h2o-llmstudio

h2o-llmstudio copied to clipboard

### 🚀 Feature Trigger UI tests as github action for PRs. Can be split into two parts: - add github workflow that runs all cpu-compatible tests - add a self-hosted...

maxjeblick

Generate settings and MoE Loss

1

This PR addresses the following: New `max_time` setting for generation allowing to specifiy a max second time per generation. Closes https://github.com/h2oai/h2o-llmstudio/issues/568 New `prompt_lookup_num_tokens` as discussed in https://twitter.com/joao_gante/status/1747322413006643259 Will likely only...

psinger

### 🚀 Feature Implement MoE Aux Loss for Router https://github.com/huggingface/transformers/blob/v4.36.1/src/transformers/models/mixtral/modeling_mixtral.py#L76

psinger

[CODE IMPROVEMENT] if user specifies a “system” column but it doesnt exist, it should error out instead of continue running silently

5

### 🔧 Proposed code refactoring if system column not in train dataframe.coljmns or in valid columns, then error out ### Motivation Otherwise user might erroneously believe they are using a...

Quetzalcohuatl

[FEATURE]when saving multiple epochs add an epoch number suffix for when save best=False

3

### 🚀 Feature Saves multiple .pth on each checkpoint. Instead of overwriting every checkpoint.pth ### Motivation Often useful to see how model performs at each epoch/savepoint. For example when training...

Quetzalcohuatl

### 🔧 Proposed code refactoring Current code is convoluted by `if rank==0` statements for logging. We should write a wrapper for that.

psinger

[BUG] App gets unresponsive after bitsandbytes training error

1

### 🐛 Bug App hangs after hitting some quantization issue in bitsandbytes. The issue seems related to bitsandbytes, raising `AssertionErrors` in random parts of the training pipeline does not result...

maxjeblick

### 🚀 Feature We can automatically add chat template to tokenizer_config.json when pushing to HF or preparing the model download. ### Motivation Fully integrated chat template in transformers https://huggingface.co/docs/transformers/chat_templating

pascal-pfeiffer

### 🚀 Feature Add a functionality to re-score the existing validation predictions of an experiment using a different (or even same) metric. ### Motivation I regularly run into the use-case...

psinger

[CODE IMPROVEMENT] Add system prompt to the chat interface

2

### 🔧 Proposed code refactoring Add system prompt to the chat interface ### Motivation Removes the need to test system prompts outside of LLM Studio

pascal-pfeiffer