title stringlengths 1 300 | score int64 0 8.54k | selftext stringlengths 0 41.5k | created timestamp[ns]date 2023-04-01 04:30:41 2026-03-04 02:14:14 ⌀ | url stringlengths 0 878 | author stringlengths 3 20 | domain stringlengths 0 82 | edited timestamp[ns]date 1970-01-01 00:00:00 2026-02-19 14:51:53 | 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 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Help Needed | 1 | [removed] | 2023-12-23T02:32:41 | https://www.reddit.com/r/LocalLLaMA/comments/18ovk3x/help_needed/ | Dragneel_passingby | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ovk3x | false | null | t3_18ovk3x | /r/LocalLLaMA/comments/18ovk3x/help_needed/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'eK-F_ltJeSQkEhSuvQ1WynvJ98Mi9uD4sGowU8DVC8k', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/cMiPHVa-MlEHpM7I-oXDxBVnueQXxegDHU0DC_E5A4Y.jpg?width=108&crop=smart&auto=webp&s=ec3208d73ce82ea73a6324da08a556492c2af74c', 'width': 108}, {'height': 162, 'url': 'h... |
My setup for using ROCm with RX 6700XT GPU on Linux | 31 | Some people have asked me to share my setup for running LLMs using ROCm, so here I am with a guide (sorry I'm late). I chose the RX 6700XT GPU for myself was because I figured it's a relatively cheap GPU with 12GB VRAM and decent performance (related discussion is here if anyone is interested: [https://www.reddit.com/r... | 2023-12-23T01:50:25 | https://www.reddit.com/r/LocalLLaMA/comments/18ourt4/my_setup_for_using_rocm_with_rx_6700xt_gpu_on/ | AgeOfAlgorithms | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ourt4 | false | null | t3_18ourt4 | /r/LocalLLaMA/comments/18ourt4/my_setup_for_using_rocm_with_rx_6700xt_gpu_on/ | false | false | self | 31 | null |
Dual k80 performance? | 2 | When I was into cryptomining there were mining calculators that would estimate what certain cards would get you over time. Is there a database of token per second performance on dirrent gpus and models?
Interested in building a sub-500 machine for miscellansous stuff and want to be able to run big things. | 2023-12-23T01:50:08 | https://www.reddit.com/r/LocalLLaMA/comments/18ourlm/dual_k80_performance/ | thebadslime | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ourlm | false | null | t3_18ourlm | /r/LocalLLaMA/comments/18ourlm/dual_k80_performance/ | false | false | self | 2 | null |
Repetition Sampler Settings -- Thoughts? | 5 | What does everyone prefer to use for their repetition sampler settings, especially through SillyTavern? We have repetition penalty, frequency penalty, presence penalty, and no-repeat ngram size to work with. Do you prefer to run just one or do you favor a combination? Have you found a killer combination that produces g... | 2023-12-23T01:47:06 | https://www.reddit.com/r/LocalLLaMA/comments/18oupiv/repetition_sampler_settings_thoughts/ | sophosympatheia | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oupiv | false | null | t3_18oupiv | /r/LocalLLaMA/comments/18oupiv/repetition_sampler_settings_thoughts/ | false | false | self | 5 | null |
mistral-ft-optimized feels like another large step up. | 126 | It's super good at everything and very obedient, my new fav:
I love how they start by saying that model merging seems like a really weird and counterintuitive idea which you would not think would work at-all well (exactly how I feel)
but then they go on to say its actually crazy good and they tried 9000 combinations ... | 2023-12-23T01:12:08 | https://www.reddit.com/r/LocalLLaMA/comments/18ou1n7/mistralftoptimized_feels_like_another_large_step/ | Revolutionalredstone | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ou1n7 | false | null | t3_18ou1n7 | /r/LocalLLaMA/comments/18ou1n7/mistralftoptimized_feels_like_another_large_step/ | false | false | self | 126 | {'enabled': False, 'images': [{'id': 'J4DgURlSpfuA2MW-ft5Zc63DOoo8QihEpKK3v1GsXCk', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/KcGVocZiPjSuyLkFZ_d1WPd_6R9KLcGFq7ioUprT990.jpg?width=108&crop=smart&auto=webp&s=80144a2387615e620dee421df127713c022972e2', 'width': 108}, {'height': 117, 'url': 'h... |
p40/p100 | 2 | what's giving more performance right now a p100 running exllama2/fp16 or p40 running whatever it is it runs?
​
also can you mix them for inferencing like you see people doing on consumer cards??
​
Thanks in advance.
​
[View Poll](https://www.reddit.com/poll/18ou0pc) | 2023-12-23T01:10:50 | https://www.reddit.com/r/LocalLLaMA/comments/18ou0pc/p40p100/ | Dyonizius | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ou0pc | false | null | t3_18ou0pc | /r/LocalLLaMA/comments/18ou0pc/p40p100/ | false | false | self | 2 | null |
Anyone got some good resources for working with LLama CPP on python? | 4 | The documentation here:https://llama-cpp-python.readthedocs.io/en/latest/#adjusting-the-context-window
is pretty bare bones, and I'm trying to figure out the class attributes and stuff.
I'm wondering if anyone knows of anything that could help disambiguate everything, it's fun to dive through, but there must be somet... | 2023-12-23T01:02:04 | https://www.reddit.com/r/LocalLLaMA/comments/18otumd/anyone_got_some_good_resources_for_working_with/ | Nano_9a9o | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18otumd | false | null | t3_18otumd | /r/LocalLLaMA/comments/18otumd/anyone_got_some_good_resources_for_working_with/ | false | false | self | 4 | null |
What tools are you using to run LLM's locally? | 16 | I set up ollama on my mac and I am running Mistral 7B and Microsoft Phi (2.7B) locally on it.
I was going to run it on my PC but also curious what other tooling is available to training/RAG so I can do more than just run the model.
What tooling are you using? LM Studio? PowerInfer? Llama.cpp? Talk about what you are... | 2023-12-22T23:57:49 | https://www.reddit.com/r/LocalLLaMA/comments/18oskbm/what_tools_are_you_using_to_run_llms_locally/ | HarkonnenSpice | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oskbm | false | null | t3_18oskbm | /r/LocalLLaMA/comments/18oskbm/what_tools_are_you_using_to_run_llms_locally/ | false | false | self | 16 | null |
Mixtral MoE ELI5: How are the responses a higher quality than a 7b? | 74 | I have two general high level understandings of Mixtral: It is 8 x 7b models, and that it uses two at a time to improve performance, so instead of running like a 56b it's really running like a 14b.
Now, I've seen lots of people saying it is producing higher quality results than Llama 2 70b... but I can't understand th... | 2023-12-22T23:53:14 | https://www.reddit.com/r/LocalLLaMA/comments/18osgzt/mixtral_moe_eli5_how_are_the_responses_a_higher/ | SomeOddCodeGuy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18osgzt | false | null | t3_18osgzt | /r/LocalLLaMA/comments/18osgzt/mixtral_moe_eli5_how_are_the_responses_a_higher/ | false | false | self | 74 | null |
Issue converting a Falcon 7-b based model from HuggingFace to gguf with llama.cpp. | 5 | Hello everyone. Thanks for taking the time to read my post.
I'm new to this field, so please be easy on me. I was trying to use the only spanish focused model I found "[Aguila-7b](https://huggingface.co/projecte-aina/aguila-7b)" as base model for localGPT, in order to experiment with some legal pdf documents (I'm a la... | 2023-12-22T23:44:41 | https://www.reddit.com/r/LocalLLaMA/comments/18osana/issue_converting_a_falcon_7b_based_model_from/ | Christ0ph_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18osana | false | null | t3_18osana | /r/LocalLLaMA/comments/18osana/issue_converting_a_falcon_7b_based_model_from/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'Gf1BvLYso-ZWo2o3aBoX2rQE2MbDdXKTyVfKwL7HnBU', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/RQ6XtgidynTC1HY9x-n8zVQyVkzuQZvOwGDoCpDxVko.jpg?width=108&crop=smart&auto=webp&s=080ac2e9765b7afcb10b841eb788e05fb9d33d7a', 'width': 108}, {'height': 116, 'url': 'h... |
Mistral Medium on Poe | 7 | Seems to have little to no censorship (gave an audience suitability warning before writing NSFW content) [https://poe.com/Mistral-Medium](https://poe.com/Mistral-Medium) | 2023-12-22T23:43:44 | https://www.reddit.com/r/LocalLLaMA/comments/18os9zf/mistral_medium_on_poe/ | WaterdanceAC | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18os9zf | false | null | t3_18os9zf | /r/LocalLLaMA/comments/18os9zf/mistral_medium_on_poe/ | false | false | self | 7 | null |
In a test, Mixtral Q4_0 is better than Llama 2 70B, Gemini Pro, Claude 2 and is on par with GPT-4! This model has made me optimistic about the future of open-source LLMs | 6 | 2023-12-22T23:23:41 | https://www.reddit.com/gallery/18orv70 | nderstand2grow | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18orv70 | false | null | t3_18orv70 | /r/LocalLLaMA/comments/18orv70/in_a_test_mixtral_q4_0_is_better_than_llama_2_70b/ | false | false | 6 | null | ||
Anyone have experience running mistral 8x7b with 48 gb of unified memory? | 5 | For anyone with the standard M3 Max 48 gb, I was wondering if you could share your experience running mistral 8x7b using metal. From what I’ve read, Apple seems to limit the amount of vram to 75% of the total amount of unified memory, so I’m assuming 36 gb will available, which is enough to run the 5 bit quantization, ... | 2023-12-22T23:14:27 | https://www.reddit.com/r/LocalLLaMA/comments/18oro1i/anyone_have_experience_running_mistral_8x7b_with/ | complexityzero | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oro1i | false | null | t3_18oro1i | /r/LocalLLaMA/comments/18oro1i/anyone_have_experience_running_mistral_8x7b_with/ | false | false | self | 5 | null |
oobabooga/webui/openapi: is it possible to use "start reply with" with the API? | 1 | Hi, I'm trying to use the API from oobabooga's repo.
