swactor/examples/pipeline-parallel-inference/profiles/example.env

# pipeline-parallel-inference — run profile (.env style)
# ----------------------------------------------------------------------------
# This is the single source of truth for the knobs that used to be scattered
# across the binaries and one operator's shell. Copy it to set up a run:
#
#   cp profiles/example.env profiles/local.env   # untracked; put real secrets here
#   PP_PROFILE=profiles/local.env cargo run --bin pp-orchestrator -- --seed ...
#
# When PP_PROFILE is unset, profiles/local.env is loaded automatically if it
# exists. The run scripts pick this up too (the binaries load it at startup).
#
# PRECEDENCE (lowest -> highest):
#   compiled default  <  this profile  <  real process env  <  CLI flag
# i.e. anything already exported in your shell wins over the file, and a
# --flag wins over everything. The profile only fills in what is unset.
#
# Lines are KEY=VALUE; `export ` prefix and surrounding quotes are tolerated;
# `#` comments and blank lines are ignored.
# ----------------------------------------------------------------------------

# --- Networking / relay -----------------------------------------------------
# Custom iroh relay the whole cluster homes onto (a swactor-iroh-relay on a
# VPS). Leave unset to use the iroh default relays. Placeholder below — put
# YOUR real relay only in profiles/local.env (gitignored), not here.
#SWACTOR_IROH_RELAY_URL=https://relay.example.com

# --- Topology ---------------------------------------------------------------
# Number of pipeline stages (>= 2). pp-orchestrator also accepts --num-stages.
#NUM_STAGES=2

# --- Compute target (vast.ai / docker) --------------------------------------
# Container image to run on each node. The run scripts already read PP_IMAGE;
# pp-orchestrator's --image default now reads it too. Set to your registry tag.
#PP_IMAGE=swactor-pp-gpu:latest
# GPU class requested when leasing on vast.ai (pp-orchestrator --gpu overrides).
#PP_GPU=RTX 3060

# --- Workload (what each stage computes) ------------------------------------
# Model identifier handed to the worker.
#MODEL=
# Python worker script + interpreter (per-stage compute). WORKER_SCRIPT is the
# pp-worker default; pp-orchestrator --worker ships a script to remote nodes.
#WORKER_SCRIPT=./pp_tinygrad_worker.py
#WORKER_CMD=python3
#PYTHON=python3
# Stub mode: skip the real worker, echo activations (fast local smoke runs).
#PP_WORKER_STUB=1
# Inference request prompt + token budget (pp-orchestrator --prompt/--max-tokens).
#MAX_TOKENS=64

# --- Timeouts (seconds; sane defaults baked in — override only if needed) ---
#PP_CONVERGE_TIMEOUT_SECS=1200
#PP_WORKER_READY_TIMEOUT_SECS=1800
#PP_NEIGHBOR_RESOLVE_TIMEOUT_SECS=1800
#PP_PIPELINE_WIRED_TIMEOUT_SECS=1800
#PP_AWAIT_RESPONSE_TIMEOUT_SECS=600
#PP_RESOLVE_TIMEOUT_SECS=1200

# --- Dashboards (optional) ---------------------------------------------------
#PP_DASHBOARD=1
#PP_DASHBOARD_PORT=9090

# Secrets (orchestrator / stage identity). Keep real values in local.env only.
#PP_ORCH_SECRET=
#PP_STAGE_SECRET=