qwen math training code (#435)

* qwen math training code

* pre-commit

training/qwen-math/scripts/set/prepare.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+
+MAMBA_ENV="tina"
+eval "$(mamba shell hook --shell bash)" && mamba activate "${MAMBA_ENV}"
+echo "START TIME: $(date)"
+echo "PYTHON ENV: $(which python)"
+
+source "./scripts/set/set_vars.sh"
+
+PY_SCRIPT="./scripts/set/run_download_model.py"
+
+echo ""
+echo "Running script: ${PY_SCRIPT}"
+echo ""
+
+python "${PY_SCRIPT}"
+
+echo "END TIME: $(date)"
+echo "DONE"

training/qwen-math/scripts/set/run_download_model.py

@@ -0,0 +1,9 @@
+import os
+
+from huggingface_hub import snapshot_download
+
+if __name__ == "__main__":
+    CKPT_DIR = os.environ["CKPT_DIR"]
+
+    print("Downloading model ...")
+    snapshot_download(repo_id="Qwen/Qwen2.5-3B-Instruct", local_dir=f"{CKPT_DIR}/models/Qwen2.5-3B-Instruct/base")

training/qwen-math/scripts/set/set_env.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# python 3.10 + cuda 11.8.0
+# the execution order the following commands matter
+
+export MKL_NUM_THREADS=1
+export NUMEXPR_NUM_THREADS=1
+export OPENBLAS_NUM_THREADS=1
+export OMP_NUM_THREADS=1
+
+conda clean -a -y
+mamba clean -a -y
+pip install --upgrade pip
+pip cache purge
+
+# cuda, gcc/g++, torch
+# mamba install cuda -c nvidia/label/cuda-11.8.0 -y
+# mamba install gcc gxx -c conda-forge -y
+pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118
+
+# xformers
+pip install xformers==0.0.28.post3 --index-url https://download.pytorch.org/whl/cu118
+
+# vLLM pre-compiled with CUDA 11.8
+pip install https://github.com/vllm-project/vllm/releases/download/v0.7.2/vllm-0.7.2+cu118-cp38-abi3-manylinux1_x86_64.whl
+
+pip install deepspeed
+pip install flash-attn==2.7.3 --no-build-isolation
+pip install peft
+
+pip install "trl==0.15.2"
+pip install latex2sympy2_extended
+pip install "math_verify==0.5.2"
+pip install word2number
+pip install scipy
+
+pip install wandb
+pip install plotly
+pip install matplotlib
+pip install seaborn

training/qwen-math/scripts/set/set_env_eval.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+# python 3.11 & cuda 11.8
+
+export MKL_NUM_THREADS=1
+export NUMEXPR_NUM_THREADS=1
+export OPENBLAS_NUM_THREADS=1
+export OMP_NUM_THREADS=1
+
+conda clean -a -y
+mamba clean -a -y
+pip install --upgrade pip
+pip cache purge
+
+# mamba install cuda -c nvidia/label/cuda-11.8.0 -y
+# mamba install gcc gxx -c conda-forge -y
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
+
+pip install xformers --index-url https://download.pytorch.org/whl/cu118
+pip install vllm
+pip install flash-attn --no-build-isolation
+
+pip install accelerate
+pip install datasets
+pip install deepspeed
+pip install distilabel[vllm,ray,openai]
+pip install e2b-code-interpreter
+pip install einops
+pip install flake8
+pip install huggingface_hub
+pip install hf_transfer
+pip install isort
+pip install langdetect
+pip install latex2sympy2_extended
+pip install liger_kernel
+pip install "math_verify==0.5.2"
+pip install packaging
+pip install parameterized
+pip install peft
+pip install pytest
+pip install python-dotenv
+pip install ruff
+pip install safetensors
+pip install sentencepiece
+pip install transformers
+pip install trl@git+https://github.com/huggingface/trl.git
+pip install wandb

training/qwen-math/scripts/set/set_vars.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+
+export CUDA_LAUNCH_BLOCKING=1
+export DS_LOG_LEVEL=error
+export TOKENIZERS_PARALLELISM=false
+
+export NCCL_P2P_DISABLE=1
+export NCCL_SHM_DISABLE=1
+export NCCL_IB_DISABLE=1
+
+export MKL_THREADING_LAYER=GNU
+export MKL_NUM_THREADS=1
+export NUMEXPR_NUM_THREADS=1
+export OPENBLAS_NUM_THREADS=1
+export OMP_NUM_THREADS=1
+
+## basic setup for the env
+export PROJECT_PREFIX="/root/projects" # e.g. /home/username/projects
+export SCRATCH_PREFIX="/root/scratch" # e.g. /home/username/scratch
+mkdir -p "${PROJECT_PREFIX}" "${SCRATCH_PREFIX}"
+
+export PROJECT_NAME="rg-math"
+export CORE_POSTFIX="tina"
+export PROJECT_DIR="${PROJECT_PREFIX}/${PROJECT_NAME}"
+export PYTHONPATH="${PROJECT_DIR}":$PYTHONPATH
+export PYTHONPATH="${PROJECT_DIR}/${CORE_POSTFIX}":$PYTHONPATH
+mkdir -p "${PROJECT_PREFIX}/${PROJECT_NAME}"
+
+export CKPT_DIR="${PROJECT_DIR}/ckpts"
+export DATA_DIR="${PROJECT_DIR}/datasets"
+export OUTPUT_DIR="${PROJECT_DIR}/outputs"
+export LOGGING_DIR="${PROJECT_DIR}/logs"
+mkdir -p "${CKPT_DIR}" "${DATA_DIR}" "${OUTPUT_DIR}" "${LOGGING_DIR}"
+
+## wandb setup
+# export WANDB_API_KEY="TODO"
+export WANDB_PROJECT="${PROJECT_NAME}"
+export WANDB_DIR="${OUTPUT_DIR}"
+
+wandb login $WANDB_API_KEY
+
+export CACHE_DIR="${PROJECT_DIR}/.cache"
+export WANDB_CACHE_DIR="${CACHE_DIR}"
+export TRITON_CACHE_DIR="${CACHE_DIR}/triton_cache"
+
+## huggingface setup
+# export HF_TOKEN="TODO"
+git config --global credential.helper store
+huggingface-cli login --token $HF_TOKEN --add-to-git-credential
+
+export HF_HOME="${CACHE_DIR}/huggingface"
+export HUGGINGFACE_HUB_CACHE="${HF_HOME}/hub"
+export HF_DATASETS_CACHE="${HF_HOME}/datasets"