Llama.cpp for fun (and maybe profit) - 30 minute

Presented at https://www.meetup.com/ai-and-dl-for-enterprise/events/299910519/ to show llama.cpp running quantised models locally, no internet connection (and no data exfiltration) to summarise text to keywords (llama2 7b), solve physics (phi-2), augment and fact-extract from images (llava), summarise code behaviour (codellama 34b), dig into the Python API to extract embeddings and see next-token generation and probability assignments to reveal the underlying training data structure. Includes a discussion on how quantisation works.
By: https://ianozsvald.com/ and it'll be written up at https://notanumber.email/

Transcript

1. llama.cpp for fun and (maybe) profit AI & DL for Enterprise 2024-04 talk @IanOzsvald – ianozsvald.com
2. Strategist/Trainer/Speaker/Author 20+ years Trainer – new Fast Pandas course Interim Chief Data Scientist By [ian]@ianozsvald[.com] Ian Ozsvald 2nd Edition!
3. When should we be using LLMs? Strategic Team Advisor By [ian]@ianozsvald[.com] Ian Ozsvald
4. No need for a GPU+VRAM Llama.cpp runs on CPU+RAM Nothing sent off your machine llama.cpp By [ian]@ianozsvald[.com] Ian Ozsvald X
5. Experiment with models as they’re published Use client data/src code – no data sent off your machine Why use local models? By [ian]@ianozsvald[.com] Ian Ozsvald
6. Prototype ideas! By [ian]@ianozsvald[.com] Ian Ozsvald llama-2-7b-chat.Q5_K_M.gguf 5GB on disk and in RAM, near real time
7. See the wackyness early on What’s your strategy to catch varied outputs? Why use local models? By [ian]@ianozsvald[.com] Ian Ozsvald
8. MS Phi2 can “reason” (IS IT RIGHT?) By [ian]@ianozsvald[.com] Ian Ozsvald I had confident answers: 125.2m/s (good Python) 17.2m/s (partial Python with comments that had mistakes), 40m/s and 31.3m/s (as teacher) Which one to believe? My model is quantised (Q5) but random variation exists anyway… The MS post didn’t disclose the prompt they used https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/
9. Similar to JPG compression Shrink the trained model 32→16→8→7/6/5/4/3/2 bits Fewer bits→worse text completion “Q5 generally an acceptable level” Quantisation By [ian]@ianozsvald[.com] Ian Ozsvald
10. Quantisation By [ian]@ianozsvald[.com] Ian Ozsvald Original fp16 models Better Bigger models with higher quantisation still has lower perplexity than simpler, less quantised models Choose the biggest you can K-quants PR https://github.com/ggerganov/llama.cpp/pull/1684
11. Experiment with multi-modal e.g. OCR and checking photo meets requirements What about image queries? By [ian]@ianozsvald[.com] Ian Ozsvald
12. Llava multi-modal Extract facts from images? By [ian]@ianozsvald[.com] Ian Ozsvald llava-v1.5-7b-Q4_K.gguf 4GB on disk & RAM 5s for example llama.cpp provides ./server
13. Trial code-support Code review? “Is this test readable?” What do you do with code and LLMs? Can they help with coding? By [ian]@ianozsvald[.com] Ian Ozsvald
14. Can you explain this function please? By [ian]@ianozsvald[.com] Ian Ozsvald codellama-34b-python.Q5_K_M.gguf 23GB on disk & RAM, 30s for example Can we use this as a “code reviewer” for internal code? codellama answer: “The function test_uniform_distribution creates a list of 10 zeros, then increments the position in that list indicated by the murmurhash3_32() digest of i. It does this 100000 times and then checks if the means of those incremented values are uniformly distributed (i.e., if they're all roughly the same).” (surprisingly clear!) https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/utils/tests/test_murmurhash.py
15. Give test functions (e.g. Pandas) to codellama Ask it “is this a good test function?” Try to get it to propose new test functions Check using pytest and coverage tools Shortcut human effort at project maintenance? My experiment for code assist By [ian]@ianozsvald[.com] Ian Ozsvald
16. Using the Python API we can learn how it works Get embeddings API leads to understanding By [ian]@ianozsvald[.com] Ian Ozsvald
17. Q&A model trained on “Let, What, Suppose, Calculate, Solve” as very-likely first tokens API leads to understanding By [ian]@ianozsvald[.com] Ian Ozsvald log(p(1)) == 0 log(p(0.5)) == -0.7
18. Run quantised models on client data locally Experience the wackyness – mitigation? Use Python API to see tokens+perplexity+more Why try llama.cpp? By [ian]@ianozsvald[.com] Ian Ozsvald
19. Do you want to talk about training or DS strategy? Discuss: How do we measure correctness? What’s the worst (!) that could go wrong with your projects? Summary By [ian]@ianozsvald[.com] Ian Ozsvald
20. Appendix - Attack via asciiart By [ian]@ianozsvald[.com] Ian Ozsvald
21. Appendix – Ask Mixtral to challenge my Monte Carlo estimation approach By [ian]@ianozsvald[.com] Ian Ozsvald Mixtral gave 5 points and some items I should be careful about, ChatGPT 3.5 gave 7 points, both felt similar
22. WizardCoder is good (tuned llama2) By [ian]@ianozsvald[.com] Ian Ozsvald wizardcoder-python-34b -v1.0.Q5_K_S.gguf 22GB on disk & RAM 15s for example You can replace CoPilot with this for completions