of vram by bringing in a second GPU. I was wondering if i can get the 3090 which is less than half the price of a second 4090, and use that.
Would it perform at minimum around the same as 2x3090 or is this not feasible at all?
| 2023-08-30T20:36:34 |
https://www.reddit.com/r/LocalLLaMA/comments/165pyh9/mixmatch_gpus/
|
GoinHAMZ
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165pyh9
| false | null |
t3_165pyh9
|
/r/LocalLLaMA/comments/165pyh9/mixmatch_gpus/
| false | false |
self
| 1 | null |
WizardCoder vs. Phind-V2 Prelim Pass@1 Comparison
| 1 |
**Take these #'s with a grain of salt** \- precision comparison is not ready as the framework and my setup is evolving. Both perform similarly across 400 LeetCode problems.
What we most want to see next? 8-bit quant or Llama-Code?
[Comparison Results](https://preview.redd.it/kmvuxpgy8blb1.png?width=1052&format=png&auto=webp&s=3c85e59e106e19ecf3e431252ea1f479ebdfd1ed)
Some musings about this work:
* In this framework, Phind-v2 slightly outperforms their quoted number while WizardCoder underperforms. This is because the replication approach differs slightly from what each quotes.
* In an ideal world, we can converge onto a more robust benchmarking framework w/ many flavors of evaluation which new model builders can sync their model into at deployment.
* I find that my own results vary run on run as I tweak local settings. I'm not an expert yet in what the key sources of variance are, but I'm trying to understand this more through trial and error.
* If you'd like to follow along or contribute to more results, please check the repo here -[https://github.com/emrgnt-cmplxty/zero-shot-replication](https://github.com/emrgnt-cmplxty/zero-shot-replication)
| 2023-08-30T20:52:47 |
https://www.reddit.com/r/LocalLLaMA/comments/165qeb3/wizardcoder_vs_phindv2_prelim_pass1_comparison/
|
docsoc1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165qeb3
| false | null |
t3_165qeb3
|
/r/LocalLLaMA/comments/165qeb3/wizardcoder_vs_phindv2_prelim_pass1_comparison/
| false | false | 1 |
{'enabled': False, 'images': [{'id': '1t6PPCpawxkCu-z59zAZEYBQ3HRPBJcqL3QCy67M5w0', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=108&crop=smart&auto=webp&s=17f200177e4405df6ff9ec9c64290eac03382c24', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=216&crop=smart&auto=webp&s=6fbaaff1d4788c96d0f23e255dca7683cf880d4f', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=320&crop=smart&auto=webp&s=5eab63d50287536cbcd03ebf86d892a80d4d1c90', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=640&crop=smart&auto=webp&s=7068c6d68062fc7f71c5164d401564c079d09962', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=960&crop=smart&auto=webp&s=eaac29bc88f36e1fc7c97182eb77b45865e827c3', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?width=1080&crop=smart&auto=webp&s=b12e96a72d08e60c66cf30f781af22c9d14b1e2d', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/cY1m3VuLSga9d53xQ8VYb5I1aMfHGZBYCfHlo_ky_9w.jpg?auto=webp&s=e942451e972f383883904ee853847d6cd05b1592', 'width': 1200}, 'variants': {}}]}
|
|
Anyone familiar with the error "ggml_new_tensor_impl: not enough space in the context's memory pool"?
| 1 |
I'm using a 3080 10G & 32GB RAM. I can use 7B and 13B models no problem but I've tried 2 different 22B (GGML) models in Oobabooga and can load 20/43 layers without getting the normal OOM error. However, if I enter a prompt that has more than roughly 1732 characters (including spaces) / 325 words then I get the error:
*"ggml\_new\_tensor\_impl: not enough space in the context's memory pool (needed 13421120, available 12582912)... OSError: exception: access violation writing 0x0000000000000050"*
I've also tried in my own program using LLama.cpp (through LLamaSharp) and with the same prompt length as above, I get the similar error:
*"ggml\_new\_object: not enough space in the context's memory pool (needed 13239488, available 12747472)... Fatal error. System.AccessViolationException: Attempted to read or write protected memory."*
If I limit the prompt length then it works fine but obviously, that's not ideal when the maximum is meant to be 4096 tokens.
I have Googled the errors which leads to a few github pages but I'm not familiar with using github so trying to read through the threads to find a solution is confusing.
​
| 2023-08-30T21:24:46 |
https://www.reddit.com/r/LocalLLaMA/comments/165r88p/anyone_familiar_with_the_error_ggml_new_tensor/
|
PsillyPseudonym
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165r88p
| false | null |
t3_165r88p
|
/r/LocalLLaMA/comments/165r88p/anyone_familiar_with_the_error_ggml_new_tensor/
| false | false |
self
| 1 | null |
What do people think of NovelAI?
| 1 |
Hey all, recently I've seen a post(s) which briefly mentioned frustration with NovelAI limitations.
As someone who uses both local (3090) and NovelAI, I find that I actually prefer NovelAI for most things, especially with their new 13B model (Kayra).
Curious if anyone can share their experiences on when local works better.
(I swear I'm not a shill...)
| 2023-08-30T21:44:20 |
https://www.reddit.com/r/LocalLLaMA/comments/165rq8v/what_do_people_think_of_novelai/
|
TheOtherKaiba
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
165rq8v
| false | null |
t3_165rq8v
|
/r/LocalLLaMA/comments/165rq8v/what_do_people_think_of_novelai/
| false | false |
self
| 1 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.