justheuristic

Follow

💭

on github.com/learning-at-home

justheuristic justheuristic

💭

on github.com/learning-at-home

Follow

building the hivemind

320 followers · 6 following

Achievements

BetaSend feedback

Achievements

BetaSend feedback

Organizations

Block or Report

Block or report justheuristic

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse

Pinned

yandexdataschool/Practical_RL Public

A course in reinforcement learning in the wild

Jupyter Notebook 5.4k 1.6k
prefetch_generator Public

Simple package that makes your generator work in background thread

Python 257 22
yandexdataschool/nlp_course Public

YSDA course in Natural Language Processing

Jupyter Notebook 8.7k 2.4k
pytorch/pytorch Public

Tensors and Dynamic neural networks in Python with strong GPU acceleration

Python 69k 18.9k

631 contributions in the last year

Contribution Graph
	Jul			Aug				Sep				Oct					Nov				Dec				Jan					Feb				Mar				Apr					May				Jun				Jul
Sun
Mon
Tue
Wed
Thu
Fri
Sat

Learn how we count contributions

Activity overview

Contributed to bigscience-workshop/petals, BlackSamorez/tensor_parallel, TimDettmers/bitsandbytes and 20 other repositories

Contribution activity

July 2023

Created 8 commits in 2 repositories

Created 1 repository

justheuristic/peft Python Jul 19

Created a pull request in Vahe1994/SpQR that received 3 comments

Rollback model loading to match the code from the paper

This code fixes bad perplexity that was found with the following config CUDA_VISIBLE_DEVICES=3 OMP_NUM_THREADS=16 MKL_NUM_THREADS=16 python main.py…

+11 −3 • 3 comments

Opened 6 other pull requests in 2 repositories

bigscience-workshop/petals 5 merged

Inherit bitsandbytes compute dtype correctly (override peft quirk) Jul 19
Fix handler memory leak, get rid of mp.Manager Jul 19
Switch adapters slightly faster Jul 14
Remove unused import os Jul 14
Estimate adapter memory overhead in choose_num_blocks() Jul 12

huggingface/peft 1 open

[minor bugfix] prevent lora Linear4bit from sneakily changing compute_dtype to float32 Jul 19

Reviewed 14 pull requests in 1 repository

bigscience-workshop/petals 14 pull requests

Require accelerate>=0.20.3 as transformers do Jul 19
Support Llama 2 Jul 19
Fix handler memory leak, get rid of mp.Manager Jul 19
Implement shortest-path routing for inference Jul 17
Make a server ping next servers Jul 15
Switch adapters slightly faster Jul 14
Estimate adapter memory overhead in choose_num_blocks() Jul 13
Spam less in server logs Jul 13
Import petals.utils.peft only when needed to avoid unnecessary import of bitsandbytes Jul 12
Fix convergence issues and switch to LLaMA in the SST-2 example Jul 12
Use bnb==0.40.0.post4 to fix bias bug, use bfloat16 by default Jul 12
Support peft LoRA adapters Jul 12
Implement direct server-to-server communication Jul 10
Support loading blocks in 4-bit (QLoRA NF4 format, disabled by default) Jul 3

Opened 1 issue in 1 repository

bigscience-workshop/petals 1 open

rpc_push may fail due to exceeding max message size Jul 19