In the webui, there is a way to force the model to start a sentence with some tokens, it can be used in all modes (chat, chat-instruct, instruct). Do users of this API know if there's a way to easily do this with [the API](https://github.com/oobabooga/text-generatio... | 2023-12-22T22:49:08 | https://www.reddit.com/r/LocalLLaMA/comments/18or4lz/oobaboogawebuiopenapi_is_it_possible_to_use_start/ | IntelArtiGen | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18or4lz | false | null | t3_18or4lz | /r/LocalLLaMA/comments/18or4lz/oobaboogawebuiopenapi_is_it_possible_to_use_start/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'f2AC_sInspxC58l52vL1a3i-BvmDYOFsEDjsMZmszcU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/9Qrgl6lEDen3-J2d3cBZBrvrwBfYHXxyacdPN-m3SCc.jpg?width=108&crop=smart&auto=webp&s=7b756860693980e1386ac4e8c6778adc8f4d80af', 'width': 108}, {'height': 108, 'url': 'h... |
LlamaCPP python server vs LlamaCPP python class | 1 | Hi all,
I noticed a significant difference in performance when between using the api of LlamaCPP python server and the llamaCPP python class (llm = llamaCPP{}) when using the same model.
Using the API is so much faster.
Why is this, or am I missing something in the configuration of the class?
Ps: I’m using CPU only... | 2023-12-22T22:17:38 | https://www.reddit.com/r/LocalLLaMA/comments/18oqgpr/llamacpp_python_server_vs_llamacpp_python_class/ | e-nigmaNL | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oqgpr | false | null | t3_18oqgpr | /r/LocalLLaMA/comments/18oqgpr/llamacpp_python_server_vs_llamacpp_python_class/ | false | false | self | 1 | null |
I downloaded dolphin-mixtral:latest on my mac, is there any way to delete it? | 1 | I downloaded it to see how it was but now that I am finished with it, I can't find a way to delete it. | 2023-12-22T21:48:47 | https://www.reddit.com/r/LocalLLaMA/comments/18opudc/i_downloaded_dolphinmixtrallatest_on_my_mac_is/ | fungusdickhead | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18opudc | false | null | t3_18opudc | /r/LocalLLaMA/comments/18opudc/i_downloaded_dolphinmixtrallatest_on_my_mac_is/ | false | false | self | 1 | null |
Does ChatGPT/GPT4 give accurate answers for the subject of LLM models and running them? Or is the subject too new & answers are unreliable? | 1 | I've been asking it a lot of questions about how models work, various concepts/things mentioned in the docs of models or in their config files - but those of you who actually know what you're talking about - does it give accurate answers on these topics or are they too new for it to be able to provide useful/reliable i... | 2023-12-22T21:41:25 | https://www.reddit.com/r/LocalLLaMA/comments/18opoqi/does_chatgptgpt4_give_accurate_answers_for_the/ | TopRecognition9302 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18opoqi | false | null | t3_18opoqi | /r/LocalLLaMA/comments/18opoqi/does_chatgptgpt4_give_accurate_answers_for_the/ | false | false | self | 1 | null |
Does it make sense to pair an RTX 4090 with a 16GB 4060TI to get 40GB of VRAM? | 13 | Exactly what it says on the tin. Has anyone tried this? How is model performance affected when split across 2 different GPUs?
Looking at the price of 3090s this seems to be the least expensive way to increase my model capacity locally. | 2023-12-22T21:40:18 | https://www.reddit.com/r/LocalLLaMA/comments/18opntr/does_it_make_sense_to_pair_an_rtx_4090_with_a/ | antialtinian | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18opntr | false | null | t3_18opntr | /r/LocalLLaMA/comments/18opntr/does_it_make_sense_to_pair_an_rtx_4090_with_a/ | false | false | self | 13 | null |
VRAM vs raw computing power vs other features - how do these impact performance? What aspects are significant and when? | 9 | I'm new to this and trying to understand the significance of various aspects of how running these local LLaMA's works.
For example is VRAM only needed for speed? OR does processing power also matter? Does processing power for each card individually matter - or only the primary one? (e.g. 3 x p40 vs 1x3090+2p40 vs 3x30... | 2023-12-22T21:31:54 | https://www.reddit.com/r/LocalLLaMA/comments/18ophi2/vram_vs_raw_computing_power_vs_other_features_how/ | TopRecognition9302 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ophi2 | false | null | t3_18ophi2 | /r/LocalLLaMA/comments/18ophi2/vram_vs_raw_computing_power_vs_other_features_how/ | false | false | self | 9 | null |
Are there any guides for using triton inference server with something like Oobabooga? | 2 | I’d like to try using triton to centralize my models and maybe make better use of resources but I can’t seem to find any specific info on setting it up. | 2023-12-22T21:28:10 | https://www.reddit.com/r/LocalLLaMA/comments/18opejx/are_there_any_guides_for_using_triton_inference/ | Jcwscience | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18opejx | false | null | t3_18opejx | /r/LocalLLaMA/comments/18opejx/are_there_any_guides_for_using_triton_inference/ | false | false | self | 2 | null |
Good model for code generation? | 6 | Out of the following list: codellama, phind-codellama, wizardcoder, deepseek-coder, codeup & starcoder.
​
Which model out of this list would be the best for code generation? More specifically, (modern) PHP and its Laravel framework, JavaScript and some styling (TailwindCSS and such).
​
Thanks in adva... | 2023-12-22T21:07:45 | https://www.reddit.com/r/LocalLLaMA/comments/18ooz25/good_model_for_code_generation/ | s7mbre | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ooz25 | false | null | t3_18ooz25 | /r/LocalLLaMA/comments/18ooz25/good_model_for_code_generation/ | false | false | self | 6 | null |
GPUs and Local Large Language models, Are there Possible Tricks for the big ones? AMD vs Nvidia? NVLink? Rocm? XGMI? | 3 | I'd like to know what you guys tried...
I want to someday develop an AI powered project using specifically one or a few of the mistral-based models (probably the Mixtral 8x7B at this point) and in order to develop it I wanted to figure out what to buy really.
Budget is not great, but despite that a bit of waiting and... | 2023-12-22T21:04:26 | https://www.reddit.com/r/LocalLLaMA/comments/18oowgq/gpus_and_local_large_language_models_are_there/ | gabrielesilinic | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oowgq | false | null | t3_18oowgq | /r/LocalLLaMA/comments/18oowgq/gpus_and_local_large_language_models_are_there/ | false | false | self | 3 | null |
How To Finetune Mixtral On Consumer Hardware | 1 | [removed] | 2023-12-22T20:28:43 | https://www.reddit.com/r/LocalLLaMA/comments/18oo4qy/how_to_finetune_mixtral_on_consumer_hardware/ | l33thaxman | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oo4qy | false | null | t3_18oo4qy | /r/LocalLLaMA/comments/18oo4qy/how_to_finetune_mixtral_on_consumer_hardware/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '5TBLrZK3iIM1cLJevLBm4J6h-X_V8kgyzetDkZyX2eY', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/rVVy07OBzXJzNvCcMHGN3qzkz3qAkYb6gjhFOTy5waE.jpg?width=108&crop=smart&auto=webp&s=2437d55dea74ed25a45026a47bc800fad06bd066', 'width': 108}, {'height': 162, 'url': 'h... |
Cheapest 7b GPU Inference Setup? | 6 | Right now I'm using runpod, colab or inference APIs for GPU inference. I've tried CPU inference and it's a little too slow for my use cases.
If I had to put together a PC purely for GPU inference (7b models), what's the cheapest setup I can have? Cheapest = both in terms of purchase cost and power utilization.
I was ... | 2023-12-22T20:11:54 | https://www.reddit.com/r/LocalLLaMA/comments/18onsb6/cheapest_7b_gpu_inference_setup/ | noellarkin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18onsb6 | false | null | t3_18onsb6 | /r/LocalLLaMA/comments/18onsb6/cheapest_7b_gpu_inference_setup/ | false | false | self | 6 | null |
LLM to Brainstorm Videogame Quests (Rtx 4090) | 2 | Hello,
I'm quite new at LLMs, and I have a usecase where I'd like a local model to help brainstorm quests, characters, and maybe make a first draft of dialogue for characters. The game is serious, though it is not 18+ NSFW.
My computer has a 4090 24gb VRAM. Offloading to GPU gets me an OK token speed.
