h2o-llmstudio
h2o-llmstudio copied to clipboard
h2oai
H2O LLM Studio - a framework and no-code GUI for fine-tuning LLMs. Documentation: https://docs.h2o.ai/h2o-llmstudio/
I trained a 33b model with DeepSpeed on 40GB cards. Based on the traceback, the model seems to be too large to fit into one GPU. Is it possible to...
### 🚀 Feature Similar to EOS token, we should offer an option to add BOS token to the beginning. Might be useful for models like Gemma.
### 🚀 Feature Support for specification of a minimum learning rate ### Motivation Often in the research literature minimum learning rates are set when fine-tuning a model using a cosine...
### 🐛 Bug Here: https://github.com/h2oai/h2o-llmstudio/blob/a9d72fffef370c7fd0dd9f29ece06ab45b6c5815/train.py#L562 Let's say total data batches are 160 and gradient accumulation is 10. Optimization step is happening only 10 times. But here scheduler is called every...
### 🚀 Feature Instead of using 6 or more different settings to control how the input can be transformed for the LLM, we can allow the use of chat templates....
### 🚀 Feature Add support for Weight-Decomposed Low-Rank Adaptation (DoRA). "DoRA decomposes the pre-trained weight into two components, magnitude and direction, for fine-tuning, specifically employing LoRA for directional updates to...
### 🐛 Bug The generated `tokenizer_config.json` has `add_bos_token=true` while H2O LLM Studio is training with `add_special_tokens=False`. Using the default AutoTokenizer, this leads to different behaviors. We should be explicit/correct about...
### 🚀 Feature New advancements bringing quantized LoRA and FSDP together. https://github.com/AnswerDotAI/fsdp_qlora ### Motivation Train larger models on consumer GPUs or older generation Data Center GPUs such as V100 Lets...
Reopened the same PR #556 with correct local branch as suggested by @psinger. Adding support for custom loss functions aimed at improving the length consistency in responses generated by fientuned...
