File size: 3,246 Bytes
0da1f00 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | #!/usr/bin/env bash
# One-shot setup for a fresh 8xB200 node. Idempotent: safe to re-run.
#
# bash /code/users/alexsu/project/fastkernels/setup_node.sh
#
# Everything this touches except /code is node-local and disappears when the
# node is recycled, which is why this script exists.
set -uo pipefail
set -x
LDAP=alexsu
REPO="/code/users/${LDAP}/project/fastkernels"
TOKEN_FILE="/code/users/${LDAP}/hf_token"
# 1. Fail fast if the GPUs aren't visible.
nvidia-smi -L || { echo "GPU not visible, aborting"; exit 1; }
# 2. Convenience symlinks.
ln -sfn /data-fast "${HOME}/data-fast"
ln -sfn "/code/users/${LDAP}" "${HOME}/${LDAP}"
ln -sfn "${REPO}" "${HOME}/fastkernels"
# 3. HF token: the copy on FSX is the source of truth.
mkdir -p /data-fast/huggingface
cp "${TOKEN_FILE}" /data-fast/huggingface/token
chmod 600 /data-fast/huggingface/token
# 4. Shell env.
if ! grep -q "FASTKERNELS_ENV_SETUP" "${HOME}/.bashrc"; then
cat >> "${HOME}/.bashrc" <<'BASHRC_EOF'
# --- FASTKERNELS_ENV_SETUP ---
ulimit -c 0
export HF_HOME=/home/yak/data-fast/huggingface
export HF_XET_HIGH_PERFORMANCE=1
export CPLUS_INCLUDE_PATH="/usr/local/cuda/include/cccl:${CPLUS_INCLUDE_PATH:-}"
export C_INCLUDE_PATH="/usr/local/cuda/include/cccl:${C_INCLUDE_PATH:-}"
export FASTKERNELS_RESULTS_DIR=/code/users/alexsu/vllm-validate
export FK_VENV_ROOT=/data-fast/fk-venvs
export UV_CACHE_DIR=/data-fast/uv-cache
# Drop the K8s service vars; vLLM tries to parse anything named VLLM_*.
unset $(env | cut -d= -f1 | grep -E '^(SOYOUNG_)?(VLLM|QWEN)_[A-Z0-9]+_.*(SERVICE|PORT)') 2>/dev/null
# --- end ---
BASHRC_EOF
fi
# 5. GitHub over 443; port 22 is blocked on this cluster. The github.com block
# must come before the Host * wildcard or it never applies.
mkdir -p "${HOME}/.ssh"; touch "${HOME}/.ssh/config"
if ! grep -q '^Host github.com' "${HOME}/.ssh/config"; then
{ printf 'Host github.com\n HostName ssh.github.com\n Port 443\n ForwardAgent yes\n\n';
cat "${HOME}/.ssh/config"; } > "${HOME}/.ssh/config.new"
mv "${HOME}/.ssh/config.new" "${HOME}/.ssh/config"
chmod 600 "${HOME}/.ssh/config"
fi
ssh -T -o StrictHostKeyChecking=accept-new git@github.com || true
# 6. Directories: results on FSX, venvs and uv cache on the node-local disk.
mkdir -p "/code/users/${LDAP}/vllm-validate" /data-fast/fk-venvs /data-fast/uv-cache
export UV_CACHE_DIR=/data-fast/uv-cache
# 7. Repo (persists on FSX, so this is a no-op after the first node).
if [ ! -d "${REPO}/.git" ]; then
git clone --recursive ssh://git@github.com/Snowflake-AI-Research/fastkernels.git "${REPO}"
fi
cd "${REPO}"
# 8. Dependencies. pyproject pins torch/vllm/transformers and the ABI-matched
# flash-attn wheel, so `-e .` is all that's needed -- but deep-gemm builds
# without build isolation, so torch has to already be importable.
uv pip install torch==2.11.0
uv pip install -e .
set +x
echo "=== verify ==="
python -c "import torch, vllm, flash_attn, fastkernels, transformers; \
print('torch', torch.__version__, '| cuda?', torch.cuda.is_available(), torch.cuda.device_count()); \
print('vllm', vllm.__version__, '| fa', flash_attn.__version__, '| hf', transformers.__version__)"
echo "=== done. open a new shell (exec bash -l) before running anything. ==="
|