I've just star... | 2023-12-22T19:46:41 | https://www.reddit.com/r/LocalLLaMA/comments/18on8qx/llm_to_brainstorm_videogame_quests_rtx_4090/ | PwanaZana | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18on8qx | false | null | t3_18on8qx | /r/LocalLLaMA/comments/18on8qx/llm_to_brainstorm_videogame_quests_rtx_4090/ | false | false | self | 2 | null |
Anyone working on a "tiny" version of Mixtral-8x7b | 6 | Hi everyone! I'm very excited about the potential use cases for LLMs running locally on mobile devices and I'm curious to know if anyone knows of any efforts to make a "tiny" version of Mixtral-8x7b, similar to \[TinyLlama\]([https://github.com/jzhang38/TinyLlama?tab=readme-ov-file](https://github.com/jzhang38/TinyLlam... | 2023-12-22T19:43:19 | https://www.reddit.com/r/LocalLLaMA/comments/18on651/anyone_working_on_a_tiny_version_of_mixtral8x7b/ | bonthebruh | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18on651 | false | null | t3_18on651 | /r/LocalLLaMA/comments/18on651/anyone_working_on_a_tiny_version_of_mixtral8x7b/ | false | false | self | 6 | {'enabled': False, 'images': [{'id': 'MjBjWvUxV4ERbTDugierSndbZWXi65aMw55GPNFbiuU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/YLgyZxp-fSC3Dqanw7F0W_mdAEvtM6ZhqfuRXmApItM.jpg?width=108&crop=smart&auto=webp&s=cb34deef57d799efd0c0fdd9d17c1f0bb0f20ec2', 'width': 108}, {'height': 108, 'url': 'h... |
Run Mixtral LLM locally in seconds with Ollama! | 1 | [removed] | 2023-12-22T19:40:32 | https://www.reddit.com/r/LocalLLaMA/comments/18on421/run_mixtral_llm_locally_in_seconds_with_ollama/ | dev-spot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18on421 | false | null | t3_18on421 | /r/LocalLLaMA/comments/18on421/run_mixtral_llm_locally_in_seconds_with_ollama/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '6NFx-wVE7S1a2iiGNyMj-XyeMr6aJj59vIlIcbRspoA', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/Razh7uBASwKlxOqfSwYoEOA4vK1LzTvZvA8L9Cx0aLc.jpg?width=108&crop=smart&auto=webp&s=4fd343d6af76e0d166db35125d2c2db57505d68b', 'width': 108}, {'height': 162, 'url': 'h... |
How do I enable longer responses? | 1 | I'm a complete noob and I'm sorry if this is a simple question, but I can't seem to figure out the answer.
I have llama.cpp running locally with the Mixtral-8x7B-Instruct-v0.1-GGUF model. It's doing amazing. But when I give it a prompt with a lot of text (i.e. a newsletter article) and ask it to re-write the arti... | 2023-12-22T19:22:06 | https://www.reddit.com/r/LocalLLaMA/comments/18omq7d/how_do_i_enable_longer_responses/ | FixMoreWhineLess | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18omq7d | false | null | t3_18omq7d | /r/LocalLLaMA/comments/18omq7d/how_do_i_enable_longer_responses/ | false | false | self | 1 | null |
Best 7b model for RAG use cases | 1 | [removed] | 2023-12-22T19:04:28 | https://www.reddit.com/r/LocalLLaMA/comments/18omcqs/best_7b_model_for_rag_use_cases/ | faridukhan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18omcqs | false | null | t3_18omcqs | /r/LocalLLaMA/comments/18omcqs/best_7b_model_for_rag_use_cases/ | false | false | self | 1 | null |
Ditch ChatGPT and try one of these downloadable chatbot options | 1 | [removed] | 2023-12-22T18:37:50 | https://www.reddit.com/r/LocalLLaMA/comments/18olr9u/ditch_chatgpt_and_try_one_of_these_downloadable/ | heaversm | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18olr9u | false | null | t3_18olr9u | /r/LocalLLaMA/comments/18olr9u/ditch_chatgpt_and_try_one_of_these_downloadable/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'BGXfaUMntPZWzYo99FfkrbqxveeayLkICP2FRV6iEYA', 'resolutions': [{'height': 41, 'url': 'https://external-preview.redd.it/V0bMuIoeIfXcCnG6uE4-9PiNgKWrVXkZZsdOyfiZwH0.jpg?width=108&crop=smart&auto=webp&s=027a4a136867b336cfe98d6c39e0251e1ba0b055', 'width': 108}, {'height': 83, 'url': 'ht... |
Mistral Instruct Finetune Dataset Format | 1 | I'm sorry for this very newbie question, I want to know if my dataset is in correct format for finetuning mistral 7b instruct v1 model or not, I'm using together ai provider so it must following jsonl format with {"text":...}. My current dataset looks like this inside of the "text" key value, is it correct format ?
... | 2023-12-22T18:21:32 | https://www.reddit.com/r/LocalLLaMA/comments/18oldy7/mistral_instruct_finetune_dataset_format/ | laveriaroha | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oldy7 | false | null | t3_18oldy7 | /r/LocalLLaMA/comments/18oldy7/mistral_instruct_finetune_dataset_format/ | false | false | self | 1 | null |
Closest thing to an Assistant? Sending emails, setting reminders...Like the upcoming Humane Ai Pin | 20 | Hi!
I've dabbling with LLM's and I'm really curious if there is a way to make it act as an assistant. At the moment from my testing it remembers "appointments" and with websearch (using Silly Tavern websearch extension) it knows what time is it. It also know in what timezone I'm on from my Persona Description.
It's ... | 2023-12-22T18:10:38 | https://www.reddit.com/r/LocalLLaMA/comments/18ol4yg/closest_thing_to_an_assistant_sending_emails/ | bullerwins | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ol4yg | false | null | t3_18ol4yg | /r/LocalLLaMA/comments/18ol4yg/closest_thing_to_an_assistant_sending_emails/ | false | false | self | 20 | null |
YSK that AGI is a negotiating chip for OpenAI | 1 | Their contract states that MSFT is entitled to everything they create EXCEPT for any AGI. That's why "AGI has been achieved internally" was a big deal, and also why the bickering has gone on for so long & so hotly. Follow the money.
https://cryptoslate.com/agi-is-excluded-from-ip-licenses-with-microsoft-should-it-be-a... | 2023-12-22T17:41:19 | https://www.reddit.com/r/LocalLLaMA/comments/18okgsp/ysk_that_agi_is_a_negotiating_chip_for_openai/ | platistocrates | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18okgsp | false | null | t3_18okgsp | /r/LocalLLaMA/comments/18okgsp/ysk_that_agi_is_a_negotiating_chip_for_openai/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'rDHczkN8U1ZAbr94k5IbDJl0kk5WFlbHABiTdZz0gFM', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/bippubJtYiJp4cQhT56X26I2f_cJ5lN6YtQeClY4IIo.jpg?width=108&crop=smart&auto=webp&s=2ae0539b9b990b9ed61e2ae968cea64b9f494e5c', 'width': 108}, {'height': 113, 'url': 'h... |
Apple's MLX Framework adds quantization support | 34 | When Apple first released the MLX framework for Apple Silicon macs it didn't support quantized models and so RAM requirements were more than a lot of people were used to. They've recently shipped quantization support. The [mlx examples](https://github.com/ml-explore/mlx-examples) have been updated to include quantizati... | 2023-12-22T17:38:12 | https://www.reddit.com/r/LocalLLaMA/comments/18oke4y/apples_mlx_framework_adds_quantization_support/ | FlishFlashman | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oke4y | false | null | t3_18oke4y | /r/LocalLLaMA/comments/18oke4y/apples_mlx_framework_adds_quantization_support/ | false | false | self | 34 | {'enabled': False, 'images': [{'id': 'NRitNAa83fFqArIPC9NwWMzw33e-3y8kSaQuduYdmHw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/k3XhFS0uf3T6hJH2r1bU3krBYMIYmYxXYP3cImVkTjo.jpg?width=108&crop=smart&auto=webp&s=4a1268ba73f625630c853841eb701b60e9da430d', 'width': 108}, {'height': 108, 'url': 'h... |
YSK that AGI is a negotiating chip for OpenA | 1 | YSK that AGI is a negotiating chip for OpenAI.
Their contract states that MSFT is entitled to everything they create EXCEPT for any AGI. That's why "AGI has been achieved internally" was a big deal, and also why the bickering has gone on for so long & so hotly. Follow the money. | 2023-12-22T17:32:58 | https://www.reddit.com/r/LocalLLaMA/comments/18ok9ur/ysk_that_agi_is_a_negotiating_chip_for_opena/ | platistocrates | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ok9ur | false | null | t3_18ok9ur | /r/LocalLLaMA/comments/18ok9ur/ysk_that_agi_is_a_negotiating_chip_for_opena/ | false | false | default | 1 | null |
Opinions on Mixtral 0.1 8x7b and Mistral 0.2 7b | 57 | The following is just some personal thoughts on these models (Mixtral 0.1 8x7b and Mistral 0.2 7b) from a random user (me), and should not be taken as facts. Opinions among users on how they perform can vary depending on hardware, settings, versions, use-cases, preferences, etc.
I thought it could be fun to share some... | 2023-12-22T17:27:49 | https://www.reddit.com/r/LocalLLaMA/comments/18ok5or/opinions_on_mixtral_01_8x7b_and_mistral_02_7b/ | Admirable-Star7088 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ok5or | false | null | t3_18ok5or | /r/LocalLLaMA/comments/18ok5or/opinions_on_mixtral_01_8x7b_and_mistral_02_7b/ | false | false | self | 57 | null |
Does having more resources leads to better outcomes in terms of quality of output, or am i dreaming this? | 3 | I've installed lmstudio on one machine with 16gb ram, and another with 64gb ram. Both have the same GPU and 8gb vram.
