GregoryKanevsky (@DataGreg) header
GregoryKanevsky (@DataGreg) profile picture

recovering R-addict, dataholic, visual-maniac, so rather plain person

Dallas, TX
Joined September 2018
Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)Photo by GregoryKanevsky (@DataGreg)
Context window is one of key LLM features, but what vendors declare almost never the same as effective context length. Check out this graph showing @AI21's #Jamba's declared vs. effective #contextwindow
22
GregoryKanevsky retweeted
Hack by popular demand! Open-Source Agent #Hackathon with @AndrewYNg & @DeepLearningAI at @agihouse_org Saturday, *August 24th* Stellar schedule planned w/ secret guest speakers, alongside co-hosts, @GroqInc & @langchain. Reserve your spot now --> eu1.hubs.ly/H0bJRt70
1
6
16
4,154
GregoryKanevsky retweeted
📽️ New 4 hour (lol) video lecture on YouTube: "Let’s reproduce GPT-2 (124M)" youtu.be/l8pRSuU81PU The video ended up so long because it is... comprehensive: we start with empty file and end up with a GPT-2 (124M) model: - first we build the GPT-2 network - then we optimize it to train very fast - then we set up the training run optimization and hyperparameters by referencing GPT-2 and GPT-3 papers - then we bring up model evaluation, and - then cross our fingers and go to sleep. In the morning we look through the results and enjoy amusing model generations. Our "overnight" run even gets very close to the GPT-3 (124M) model. This video builds on the Zero To Hero series and at times references previous videos. You could also see this video as building my nanoGPT repo, which by the end is about 90% similar. Github. The associated GitHub repo contains the full commit history so you can step through all of the code changes in the video, step by step. github.com/karpathy/build-na… Chapters. On a high level Section 1 is building up the network, a lot of this might be review. Section 2 is making the training fast. Section 3 is setting up the run. Section 4 is the results. In more detail: 00:00:00 intro: Let’s reproduce GPT-2 (124M) 00:03:39 exploring the GPT-2 (124M) OpenAI checkpoint 00:13:47 SECTION 1: implementing the GPT-2 nn.Module 00:28:08 loading the huggingface/GPT-2 parameters 00:31:00 implementing the forward pass to get logits 00:33:31 sampling init, prefix tokens, tokenization 00:37:02 sampling loop 00:41:47 sample, auto-detect the device 00:45:50 let’s train: data batches (B,T) → logits (B,T,C) 00:52:53 cross entropy loss 00:56:42 optimization loop: overfit a single batch 01:02:00 data loader lite 01:06:14 parameter sharing wte and lm_head 01:13:47 model initialization: std 0.02, residual init 01:22:18 SECTION 2: Let’s make it fast. GPUs, mixed precision, 1000ms 01:28:14 Tensor Cores, timing the code, TF32 precision, 333ms 01:39:38 float16, gradient scalers, bfloat16, 300ms 01:48:15 torch.compile, Python overhead, kernel fusion, 130ms 02:00:18 flash attention, 96ms 02:06:54 nice/ugly numbers. vocab size 50257 → 50304, 93ms 02:14:55 SECTION 3: hyperpamaters, AdamW, gradient clipping 02:21:06 learning rate scheduler: warmup + cosine decay 02:26:21 batch size schedule, weight decay, FusedAdamW, 90ms 02:34:09 gradient accumulation 02:46:52 distributed data parallel (DDP) 03:10:21 datasets used in GPT-2, GPT-3, FineWeb (EDU) 03:23:10 validation data split, validation loss, sampling revive 03:28:23 evaluation: HellaSwag, starting the run 03:43:05 SECTION 4: results in the morning! GPT-2, GPT-3 repro 03:56:21 shoutout to llm.c, equivalent but faster code in raw C/CUDA 03:59:39 summary, phew, build-nanogpt github repo
391
2,147
414
15,208
1,529,102
If you're a developer, this is probably the first chance you'll get to hack using a model that was developed with Mamba. Prizes include up to $5K in credits to build apps with Jamba on AI21 Studio (and lots of fun swag too!). Come meet the AI21 team and hack with us.
16
Come join @AI21Labs #Jambathon at the @AGIHouseSF on June 22nd to hack away with Jamba, the groundbreaking fusion of Mamba and Transformer architectures. agihouse-app.web.app/events/…
1
38
Replying to @ph_singer
just like git never ceases to frustrate me, unix never ceases to amaze
126
Setup shell > echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.zshrc > echo '[[ -d $PYENV_ROOT/bin ]] && export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.zshrc > echo 'eval "$(pyenv init -)"' >> ~/.zshrc Activate virtualenv automatically in your project > pyenv local [virtual-env-name]
39
Upgrade pip > pip install --upgrade pip Install Python > pyenv install -s 3.10.13 Remove virtual env > pyenv virtualenv-delete [virtual-env-name]
1
63
Git remove local branches after they were deleted in remote repo likely by closing PRs : > git fetch --all -p; git branch -vv | grep ": gone]" | awk '{ print $1 }' | xargs -n 1 git branch -d
36
Install Python version > pyenv install --list | grep -F 3.10 > pyenv install -v 3.10.7
1
55
Check what is current virtual environment > pip -V
1
43
All while the virtual environment is active. Export installed packages into requirements.txt: > pip freeze > requirements.txt Uninstall a package <pkg_name> > pip uninstall <pkg_name>
1
62
Using pyenv + virtualenv : > pyenv versions # to list python versions, once > pyenv virtualenv [python version] [virtual env name] # create virtualenv, once > pyenv activate [virtual env name] # each time > pip install -r requirements.txt # once > pyenv deactivate # each time
1
78
Prompt is All You Need
39
Replying to @tunguz
looks like someone got lazy and ChatGPT-ed this sticker 😀
10
pyenv + virtualenv : > pyenv versions (once) > pyenv virtualenv <desired python version> <virtual env name> (once) > pyenv activate <virtual env name> (every time) > pip install -r requirements.txt (once) > pyenv deactivate (every time)
1
118
Artificial Intelligence: The Good, The Bad and The Ugly - Free public lecture on May 24 by Professor Abu-Mostafa of @Caltech to explain the science of AI in plain language and assess extreme scenarios. work.caltech.edu/watson #MachineLearning #ArtificialIntelligence #AI #chatGPT
9
37
2
102
7,058
is anything missing?
48
This is his last weekend with us
1
51