tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-07 12:52:16 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

aling scripts with speedrun.sh

Browse Source
This commit is contained in:
Shizhe Diao 2025-10-23 06:33:23 -07:00
parent a5f3ccc3ca
commit ed8d73a154
1 changed files with 20 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
speedrun_nemotron.sh
View File

@ -1,23 +1,24 @@
#!/bin/bash #!/bin/bash
# This script is the "Best ChatGPT clone that $100 can buy", # This script is based on the "Best ChatGPT clone that $100 can buy" and modified for the nemotron datasets.
# It is designed to run in ~4 hours on 8XH100 node at $3/GPU/hour. # It is designed to run in ~4 hours on 8XH100 node at $3/GPU/hour.
# 1) Example launch (simplest): # 1) Example launch (simplest):
# bash speedrun.sh # bash speedrun_nemotron.sh
# 2) Example launch in a screen session (because the run takes ~4 hours): # 2) Example launch in a screen session (because the run takes ~4 hours):
# screen -L -Logfile speedrun.log -S speedrun bash speedrun.sh # screen -L -Logfile speedrun_nemotron.log -S speedrun_nemotron bash speedrun_nemotron.sh
# 3) Example launch with wandb logging, but see below for setting up wandb first: # 3) Example launch with wandb logging, but see below for setting up wandb first:
# WANDB_RUN=speedrun screen -L -Logfile speedrun.log -S speedrun bash speedrun.sh # WANDB_RUN=speedrun_nemotron screen -L -Logfile speedrun_nemotron.log -S speedrun_nemotron bash speedrun_nemotron.sh
set -x set -x
DATA_NAME=climbmix_small PRE_TRAINING_DATA_NAME=nemotron_climbmix
DATA_DIR=/lustre/fsw/portfolios/nvr/users/sdiao/nanochat/data/$DATA_NAME
POST_TRAINING_DATA_NAME=nemotron POST_TRAINING_DATA_NAME=nemotron
TOKENIZER_NAME="tokenizer_${PRE_TRAINING_DATA_NAME}"
# Default intermediate artifacts directory is in ~/.cache/nanochat # Default intermediate artifacts directory is in ~/.cache/nanochat
export OMP_NUM_THREADS=1 export OMP_NUM_THREADS=1
export NANOCHAT_BASE_DIR="/lustre/fsw/portfolios/nvr/users/sdiao/nanochat/.cache" export NANOCHAT_BASE_DIR="$HOME/.cache/nanochat"
export PRE_TRAINING_DATA_DIR="$NANOCHAT_BASE_DIR/${PRE_TRAINING_DATA_NAME}"
mkdir -p $NANOCHAT_BASE_DIR mkdir -p $NANOCHAT_BASE_DIR
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
@ -28,7 +29,7 @@ command -v uv &> /dev/null || curl -LsSf https://astral.sh/uv/install.sh | sh
# create a .venv local virtual environment (if it doesn't exist) # create a .venv local virtual environment (if it doesn't exist)
[ -d ".venv" ] || uv venv [ -d ".venv" ] || uv venv
# install the repo dependencies # install the repo dependencies
uv sync uv sync --extra gpu
# activate venv so that `python` uses the project's venv instead of system python # activate venv so that `python` uses the project's venv instead of system python
source .venv/bin/activate source .venv/bin/activate
@ -38,13 +39,11 @@ source .venv/bin/activate
# 1) Make sure to first log in to wandb, e.g. run: # 1) Make sure to first log in to wandb, e.g. run:
# `wandb login` # `wandb login`
# 2) Set the WANDB_RUN environment variable when running this script, e.g.: # 2) Set the WANDB_RUN environment variable when running this script, e.g.:
# `WANDB_RUN=d26 bash speedrun.sh` # `WANDB_RUN=d26 bash speedrun_nemotron.sh`
# if [ -z "$WANDB_RUN" ]; then if [ -z "$WANDB_RUN" ]; then
# # by default use "dummy" : it's handled as a special case, skips logging to wandb # by default use "dummy" : it's handled as a special case, skips logging to wandb
# WANDB_RUN=dummy WANDB_RUN=dummy
# fi fi
export WANDB_API_KEY="ec7a9c0701d404122e4fc5c7c7518ed17f5b03ca"
export WANDB_RUN=${DATA_NAME}_d20_1node
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# During the course of the run, we will be writing markdown reports to the report/ # During the course of the run, we will be writing markdown reports to the report/
@ -62,19 +61,10 @@ source "$HOME/.cargo/env"
# Build the rustbpe Tokenizer # Build the rustbpe Tokenizer
uv run maturin develop --release --manifest-path rustbpe/Cargo.toml uv run maturin develop --release --manifest-path rustbpe/Cargo.toml
# # Download the first ~2B characters of pretraining dataset # download and process the nemotron_climbmix data
# # look at dev/repackage_data_reference.py for details on how this data was prepared python data/nemotron_climbmix/process_climbmix.py --data_dir=$PRE_TRAINING_DATA_DIR
# # each data shard is ~250M chars
# # so we download 2e9 / 250e6 = 8 data shards at this point python -m scripts.tok_train --max_chars=2000000000 --data_dir=$PRE_TRAINING_DATA_DIR --tokenizer_name=$TOKENIZER_NAME
# # each shard is ~100MB of text (compressed), so this is about ~800MB of data on disk
# python -m nanochat.dataset -n 8
# # Immediately also kick off downloading more shards in the background while tokenizer trains
# # See comment below for why 240 is the right number here
# python -m nanochat.dataset -n 240 &
# DATASET_DOWNLOAD_PID=$!
# train the tokenizer with vocab size 2**16 = 65536 on ~2B characters of data
export TOKENIZER_NAME="tokenizer_${DATA_NAME}"
python -m scripts.tok_train --max_chars=2000000000 --data_dir=$DATA_DIR --tokenizer_name=$TOKENIZER_NAME
# evaluate the tokenizer (report compression ratio etc.) # evaluate the tokenizer (report compression ratio etc.)
python -m scripts.tok_eval --tokenizer_name=$TOKENIZER_NAME python -m scripts.tok_eval --tokenizer_name=$TOKENIZER_NAME
@ -100,9 +90,9 @@ echo "Waiting for dataset download to complete..."
wait $DATASET_DOWNLOAD_PID wait $DATASET_DOWNLOAD_PID
# pretrain the d20 model # pretrain the d20 model
torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=20 --run=$WANDB_RUN --data_dir=$DATA_DIR --tokenizer_name=$TOKENIZER_NAME --model_tag=$WANDB_RUN torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=20 --run=$WANDB_RUN --data_dir=$PRE_TRAINING_DATA_DIR --tokenizer_name=$TOKENIZER_NAME --model_tag=$WANDB_RUN
# evaluate the model on a larger chunk of train/val data and draw some samples # evaluate the model on a larger chunk of train/val data and draw some samples
torchrun --standalone --nproc_per_node=8 -m scripts.base_loss --data_dir=$DATA_DIR --tokenizer_name=$TOKENIZER_NAME --model_tag=$WANDB_RUN torchrun --standalone --nproc_per_node=8 -m scripts.base_loss --data_dir=$PRE_TRAINING_DATA_DIR --tokenizer_name=$TOKENIZER_NAME --model_tag=$WANDB_RUN
# evaluate the model on CORE tasks # evaluate the model on CORE tasks
torchrun --standalone --nproc_per_node=8 -m scripts.base_eval --model_tag=$WANDB_RUN torchrun --standalone --nproc_per_node=8 -m scripts.base_eval --model_tag=$WANDB_RUN