The speed difference is mostly on first prompt, with a slight improvement on larger model (mostly consisting of being able to load them in the first place hehe) - **BUT** on models such as mythomax l2... | 2023-12-22T17:19:43 | https://www.reddit.com/r/LocalLLaMA/comments/18ojz9n/does_having_more_resources_leads_to_better/ | RadioSailor | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ojz9n | false | null | t3_18ojz9n | /r/LocalLLaMA/comments/18ojz9n/does_having_more_resources_leads_to_better/ | false | false | self | 3 | null |
Why aren’t LoRA’s a big thing i the LLM realm? | 122 | Sorry if that’s a stupid question, but it seems wierdly underutilized to have LoRA’s for different tasks. E.g. a LoRA on top for storywriting, one for sales emails etc. Is that just worse than finetuning a model or what is your explanation?
In the StableDiffusion realm it is certainly much more utilized and with great ... | 2023-12-22T16:47:36 | https://www.reddit.com/r/LocalLLaMA/comments/18oj983/why_arent_loras_a_big_thing_i_the_llm_realm/ | ___defn | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oj983 | false | null | t3_18oj983 | /r/LocalLLaMA/comments/18oj983/why_arent_loras_a_big_thing_i_the_llm_realm/ | false | false | self | 122 | null |
Should I adjust my learning rate? | 1 | Hi, I am using axolotl to train Llama-2-7b-chat-hf with qlora. My training dataset is \~1.3M rows of data and I'm using the following training parameters:
* micro\_batch\_size: 2
* num\_epochs: 4
* learning\_rate: 0.0002
* optimizer: paged\_adamw\_32bit
This is my training loss chart after 1.3 epochs:
https://pre... | 2023-12-22T16:46:55 | https://www.reddit.com/r/LocalLLaMA/comments/18oj8ol/should_i_adjust_my_learning_rate/ | RizzMiser | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oj8ol | false | null | t3_18oj8ol | /r/LocalLLaMA/comments/18oj8ol/should_i_adjust_my_learning_rate/ | false | false | 1 | null | |
Ollama iOS mobile app (open source) | 26 | 2023-12-22T16:32:11 | https://v.redd.it/mebj3z6eiv7c1 | 1amrocket | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18oiwno | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/mebj3z6eiv7c1/DASHPlaylist.mpd?a=1705854745%2CODFmNjQ2NTNmNjQ1ZmU0ZjQwMTAwZGExNmMxNTBlNThjZDUxZGE1MzIyZTU5NDQ0ZjFlM2VlN2U2ZTg1MGIwZQ%3D%3D&v=1&f=sd', 'duration': 89, 'fallback_url': 'https://v.redd.it/mebj3z6eiv7c1/DASH_1080.mp4?source=fallback', 'h... | t3_18oiwno | /r/LocalLLaMA/comments/18oiwno/ollama_ios_mobile_app_open_source/ | false | false | 26 | {'enabled': False, 'images': [{'id': 'YzJ6aTRvM2xpdjdjMfSuwtMFsz81_YvBeCujB9YP6NmQY-Q3EZ8LTnVSHDmg', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/YzJ6aTRvM2xpdjdjMfSuwtMFsz81_YvBeCujB9YP6NmQY-Q3EZ8LTnVSHDmg.png?width=108&crop=smart&format=pjpg&auto=webp&s=a259d7bf54972c6c995dfd6dedfa07d569182... | ||
Gemini-pro is so lazy till the point it's so useless for RAG applicatiom | 34 | Currently i have some free credits from GCP and curious to try out gemini model. But the API result from the model is just too bad compare to other large model and even worst than mistral 7B. Don't get me wrong that the gemini still answer correctly but the way it answers is so blandness and unhelpful despite I already... | 2023-12-22T16:28:48 | https://www.reddit.com/r/LocalLLaMA/comments/18oituc/geminipro_is_so_lazy_till_the_point_its_so/ | cheviethai123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oituc | false | null | t3_18oituc | /r/LocalLLaMA/comments/18oituc/geminipro_is_so_lazy_till_the_point_its_so/ | false | false | self | 34 | null |
Can you feed an entire source code to LLAMA? | 3 | i want to give LLAMA an entire source code to fine tune it with it, is it possible? and will the source code stay in my machine or will be shared with other companies? | 2023-12-22T16:16:46 | https://www.reddit.com/r/LocalLLaMA/comments/18oik3l/can_you_feed_an_entire_source_code_to_llama/ | Dependent_Tie8184 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oik3l | false | null | t3_18oik3l | /r/LocalLLaMA/comments/18oik3l/can_you_feed_an_entire_source_code_to_llama/ | false | false | self | 3 | null |
fine-tuning from logits — dumb or genius? | 11 | As OpenAI now alows logit-derived probabilities to be output( https://github.com/openai/openai-cookbook/blob/main/examples/Using\_logprobs.ipynb )
and such a probability distribution can be grabbed from the open models as well,
would it make sense to build fine-tuning datasets with those as that could speed up the ... | 2023-12-22T16:13:55 | https://www.reddit.com/r/LocalLLaMA/comments/18oihrq/finetuning_from_logits_dumb_or_genius/ | uhuge | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oihrq | false | null | t3_18oihrq | /r/LocalLLaMA/comments/18oihrq/finetuning_from_logits_dumb_or_genius/ | false | false | self | 11 | null |
Who wants to explore some edge deployment LLM e-commerce opportunities? | 1 | I was thinking recently and believe there's great market potential for a variety of edge deployment LLM products, especially with the rise of smaller open source models these past few months.
I currently run a few large ecom stores that I've built, with my specialty being SEO. Been in GPT3 beta since June 2021 and hav... | 2023-12-22T16:11:16 | https://www.reddit.com/r/LocalLLaMA/comments/18oifo9/who_wants_to_explore_some_edge_deployment_llm/ | fission4433 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oifo9 | false | null | t3_18oifo9 | /r/LocalLLaMA/comments/18oifo9/who_wants_to_explore_some_edge_deployment_llm/ | false | false | self | 1 | null |
Question about System RAM and GPU VRAM requirements for large models | 1 | Hello folks, I need some help to see if I can add GPUs to repurpose my older computer for LLM (interference mainly, maybe training later on).
My current limitation is that I have only 2 ddr4 ram slots, and can either continue with 16GBx2 or look for a set of 32GBx2 kit.
My understanding is that we can reduce system ... | 2023-12-22T15:55:57 | https://www.reddit.com/r/LocalLLaMA/comments/18oi2vd/question_about_system_ram_and_gpu_vram/ | Silver-of-Curiosity | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oi2vd | false | null | t3_18oi2vd | /r/LocalLLaMA/comments/18oi2vd/question_about_system_ram_and_gpu_vram/ | false | false | self | 1 | null |
How-to for training a model on tweets | 1 | [removed] | 2023-12-22T15:18:55 | https://www.reddit.com/r/LocalLLaMA/comments/18oh9bz/howto_for_training_a_model_on_tweets/ | SideShow_Bot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oh9bz | false | null | t3_18oh9bz | /r/LocalLLaMA/comments/18oh9bz/howto_for_training_a_model_on_tweets/ | false | false | self | 1 | null |
Noob question on MOE finetuning possibilities | 5 | I don't really understand the FT processes or MOE architecture but I was wondering if it would it be possible to only train part of the network at a time?
