Xuan-Son Nguyen
159b741427
Copilot Setup Steps / copilot-setup-steps (push) Has been cancelled
Check Pre-Tokenizer Hashes / pre-tokenizer-hashes (push) Has been cancelled
Python check requirements.txt / check-requirements (push) Has been cancelled
Python Type-Check / python type-check (push) Has been cancelled
Update Operations Documentation / update-ops-docs (push) Has been cancelled
* training: fix no KV cache * apply @ ggerganov suggestion
llama.cpp/examples/training
This directory contains examples related to language model training using llama.cpp/GGML. So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP. Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory. For CPU training, compile llama.cpp without any additional backends such as CUDA. For CUDA training, use the maximum number of GPU layers.
Flash attention is disabled during training because FLASH_ATTN_EXT has no backward pass.
Proof of concept:
export model_name=llama_3.2-1b && export quantization=f32
./build/bin/llama-finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
./build/bin/llama-perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.