maxjeblick

Link to pretrained models

3

Are you planning to publish pretrained models?

[CODE IMPROVEMENT] Store weights in AutoModelForCausalLM format

5

### 🔧 Proposed code refactoring Currently, model weights are stored in LLm Studio format which is a small wrapper around `AutoModelForCausalLM`. Instead, store model weights in `AutoModelForCausalLM` format, as well...

This PR converts pickled configurations into yaml configuration. It introduces a breaking change to the current UI (meaning that old experiments won't work). As the repo is in its initial...

### 🔧 Proposed code refactoring Move from pickle format of saving `cfg_last.p` to yaml (or similar format). ### Motivation - Human readable data format - Better comparability to possible code...

Fixes #31

### 🔧 Proposed code refactoring Update Starlette to v 0.25.0 ### Motivation Fix `Starlette allows an unauthenticated and remote attacker to specify any number of form fields or files` security...

add requirements.txt

1

This PR adds `requirements.txt` file that is generated by `pipenv requirements > requirements.txt` In future, we can add a CI test that automatically checks if `requirements.txt` is synced with `pipfile.lock`....

Add falcon peft target modules

4

This PR adds default lora target layers for falcon models. I have excluded MPT models, as they seem to require additional code changes.

[BUG] GPU ids are not checked when using cfg.yaml

2

### 🐛 Bug Starting a new experiment from `cfg.yaml` causes an error if number of gpu specified in `cfg.yaml` exceeds the number of gpus on the target machine. Within the...

[CODE IMPROVEMENT] Check default RLHF parameters

3

### 🔧 Proposed code refactoring Check if our default hyperparameters (e.g. [kl_target](https://github.com/h2oai/h2o-llmstudio/blob/main/llm_studio/python_configs/text_causal_language_modeling_config.py#L154)) are correct, see: https://github.com/lvwerra/trl/commit/b56e8b327733baa81c3ef0d6508f08e1b3e33939 and https://github.com/lvwerra/trl/issues/462 Also, RLHF training is quite unstable w.r.t. parameter choices, see e.g. issues...