Would this let you reduce VRAM requirements for FT and allow you not 'spoil' parts of the model you're happy with? | 2023-12-22T15:16:27 | https://www.reddit.com/r/LocalLLaMA/comments/18oh7gg/noob_question_on_moe_finetuning_possibilities/ | HikaruZA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oh7gg | false | null | t3_18oh7gg | /r/LocalLLaMA/comments/18oh7gg/noob_question_on_moe_finetuning_possibilities/ | false | false | self | 5 | null |
JSON mode in vLLM | 42 | I am the author of the[Outlines library](https://github.com/outlines-dev/outlines) that provides guided generation for Large Language Models. Our users frequently asked us how they could deploy JSON-guided generation to solve their use case. So I patched the vLLM library and modified their API serving file to add the p... | 2023-12-22T14:38:20 | https://www.reddit.com/r/LocalLLaMA/comments/18ogd75/json_mode_in_vllm/ | GoBayesGo | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ogd75 | false | null | t3_18ogd75 | /r/LocalLLaMA/comments/18ogd75/json_mode_in_vllm/ | false | false | self | 42 | {'enabled': False, 'images': [{'id': '20L5uo42HDnGHpNOXV8hP7kSao53LFtUV02eMCO3kgs', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/AUaEDegu6fPOBeD4BuoULnDNUvMQF39ipI02ZibyCmU.jpg?width=108&crop=smart&auto=webp&s=debf90f85c36f0f56c817f62e6a23773ea1263da', 'width': 108}, {'height': 108, 'url': 'h... |
Can AMD provide better support to LLM Community loccal SOFTWARE to use their NPU from ryzen 7 ryzen 9 7940hs 7840 hs and new 8940hs ? | 1 | [removed] | 2023-12-22T14:14:41 | https://www.reddit.com/r/LocalLLaMA/comments/18ofvjx/can_amd_provide_better_support_to_llm_community/ | maxpayne07 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ofvjx | false | null | t3_18ofvjx | /r/LocalLLaMA/comments/18ofvjx/can_amd_provide_better_support_to_llm_community/ | false | false | self | 1 | null |
Apple LLM breakthrough: LLM in a flash - Efficient Large Language Model Inference with Limited Memory | 2 | 2023-12-22T14:14:09 | https://arxiv.org/pdf/2312.11514.pdf | Xnohat | arxiv.org | 1970-01-01T00:00:00 | 0 | {} | 18ofv4x | false | null | t3_18ofv4x | /r/LocalLLaMA/comments/18ofv4x/apple_llm_breakthrough_llm_in_a_flash_efficient/ | false | false | default | 2 | null | |
Can I run 70b models with 24gb VRAM and a 3090RTX? | 6 | title | 2023-12-22T13:55:38 | https://www.reddit.com/r/LocalLLaMA/comments/18ofh58/can_i_run_70b_models_with_24gb_vram_and_a_3090rtx/ | laveshnk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ofh58 | false | null | t3_18ofh58 | /r/LocalLLaMA/comments/18ofh58/can_i_run_70b_models_with_24gb_vram_and_a_3090rtx/ | false | false | self | 6 | null |
Andrej Karpathy championing this sub while working at 'Open’ AI doesn't sit well with me....... | 1 | 2023-12-22T13:45:01 | https://v.redd.it/ze7knn2rou7c1 | Iboxelephants | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18of9ab | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/ze7knn2rou7c1/DASHPlaylist.mpd?a=1705844715%2COWIzYmE5N2MwMzRkMThkYmI4YjBiMmExZTcwZDU0NWZhNGUwNmIzYTVlMzI2YmFiNmRmMmEwZjIwY2VlYzkxOQ%3D%3D&v=1&f=sd', 'duration': 4, 'fallback_url': 'https://v.redd.it/ze7knn2rou7c1/DASH_1080.mp4?source=fallback', 'ha... | t3_18of9ab | /r/LocalLLaMA/comments/18of9ab/andrej_karpathy_championing_this_sub_while/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'cTBoc3Q3dHFvdTdjMaZ4WeSzrT0pgNhSGByXf2bB6xaHjZITawBwxMZH5MWj', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/cTBoc3Q3dHFvdTdjMaZ4WeSzrT0pgNhSGByXf2bB6xaHjZITawBwxMZH5MWj.png?width=108&crop=smart&format=pjpg&auto=webp&s=97364b37b5dc9a6d06289b16b95b5e100f656... | ||
Prompt engineering strategies to ensure output consistency? | 1 | Following is the output of Microsoft's Phi-2 (phi-2 8 bit quant gguf) running on .95 temp.
`Introduce yourself.`
`Hi there, my name is Alex and I am a student studying psychology at University of Cambridge. My main interest is in understanding how the human mind works and how we can use that knowledge to improve peo... | 2023-12-22T13:44:27 | https://www.reddit.com/r/LocalLLaMA/comments/18of8vz/prompt_engineering_strategies_to_ensure_output/ | saccharineboi | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18of8vz | false | null | t3_18of8vz | /r/LocalLLaMA/comments/18of8vz/prompt_engineering_strategies_to_ensure_output/ | false | false | self | 1 | null |
Andrej Karpathy dickriding this sub while working at ‘Open’ AI doesn't sit well with me. | 1 | 2023-12-22T13:42:47 | https://v.redd.it/1zr6c4qcou7c1 | Iboxelephants | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18of7ok | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/1zr6c4qcou7c1/DASHPlaylist.mpd?a=1705844583%2CNTc0ZTRiMmI2NzRhNzRmODcyN2FkMDg0NzdhMWJmMDQyN2U4NTYzOTNiMmFhNTQ4OTY4NzA0ZWI5ZjQxOWRiNA%3D%3D&v=1&f=sd', 'duration': 4, 'fallback_url': 'https://v.redd.it/1zr6c4qcou7c1/DASH_1080.mp4?source=fallback', 'ha... | t3_18of7ok | /r/LocalLLaMA/comments/18of7ok/andrej_karpathy_dickriding_this_sub_while_working/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'd2hzeXN0Z2NvdTdjMXoz1L1IrqfSHVj_TqXJlX7WZduhHNOBmwVM6Z0Mh630', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/d2hzeXN0Z2NvdTdjMXoz1L1IrqfSHVj_TqXJlX7WZduhHNOBmwVM6Z0Mh630.png?width=108&crop=smart&format=pjpg&auto=webp&s=45a63f988aba553974a694317bfa3c790a0d7... | ||
What linux distribution is better to run LLMs on old laptop? | 2 | I got a very old laptop in perfect condition but not in use. It got 8gb Ram and touch screen.
It's currently running windows 11.
What is the best **lightweight linux distribution** that can run koboldccp or text generation ?
**And what token rate can i expect for a q4_k_m 7b model ?**
If possible i want to use it ... | 2023-12-22T13:29:15 | https://www.reddit.com/r/LocalLLaMA/comments/18oexp7/what_linux_distribution_is_better_to_run_llms_on/ | Prior-Blood5979 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oexp7 | false | null | t3_18oexp7 | /r/LocalLLaMA/comments/18oexp7/what_linux_distribution_is_better_to_run_llms_on/ | false | false | self | 2 | null |
Apple Researchers Unveil DeepPCR: A Novel Machine Learning Algorithm that Parallelizes Typically Sequential Operations in Order to Speed Up Inference and Training of Neural Networks | 79 | 7x Faster training and 11x Faster generation
"The paper demonstrates the effectiveness of DeepPCR through various applications. It achieves speedups of up to 30× for forward and 200× for backward passes in multi-layer perceptrons. Additionally, the algorithm is applied to parallelize training of deep ResNet archite... | 2023-12-22T13:07:03 | https://www.reddit.com/r/LocalLLaMA/comments/18oeimf/apple_researchers_unveil_deeppcr_a_novel_machine/ | paryska99 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oeimf | false | null | t3_18oeimf | /r/LocalLLaMA/comments/18oeimf/apple_researchers_unveil_deeppcr_a_novel_machine/ | false | false | self | 79 | null |
Looking for a model for iterative summarisation using webui API and question about endpoints | 2 | Can you recommend any 7b GGUF model that performs exceptionally well when it comes to summarisation? I've been using OpenAI offerings for my workflow/script, and it works well, but is rather expensive. So far I've been experimanting with Dolphin. Admittedly, I'm a noob, so I'm not sure which endpoint to use for best re... | 2023-12-22T12:54:59 | https://www.reddit.com/r/LocalLLaMA/comments/18oeabn/looking_for_a_model_for_iterative_summarisation/ | black_cat90 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oeabn | false | null | t3_18oeabn | /r/LocalLLaMA/comments/18oeabn/looking_for_a_model_for_iterative_summarisation/ | false | false | self | 2 | null |
I REFUSE TO ACCEPT the idea that transformers based LLM we have won't be able to equal the best human mathematicians and reasoners. | 1 | Why?
Because scaling laws are still true.
A larger model trained the same way with similar digital data quality will always outperform a smaller model especially in reasoning and maths.
There can be no discussion in this matter. It is true and if we look at the trends it will continue to be true the more we scale.
... | 2023-12-22T12:53:19 | https://www.reddit.com/r/LocalLLaMA/comments/18oe9bv/i_refuse_to_accept_the_idea_that_transformers/ | Iboxelephants | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oe9bv | false | null | t3_18oe9bv | /r/LocalLLaMA/comments/18oe9bv/i_refuse_to_accept_the_idea_that_transformers/ | false | false | self | 1 | null |
Need help for use case with local models | 1 | [removed] | 2023-12-22T12:11:44 | https://www.reddit.com/r/LocalLLaMA/comments/18odjci/need_help_for_use_case_with_local_models/ | faridukhan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18odjci | false | null | t3_18odjci | /r/LocalLLaMA/comments/18odjci/need_help_for_use_case_with_local_models/ | false | false | self | 1 | null |
I built a unlimited free to use GPT4, Voice AI (100+ Voices) and RoomGpt( 80+ themes) in Switch. It's free and unlimited generations. Hope you guys can make use of them:). Here is the link https://play.google.com/store/apps/details?id=com.app.switchapp1 . Find clickable link in body section. | 1 | https://play.google.com/store/apps/details?id=com.app.switchapp1 | 2023-12-22T11:57:02 | https://www.reddit.com/gallery/18odaam | LengthWorking4994 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18odaam | false | null | t3_18odaam | /r/LocalLLaMA/comments/18odaam/i_built_a_unlimited_free_to_use_gpt4_voice_ai_100/ | false | false | 1 | null | |
Unique fine tuned Models, what are your favorites? | 1 | Recently I found that there was a model fine tuned on mysticism/occultism and I only got to play with it a bit last night but honestly found it amazing.
I would love to learn more models like this. Are there any out there fine tuned for D&D? Star Wars? Specific novelists?
My other question which might be better as i... | 2023-12-22T11:52:57 | https://www.reddit.com/r/LocalLLaMA/comments/18od81g/unique_fine_tuned_models_what_are_your_favorites/ | Foot-Note | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18od81g | false | null | t3_18od81g | /r/LocalLLaMA/comments/18od81g/unique_fine_tuned_models_what_are_your_favorites/ | false | false | self | 1 | null |
Couuld llama2 70b be run on a tesla m10? | 3 | I am trying to run llama2 70b and use it as a base model for further optimization.
