# Copy to .env (git-ignored) and edit. Every script runs on the head node (rank 0) from this checkout.
# Values below are examples from a two-node setup with one ConnectX-7 cable between the nodes.

# --- the two nodes -------------------------------------------------------------------------------------------
# Worker (rank 1) as ssh reaches it over the ConnectX-7 link. Passwordless ssh from the head is required.
WORKER=user@10.0.0.1
# The head's own IPv4 on that link: rank 0's rendezvous address.
HEAD_IP=10.0.0.2
# RoCE devices of the link (one cable usually exposes two: list both), the link's network interface, and the RoCE v2
# GID index whose GID carries the link's IPv4. scripts/preflight.sh prints all three for each node.
HCA=rocep1s0f1,roceP2p1s0f1
IFACE=enp1s0f1np1
GID=3

# --- weights and caches (the same absolute paths on both nodes) -----------------------------------------------
MODEL_DIR=$HOME/Models/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw
ENGRAM_DIR=$HOME/Models/DeepSeek-V4.1-Flash-1017-engram
# Each rank's weight cache (~99 GB a node, written on the first start), Engram token map, built CUDA extensions.
CACHE=$HOME/.cache/ds41-tf
# Hugging Face CLI on the head (and on the worker, for download-weights.sh --engram-on-worker).
HF=hf

# --- serving ---------------------------------------------------------------------------------------------------
IMAGE=tensorfold-dsv41:8e0966e
PORT=8888
SERVED=deepseek-v4.1-flash
CONTEXT=1048576
PARALLEL=4
MTP_DRAFTS=5
# Extra docker -e flags for both ranks, e.g. EXTRA_ENV="-e TF_LOOP_GUARD=1"; extra `tensorfold serve` flags in EXTRA_ARGS.
EXTRA_ENV=
EXTRA_ARGS=