pseudotensor
Follow
Popular repositories
5,194 contributions in the last year
Activity overview
Contributed to
h2oai/h2ogpt,
h2oai/driverlessai-recipes,
gradio-app/gradio
and 15 other
repositories
Contribution activity
June 2023
Created 243 commits in 1 repository
Created 1 repository
- pseudotensor/falcontune Python
Created a pull request in h2oai/h2ogpt that received 13 comments
Add support for text-generation-server
Add gradio servers as inference_server option Add ability to lock model name to an endpoint from CLI, instead of all independent, and then hide …
+683
−114
•
13
comments
Opened 24 other pull requests in 1 repository
h2oai/h2ogpt
24
merged
- Fixes #279
- Option to fill-up context if top_k_docs=-1
- Add QUIP et al. metrics for context-Q-A testing
- Fixes #291 -- make user_path if doesn't exist but passed, and move gradio temp file to user_path if passed to generate.
- Pass actual eos id to generate, else doesn't know how to stop early if using non-standard eos id (normally=0, falcon GM was 11)
- Update requirements and add code to get latest versions
- Update Weaviate and FAISS a bit to be closer to Chroma in h2oGPT with limitations. Add testing.
- Fixes #270
- Upgrade gradio
- Ensure generate hyperparameters are passed through to h2oai_pipelinepy for generation
- move enums and add test for export copy since keep changing what files have what structures
- Fixes #249
- Add extra clear torch cache calls so embedding on GPU doesn't stick to GPU
- use instructor embedding, and add migration of embeddings if ever changes, at least for chroma
- Improve subset words and code
- Fix nochat in UI so enter works to submit again, and if langchain mode used then shows HTML links for sources
- Chunk improve
- Make API easier, and add prompt_dict for custom control over prompt as example of new API parameter don't need to pass
- More control over max_max_new_tokens and memory behavior from generate args
- Allow update db from UI if files changed, since normally not constantly checking for new files
- More weaviate tests
- Add setup.py wheel building option
- Fix run_eval and validate parameters are all passed
- Allow for custom eval json file
Reviewed 15 pull requests in 1 repository
h2oai/h2ogpt
15 pull requests
- Update README.md
- Add ci support + wheel
- Add h2oGPT Client
-
Escape so it outputs
$LD_LIBRARY_PATH:/usr/local/cuda/lib64/ - Add documentation on how to connect to weaviate
- A couple of small updates to the documentation
- prompt form moved to a separate file
- input prompt - primary color border added + change in label text
- Submit button is now primary + more spacing between prompt area and action buttons
- venv support in gitignore&Makefile
- .gitignore updated for .idea and venv
- Support connecting to a local weaviate instance
-
Fix
TypeError: can only concatenate str (not "list") to stron startup - Use optional dependencies
- Add weaviate support
Created an issue in huggingface/peft that received 5 comments
just import of peft leads to use of cuda and introduces cuda context that makes forks not possible
Older peft did not do this, but newer does, at least since 3714aa2 and likely earlier. This means if import in global scope as normal, no forks are…
5
comments
Opened 24 other issues in 3 repositories
h2oai/h2ogpt
16
open
6
closed
- if hit stop, write uuid file that stopping.py can detect and stop, so generation really stops
- control max_time limits for API as well
- add PDF test cases
- check on performance of gradio generation
- Improve UI
- use API endpoints as model choices, including OpenAI
- add other meta data to get system info
- E ImportError: cannot import name 'DEFAULT_CIPHERS' from 'urllib3.util.ssl_'
- revert when langchain keep_separator works
- For h2oai_pipeline.py, don't truncated max_new_tokens given input, just use stopping condition to stop after total tokens reaches max
- RuntimeError: expected scalar type Half but found Float
- need to filter data stuffed into context, it pushes LLM into bad outputs
- PDF makes wraps lines, then LLM does same in response
- assert db is not None or metadatas is not None
- hit just enter and if go through db, get garbage answer in another language
- avoid verbose print from instructor loader
-
ValueError: We cannot use
holeto handle this generation the number of desired tokens exceeds the models max length - add continue button
- make pipeline look at model name and lookup prompt_type if exists
- connect to elastic search
- gradio generation of html leads to odd UI content
- Better post-quantization performance
gradio-app/gradio
1
closed
PromtEngineer/localGPT
1
open
8
contributions
in private repositories
Jun 5 – Jun 16