Does anyone have any experience using the tesla m10? Would it be possible, and if it is, anything I should know that would might be a challenge?
https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-m10/pdf/188359-Tesla-... | 2023-12-22T11:28:05 | https://www.reddit.com/r/LocalLLaMA/comments/18ocu6q/couuld_llama2_70b_be_run_on_a_tesla_m10/ | bluespy89 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ocu6q | false | null | t3_18ocu6q | /r/LocalLLaMA/comments/18ocu6q/couuld_llama2_70b_be_run_on_a_tesla_m10/ | false | false | self | 3 | null |
Llama 2 is awesome! | 8 | It is really good to be able to mess around with such an open source project! | 2023-12-22T11:18:09 | https://www.reddit.com/r/LocalLLaMA/comments/18ocoph/llama_2_is_awesome/ | azmoura | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ocoph | false | null | t3_18ocoph | /r/LocalLLaMA/comments/18ocoph/llama_2_is_awesome/ | false | false | self | 8 | null |
desperate noob needs help | 1 | [removed] | 2023-12-22T11:15:34 | https://www.reddit.com/r/LocalLLaMA/comments/18ocn9z/desperate_noob_needs_help/ | TravelOdd3712 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ocn9z | false | null | t3_18ocn9z | /r/LocalLLaMA/comments/18ocn9z/desperate_noob_needs_help/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'RhTfmtdYpD3FM9yL1KB0OZqOyIM1ddRt2J1Ctjsd_Fc', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/T1SZWoSg1O-0Nspmx3NgT-ikFs40ixi5HcnkXVZgf_g.jpg?width=108&crop=smart&auto=webp&s=e5a287b1663987da09838c6ca88936f0aac32b05', 'width': 108}, {'height': 162, 'url': 'h... |
Is there a model that works well with Russian language? | 1 | Is there a model that works well with Russian language? | 2023-12-22T11:15:26 | https://www.reddit.com/r/LocalLLaMA/comments/18ocn7l/is_there_a_model_that_works_well_with_russian/ | glorsh66 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ocn7l | false | null | t3_18ocn7l | /r/LocalLLaMA/comments/18ocn7l/is_there_a_model_that_works_well_with_russian/ | false | false | self | 1 | null |
Is there a good model to extract names from text? | 2 | Is there a good model to extract names from text? | 2023-12-22T11:14:44 | https://www.reddit.com/r/LocalLLaMA/comments/18ocmrn/is_there_a_good_model_to_extract_names_from_text/ | glorsh66 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18ocmrn | false | null | t3_18ocmrn | /r/LocalLLaMA/comments/18ocmrn/is_there_a_good_model_to_extract_names_from_text/ | false | false | self | 2 | null |
Macbook or PC for running LLMs? | 3 | So, I have a MacBook and a PC and I was wondering which would be more suited for running LLMs.
My MacBook is an M1 Pro and has 32gb of unified memory (shared cpu and gpu memory. Its 128-bit LPDDR4X SDRAM)
My PC is a 13600kf with 32 gb of 3200MT/s DDR4 and a 6700xt with 12gb of vram
So, my three options for ... | 2023-12-22T11:13:00 | https://www.reddit.com/r/LocalLLaMA/comments/18oclws/macbook_or_pc_for_running_llms/ | McpeIsSoBuggy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oclws | false | null | t3_18oclws | /r/LocalLLaMA/comments/18oclws/macbook_or_pc_for_running_llms/ | false | false | self | 3 | null |
I built a unlimited free to use GPT4, Voice AI (100+ Voices) and RoomGpt( 80+ themes) in Switch. It's free and unlimited generations. Hope you guys can make use of them:). Find the link in Comments section. | 1 | 2023-12-22T10:55:01 | https://www.reddit.com/gallery/18ocbnm | LengthWorking4994 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 18ocbnm | false | null | t3_18ocbnm | /r/LocalLLaMA/comments/18ocbnm/i_built_a_unlimited_free_to_use_gpt4_voice_ai_100/ | false | false | 1 | null | ||
Been using Goliath 120B, game changing for roleplay, but are there cheaper alternatives? | 40 | Goliath 120B is amazing, but are there smaller models that has the same level of detail and creativity? | 2023-12-22T10:47:26 | https://www.reddit.com/r/LocalLLaMA/comments/18oc7rg/been_using_goliath_120b_game_changing_for/ | Extra-Fig-7425 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oc7rg | false | null | t3_18oc7rg | /r/LocalLLaMA/comments/18oc7rg/been_using_goliath_120b_game_changing_for/ | false | false | self | 40 | null |
I tried to teach Mistral 7B a new language (Sundanese) and it worked! (sort of) | 144 | [Nero10578/Mistral-7B-Sunda-v1.0 · Hugging Face](https://huggingface.co/Nero10578/Mistral-7B-Sunda-v1.0)
Seeing that even Bing chat which is powered by chatGPT-4 couldn't speak in Sundanese when asked, I thought of trying to teach Mistral-7B Sundanese using just QLora training. It surprisingly worked out pretty well... | 2023-12-22T10:36:24 | https://www.reddit.com/r/LocalLLaMA/comments/18oc1yc/i_tried_to_teach_mistral_7b_a_new_language/ | nero10578 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oc1yc | false | null | t3_18oc1yc | /r/LocalLLaMA/comments/18oc1yc/i_tried_to_teach_mistral_7b_a_new_language/ | false | false | self | 144 | {'enabled': False, 'images': [{'id': 'Kbc4t8i3_BJ1yNpqg-shARbU8csjm5G1LtpVQYX83oQ', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/FvkYiF4N1m6tNlmNT4G-9GCgIBclpimu1SFx7b_g-Ic.jpg?width=108&crop=smart&auto=webp&s=24d2d9104f9d0e2fc4966eeb323ac2a1fb6cca96', 'width': 108}, {'height': 116, 'url': 'h... |
3B models are great when used with ChatGPT character | 16 | I recently tested a lot of 3B models and it is surprisingly usable when giving it ChatGPT character. Below is an example when I put this prompt in the memory tab; 'Gor is ChatGPT. Gor give accurate answer in direct and concise manner. Gor follow all instructions.'
This tips might be useful for those who use lower mode... | 2023-12-22T09:24:12 | https://www.reddit.com/r/LocalLLaMA/comments/18oazcz/3b_models_are_great_when_used_with_chatgpt/ | DivineAscension | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oazcz | false | null | t3_18oazcz | /r/LocalLLaMA/comments/18oazcz/3b_models_are_great_when_used_with_chatgpt/ | false | false | 16 | null | |
Cannot reproduce the exact same training results - diffusion models | 1 | Hi ! In order to collaborate with my team we want to be able to get the same training results based on a finetuned version of [https://huggingface.co/runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5).
We have real issues defining the exact same environment, even if we have the ... | 2023-12-22T09:14:09 | https://www.reddit.com/r/LocalLLaMA/comments/18oau6r/cannot_reproduce_the_exact_same_training_results/ | New_Detective_1363 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oau6r | false | null | t3_18oau6r | /r/LocalLLaMA/comments/18oau6r/cannot_reproduce_the_exact_same_training_results/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'QCpE9U9tHthu6Tk68NcMSdO4JuQy7KXbcN_EMSAAqpA', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/-nJ7f0Q0CjLyTmaxM0npqUNftpWUfhbKtuco-fYGp7M.jpg?width=108&crop=smart&auto=webp&s=dff119d09f945c4fb3ecf70685bedf3820cb50c4', 'width': 108}, {'height': 116, 'url': 'h... |
Multiple Users compute requirements | 4 | Hi,
I'm going to be hosting either OpenChat-3.5-16K or Mixtral-32k on Runpod. I'll have anywhere from 10-20 concurrent users. Can anyone give me some advice about what kind of compute I'll need to purchase? I plan to run my model on Ooba. | 2023-12-22T08:56:26 | https://www.reddit.com/r/LocalLLaMA/comments/18oakpt/multiple_users_compute_requirements/ | Inevitable_Job_492 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oakpt | false | null | t3_18oakpt | /r/LocalLLaMA/comments/18oakpt/multiple_users_compute_requirements/ | false | false | self | 4 | null |
Fine-tuned Mistral on my Telegram messages | 1 | [removed] | 2023-12-22T08:51:13 | https://www.reddit.com/r/LocalLLaMA/comments/18oai62/finetuned_mistral_on_my_telegram_messages/ | furiousteabag | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oai62 | false | null | t3_18oai62 | /r/LocalLLaMA/comments/18oai62/finetuned_mistral_on_my_telegram_messages/ | false | false | self | 1 | null |
How to add stopping strings to llama.cpp? | 3 | I've searched everywhere on how to add stopping strings to the llama.cpp webui server, but I cannot find a way to make it work. I really want to use the webui, and not the console.
Do anyone know how to add stopping strings to the webui server? There are settings inside the webui, but not for stopping strings.
