Mike Wang

[Feature] Updated HuggingFace LLM Creator to Support Device Customization

1

- As described in HuggingFace source code, `device` parameter supports multiple input types. (Link: https://github.com/huggingface/transformers/blob/v4.28.1/src/transformers/pipelines/__init__.py#L605-L607) - In this PR, it enables `cpu` only acceleration for older laptops which doesn’t have...

[test] add a new test for LLM class

2

- added unit test cases for LLM base class. - covering both realtime and async entry points.

[fix] GPT sometimes ignores the SQL format request.

2

I found GPT3.5 sometimes consistently ignores the requested SQL format - “Action *** Action Input ***”, instead provided “Action ***, “”” when there is no Action Input. This will cause...

This PR is on top of pr/4381. Working on multiple new features for Spark Agent. # Feature 1: Spark Flexible Agent **Rationale:** The classical ZeroShotAgent always forces to use a...

- simplify the validation check a little bit. - re-tested in jupyter notebook. Reviewer: @hwchase17

[jupyter] add memory example to sql_db agent

3

- Memory can largely enrich Agent convenience. - This PR added an example showing how to use Memory in SQL DB Agent. - Fixed a bug in the old SQLDatabaseToolkit...

[fix doc] AC is enabled by default

3

#### Context - fix the statement as AC is now enabled by default 13b full config https://github.com/pytorch/torchtune/blob/main/recipes/configs/llama2/13B_full.yaml#L75 7b full config https://github.com/pytorch/torchtune/blob/main/recipes/configs/llama2/7B_full.yaml#L70 #### Changelog - updated full fine tuning distributed statement...

[RFC][fix test] missing .item() in frozen nf4 test

5

#### Context - as titled. missing .item() - but there is another thing. I found the error from frozen-nf4 is larger than the one from bnb-linear #### Changelog - updated...