# Copy this file to .env and replace every REPLACE_* value for your cluster.
PROFILE=adaptive
WORKER_SSH=REPLACE_USER@10.100.10.2
WORKER_CHECKOUT=/home/REPLACE_USER/dual-spark-optimizations

# Direct 200GbE CX-7/RoCE fabric.
HEAD_ROCE_IP=10.100.10.1
WORKER_ROCE_IP=10.100.10.2
HEAD_NCCL_SOCKET_IFNAME=REPLACE_HEAD_CX7_INTERFACE
HEAD_NCCL_IB_HCA=REPLACE_HEAD_ROCE_HCA
HEAD_NCCL_IB_GID_INDEX=3
WORKER_NCCL_SOCKET_IFNAME=REPLACE_WORKER_CX7_INTERFACE
WORKER_NCCL_IB_HCA=REPLACE_WORKER_ROCE_HCA
WORKER_NCCL_IB_GID_INDEX=3
MASTER_PORT=25002

# Gated Hugging Face model. Authenticate only on Forge; never put HF_TOKEN here.
MODEL_ID=drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-v1.1-alpha-Mida-Brikie
MODEL_REVISION=14c06b6db367b20a8cbe13f28b1f767b1e983307
MODEL_DIR=/home/REPLACE_USER/models/dsv4-flash-dspark-abliterated-v1.1
WORKER_MODEL_DIR=/home/REPLACE_USER/models/dsv4-flash-dspark-abliterated-v1.1
HF_CACHE=/home/REPLACE_USER/.cache/huggingface
HF_DOWNLOAD_WORKERS=4
HF_HUB_OFFLINE=1
TRANSFORMERS_OFFLINE=1
HF_HUB_DISABLE_XET=1
SERVED_MODEL_NAME=deepseek-v4-flash-dspark-v1.1-abliterated

# The API stays loopback-only unless you intentionally change this.
VLLM_HOST=127.0.0.1
VLLM_PORT=8888
NCCL_DEBUG=WARN