Im r... | 2023-12-22T08:16:39 | https://www.reddit.com/r/LocalLLaMA/comments/18oa12k/how_to_add_stopping_strings_to_llamacpp/ | fetballe | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18oa12k | false | null | t3_18oa12k | /r/LocalLLaMA/comments/18oa12k/how_to_add_stopping_strings_to_llamacpp/ | false | false | self | 3 | null |
What models can I comfortably run with an 16 gb ram and rtx 3060? | 6 | i am new in this field, don't have any idea. Can i run dolphin mistral8x7b ?? | 2023-12-22T08:00:27 | https://www.reddit.com/r/LocalLLaMA/comments/18o9spl/what_models_can_i_comfortably_run_with_an_16_gb/ | Good-Delay3011 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o9spl | false | null | t3_18o9spl | /r/LocalLLaMA/comments/18o9spl/what_models_can_i_comfortably_run_with_an_16_gb/ | false | false | self | 6 | null |
How to use Dolphin-Mixtral uncensored model on you Mac Silicon | 1 | 2023-12-22T07:58:03 | Sponge-Theory | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 18o9rfj | false | null | t3_18o9rfj | /r/LocalLLaMA/comments/18o9rfj/how_to_use_dolphinmixtral_uncensored_model_on_you/ | false | false | nsfw | 1 | {'enabled': True, 'images': [{'id': 'Ybh3HcNCIYbOV9V9XmoAEIhpAkwafeOjcdgk8Jh1aGw', 'resolutions': [{'height': 72, 'url': 'https://preview.redd.it/dorpe4kuys7c1.png?width=108&crop=smart&auto=webp&s=0c4985bcdc261f59054fd76c7112a7da7cd2fa4d', 'width': 108}, {'height': 145, 'url': 'https://preview.redd.it/dorpe4kuys7c1.png... | ||
Why do graphic cards click when doing compute? | 11 | Whenever I run LLM inference on my Intel Arc/AMD/or Nvidia cards including Teslas. There is this click click click sound as it does the compute.
I wonder where that came from? What purpose does it serve?
Oh and smaller GPUs like Mali GPUs on devboards don't seem to click, unlike graphics cards on bigger PCs. | 2023-12-22T07:45:28 | https://www.reddit.com/r/LocalLLaMA/comments/18o9l7q/why_do_graphic_cards_click_when_doing_compute/ | A_Degenerate_Idiot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o9l7q | false | null | t3_18o9l7q | /r/LocalLLaMA/comments/18o9l7q/why_do_graphic_cards_click_when_doing_compute/ | false | false | self | 11 | null |
Phind-codellama trained on stack-overflow amongst other things. | 10 | I am running it locally via ollama on an M2 MBP. I asked it a coding question and after its response I said “thank you”. To which it replied “You’re welcome! If you find this helpful, please consider accepting the answer.” Hmmm :-). | 2023-12-22T07:41:59 | https://www.reddit.com/r/LocalLLaMA/comments/18o9ji5/phindcodellama_trained_on_stackoverflow_amongst/ | nborwankar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o9ji5 | false | null | t3_18o9ji5 | /r/LocalLLaMA/comments/18o9ji5/phindcodellama_trained_on_stackoverflow_amongst/ | false | false | self | 10 | null |
How to reduce Hallunications of the outputs generated by the LLMs. | 2 | After a LLM generates a answer is there a way to trace back from which document the generated text has been retrieved? I have stored the document information as meta data while upserting the vector embeddings into a vector database. I want attribution to where in the document corpus was the generated text from.
Also a... | 2023-12-22T07:29:39 | https://www.reddit.com/r/LocalLLaMA/comments/18o9cy8/how_to_reduce_hallunications_of_the_outputs/ | vm123313223 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o9cy8 | false | null | t3_18o9cy8 | /r/LocalLLaMA/comments/18o9cy8/how_to_reduce_hallunications_of_the_outputs/ | false | false | self | 2 | null |
Can I run LLM on Android natively? | 1 | [removed] | 2023-12-22T07:13:21 | https://www.reddit.com/r/LocalLLaMA/comments/18o94qa/can_i_run_llm_on_android_natively/ | bharathdp | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o94qa | false | null | t3_18o94qa | /r/LocalLLaMA/comments/18o94qa/can_i_run_llm_on_android_natively/ | false | false | self | 1 | null |
Running Mixtral-8x7B-Instruct-v0.1 On Server? | 2 | Hello, from using h20's [https://gpt-gm.h2o.ai](https://gpt-gm.h2o.ai) website, I have found that the Mixtral-8x7B-Instruct-v0.1 model works PERFECTLY for my use case. Is there any way I can deploy this model to a server with the same document-based capabilities as this demo website for relatively low cost (<$50 a mont... | 2023-12-22T07:08:09 | https://www.reddit.com/r/LocalLLaMA/comments/18o9203/running_mixtral8x7binstructv01_on_server/ | Critical_Pop_2216 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o9203 | false | null | t3_18o9203 | /r/LocalLLaMA/comments/18o9203/running_mixtral8x7binstructv01_on_server/ | false | false | self | 2 | null |
pip install startup advisor | 2 | Using supervised fine-tuning of the **stable lm-zephyr-3b** with 1000+ hours of tech & startup podcast transcripts, now you can `pip install localmentor` to chat startup on your local computer.
Repo: [https://github.com/remyxai/LocalMentor](https://github.com/remyxai/LocalMentor) | 2023-12-22T06:53:01 | https://www.reddit.com/r/LocalLLaMA/comments/18o8u21/pip_install_startup_advisor/ | remyxai | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o8u21 | false | null | t3_18o8u21 | /r/LocalLLaMA/comments/18o8u21/pip_install_startup_advisor/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'SqEKD0fqoDt1-uXyTjjxegoRheuvk6hjRw30XHfw5XE', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/7niuJDxqY2DTbhcaZfoGkzEa3NOX43mSThcBDL-gA9Y.jpg?width=108&crop=smart&auto=webp&s=8a4e39a6f4e516dff2dde750576b3ef8ce28b1c8', 'width': 108}, {'height': 108, 'url': 'h... |
Has anyone tried running this with a local model, maybe Mixtral? “We demonstrate generative agents by populating a sandbox environment, reminiscent of The Sims, with twenty-five agents.” It was open-sourced a while ago, but the API costs to run it would be prohibitive. | 20 | 2023-12-22T06:27:01 | https://github.com/joonspk-research/generative_agents | danysdragons | github.com | 1970-01-01T00:00:00 | 0 | {} | 18o8fhm | false | null | t3_18o8fhm | /r/LocalLLaMA/comments/18o8fhm/has_anyone_tried_running_this_with_a_local_model/ | false | false | 20 | {'enabled': False, 'images': [{'id': 'oZY7hufqIQ35fb0KnqeA-TW1NJe0knjyVE90X7JsYS4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/7UqZIbEhJsc6eGuEa2mIxPokA49F49Oi3aPR-O_9gHo.jpg?width=108&crop=smart&auto=webp&s=29359ed50bc66f0735c49d9175390789b0674b5f', 'width': 108}, {'height': 108, 'url': 'h... | ||
Need Guidance on Running LLAMA 7B/13B Models on Local K8s Cluster | 2 | Hello everyone,
I’m setting up a local k8s cluster using 5 workstations, each equipped with an AMD Ryzen™ 9 7950X CPU and 128 GB of RAM. Three of these systems have GeForce RTX 3060 GDDR6 12 GB GPUs, and two have GeForce RTX 4060 GDDR6 8 GB GPUs. All run Ubuntu desktop 22.04
I'm aiming to run LLAMA 7B and 13B model ... | 2023-12-22T06:22:49 | https://www.reddit.com/r/LocalLLaMA/comments/18o8d69/need_guidance_on_running_llama_7b13b_models_on/ | riskyolive | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o8d69 | false | null | t3_18o8d69 | /r/LocalLLaMA/comments/18o8d69/need_guidance_on_running_llama_7b13b_models_on/ | false | false | self | 2 | null |
AppAgent: Multimodal Agents as Smartphone Users | 1 | https://huggingface.co/papers/2312.13771
Abstract
Recent advancements in large language models (LLMs) have led to the creation of intelligent agents capable of performing complex tasks. This paper introduces a novel LLM-based multimodal agent framework designed to operate smartphone applications. Our framework enable... | 2023-12-22T06:13:56 | https://www.reddit.com/r/LocalLLaMA/comments/18o880p/appagent_multimodal_agents_as_smartphone_users/ | metalman123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o880p | false | null | t3_18o880p | /r/LocalLLaMA/comments/18o880p/appagent_multimodal_agents_as_smartphone_users/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'KAZAEsQ-XUC_NxiYLajYSb_vsZi5cNq3hy2UA9JibTs', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/WdiN-2m8TnTXhibxLMU71JsugnxcGKwCE8u8Q551Df4.jpg?width=108&crop=smart&auto=webp&s=274f44faa26bb905b0ec13c365f7f0849e213f80', 'width': 108}, {'height': 116, 'url': 'h... |
What's the best way to beat GPT-5 with Open Source & Distributed Compute? | 47 | According to you (the community) it looks like the biggest hurdle to beating GPT-4 & GPT-5 level models is compute.
However, using a variety of strategies, including distributed compute, it looks achievable.
The question is, does anyone have clear next steps for how we would go about this?
This is what we gathered i... | 2023-12-22T06:05:47 | https://www.reddit.com/r/LocalLLaMA/comments/18o83dm/whats_the_best_way_to_beat_gpt5_with_open_source/ | askchris | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o83dm | false | null | t3_18o83dm | /r/LocalLLaMA/comments/18o83dm/whats_the_best_way_to_beat_gpt5_with_open_source/ | false | false | self | 47 | null |
Typical Fine-tuning Settings and Parameters? | 6 | Hi all, I've recently started fine-tuning some models using transformers and have a couple questions about settings and parameters.
1. When using Lora, why do we typically only add adapters on Q and V matrices?
2. Are there rough defaults to \`lora\_r\` and \`lora\_alpha\` parameters? Any other params to try tuning?
3... | 2023-12-22T05:51:58 | https://www.reddit.com/r/LocalLLaMA/comments/18o7v66/typical_finetuning_settings_and_parameters/ | xnick77x | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o7v66 | false | null | t3_18o7v66 | /r/LocalLLaMA/comments/18o7v66/typical_finetuning_settings_and_parameters/ | false | false | self | 6 | null |
Favorite fine-tuning tool? | 8 | I've seen a lot of people here suggest Axolotl ([https://github.com/OpenAccess-AI-Collective/axolotl.git](https://github.com/OpenAccess-AI-Collective/axolotl.git)) as there go-to tool for fine-tuning models, but there also are a few other options that seem to be popular. For example, LLaMA Factory ([https://github.com/... | 2023-12-22T04:54:54 | https://www.reddit.com/r/LocalLLaMA/comments/18o6wo0/favorite_finetuning_tool/ | SirStagMcprotein | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o6wo0 | false | null | t3_18o6wo0 | /r/LocalLLaMA/comments/18o6wo0/favorite_finetuning_tool/ | false | false | self | 8 | {'enabled': False, 'images': [{'id': 'vwO_BTf6pLhUmX_TRqL3aU-kiYzWkx6M_WXTmbaBVa4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/o-rCba6np3PdOkN20kolVyIAj1pfH4xabEvxH8dPgnA.jpg?width=108&crop=smart&auto=webp&s=232091dd4f5c006766403f8b8480557fb4aa509d', 'width': 108}, {'height': 108, 'url': 'h... |
Helpful VRAM requirement table for qlora, lora, and full finetuning. | 130 | Source: https://github.com/hiyouga/LLaMA-Factory#hardware-requirement
### Hardware Requirements
| Method | Bits | 7B | 13B | 30B | 65B | 8x7B |
| ------ | ---- | ----- | ----- | ----- | ------ | ------ |
| Full | 16 | 160GB | 320GB | 600GB | 1200GB | 1000GB |
| Freeze | 16 | 20GB | 40GB | 120GB | ... | 2023-12-22T03:54:01 | https://www.reddit.com/r/LocalLLaMA/comments/18o5u0k/helpful_vram_requirement_table_for_qlora_lora_and/ | Aaaaaaaaaeeeee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o5u0k | false | null | t3_18o5u0k | /r/LocalLLaMA/comments/18o5u0k/helpful_vram_requirement_table_for_qlora_lora_and/ | false | false | self | 130 | {'enabled': False, 'images': [{'id': 'KH8rP1laXrKX7kQH4ecJCjAPVRTJEoI3PEpjxFaNGbQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/eU5sgD7_61bCe717pQaE6Qr33QDNNyCz4qXGxrssAww.jpg?width=108&crop=smart&auto=webp&s=ef60162e69c0c9707b500c028514ac6a678840fe', 'width': 108}, {'height': 108, 'url': 'h... |
eGPU | 5 | Hi I have a total noob question about local llm. I have a laptop with a GFX that has 6 GB of VRAM. Unfortunately, I can't afford a new computer, I have an offer to buy an egpu T3 cheaply, I'm aware that egpu have a performance loss. But as I understand it, e.g. With llama.cpp the model is loaded into the vram and stays... | 2023-12-22T03:19:39 | https://www.reddit.com/r/LocalLLaMA/comments/18o57uz/egpu/ | Strong_Sympathy9955 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o57uz | false | null | t3_18o57uz | /r/LocalLLaMA/comments/18o57uz/egpu/ | false | false | self | 5 | null |
The "alignment" is becoming insane | 1 | [removed] | 2023-12-22T03:17:42 | https://www.reddit.com/r/LocalLLaMA/comments/18o56ou/the_alignment_is_becoming_insane/ | petrus4 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o56ou | false | null | t3_18o56ou | /r/LocalLLaMA/comments/18o56ou/the_alignment_is_becoming_insane/ | false | false | self | 1 | null |
Is there a penalty for having more experts in MoE (Mixtral)? | 6 | If the speed is derived from having the equivalent of eight 7b models, with two being used. Wouldn’t be faster and/or more accurate with a sixteen head 4b model that can opt to use 2, 3 or 4 experts? | 2023-12-22T03:14:10 | https://www.reddit.com/r/LocalLLaMA/comments/18o54bh/is_there_a_penalty_for_having_more_experts_in_moe/ | MachineZer0 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o54bh | false | null | t3_18o54bh | /r/LocalLLaMA/comments/18o54bh/is_there_a_penalty_for_having_more_experts_in_moe/ | false | false | self | 6 | null |
What are some models you have been running? On what hardware? and why? | 7 | Finally figured out how to actually use LocalAI and Ollama and textgen, stable diffusion, etc. Fair to say my 1660 6gb isnt the best, still works, but eh, so I'm in the market for a new AI rig. An A100 would be great, but I'd only ever get my hands on one in my dreams. But I wanna hear what hardware you are running in ... | 2023-12-22T02:00:44 | https://www.reddit.com/r/LocalLLaMA/comments/18o3p82/what_are_some_models_you_have_been_running_on/ | Mafyuh | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o3p82 | false | null | t3_18o3p82 | /r/LocalLLaMA/comments/18o3p82/what_are_some_models_you_have_been_running_on/ | false | false | self | 7 | null |
Upgrade to 96 GB from 64 GB Apple M3 Max for LLM inference, pre-training? | 1 | [removed] | 2023-12-22T01:19:51 | https://www.reddit.com/r/LocalLLaMA/comments/18o2wma/upgrade_to_96_gb_from_64_gb_apple_m3_max_for_llm/ | Adventurous-Sign4520 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o2wma | false | null | t3_18o2wma | /r/LocalLLaMA/comments/18o2wma/upgrade_to_96_gb_from_64_gb_apple_m3_max_for_llm/ | false | false | 1 | null | |
How much context can mistral and its finetunes handel ? | 4 | Uts probably very simple but I am a bit lost at the moment up to which context size can mistral models be used and is there a difference between v 0.1 and 0.2 ?
Lastly I am using the open hermes 2.5 fine tune up to what size of Context can it handle ? | 2023-12-22T01:10:17 | https://www.reddit.com/r/LocalLLaMA/comments/18o2pwb/how_much_context_can_mistral_and_its_finetunes/ | Noxusequal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o2pwb | false | null | t3_18o2pwb | /r/LocalLLaMA/comments/18o2pwb/how_much_context_can_mistral_and_its_finetunes/ | false | false | self | 4 | null |
Open Source RAG - 500 users | 11 | Hi,
I have been sneaking inside this forum just looking at all the new open source models dropping and I thought theoretically what it would require implementing an open source on premise RAG solution with capacity to serve without performance jitter to 500 people.
Like a ChatGPT but I personally host the LLM solutio... | 2023-12-22T00:48:36 | https://www.reddit.com/r/LocalLLaMA/comments/18o29ju/open_source_rag_500_users/ | pathfinder6709 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o29ju | false | null | t3_18o29ju | /r/LocalLLaMA/comments/18o29ju/open_source_rag_500_users/ | false | false | self | 11 | null |
Is there any cheap mobo/cpu combo recommended for running 2-3 tesla p40s? | 2 | I've got 2 P40s but just found out my current mobo can't handle more than 1. Since my cpu is a i5-11400f it would also need a gpu so I'd need 3 slots minimum, and ideally 4 for the 3rd card.
Since I've already got everything else (ram, gpu, psu, ssd), I'm wondering if there's any recommended cheap mobo/cpu set I can b... | 2023-12-21T23:41:05 | https://www.reddit.com/r/LocalLLaMA/comments/18o0tqc/is_there_any_cheap_mobocpu_combo_recommended_for/ | TopRecognition9302 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o0tqc | false | null | t3_18o0tqc | /r/LocalLLaMA/comments/18o0tqc/is_there_any_cheap_mobocpu_combo_recommended_for/ | false | false | self | 2 | null |
Fine Tuning Mixtral Dolphin? | 3 | Anyone succeeded/any write-ups? I've only found this tutorial, which is bit poor [https://predibase.com/blog/how-to-fine-tune-mixtral-8x7b-with-open-source-ludwig](https://predibase.com/blog/how-to-fine-tune-mixtral-8x7b-with-open-source-ludwig) | 2023-12-21T23:23:27 | https://www.reddit.com/r/LocalLLaMA/comments/18o0fc0/fine_tuning_mixtral_dolphin/ | Otherwise-Tiger3359 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o0fc0 | false | null | t3_18o0fc0 | /r/LocalLLaMA/comments/18o0fc0/fine_tuning_mixtral_dolphin/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': '5cdtkjF4UabknBVHv0jzqKKGN18nxOccQ_96tWdVrrA', 'resolutions': [{'height': 57, 'url': 'https://external-preview.redd.it/pjM3BEvfnHV9vja2n6geFvR-DTXgu3ZKPMzjNK8bmH8.jpg?width=108&crop=smart&auto=webp&s=22a3be2b589ff22fb70df84c3e1d34dd609280de', 'width': 108}, {'height': 114, 'url': 'h... |
mixtral q4 k-m really braind dead for me compared to mistral 7b f16 | 22 | clblast isnt working so i cant do more than q4 for now
mistral 7b has been really great so i had hoped for mixtral to massivly out perform it but for now its the other way around despite most charts claiming even q4 b13 models should outperform it
and here we have a model which is said to outperform 70b ... | 2023-12-21T23:21:11 | https://www.reddit.com/r/LocalLLaMA/comments/18o0dm8/mixtral_q4_km_really_braind_dead_for_me_compared/ | Zealousideal_Nail288 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 18o0dm8 | false | null | t3_18o0dm8 | /r/LocalLLaMA/comments/18o0dm8/mixtral_q4_km_really_braind_dead_for_me_compared/ | false | false | self | 22 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.