env.eval.example

# ═══════════════════════════════════════════════════════════════════════════════
# Leoma EVAL SERVER — copy to .env on the GPU box (`leoma servers eval-server`).
# Downloads king + challenger weights from Hippius Hub, loads the pinned diffusers
# I2V pipeline, and scores generations against the real continuations.
# ═══════════════════════════════════════════════════════════════════════════════

EVAL_SERVER_HOST=127.0.0.1
EVAL_SERVER_PORT=9000
# Required if binding anywhere except loopback; recommended everywhere. The validator
# must use the same value. Sent as an Authorization: Bearer header.
LEOMA_EVAL_TOKEN=

# Hippius Hub (OCI model registry) auth — a token OR username/password.
HIPPIUS_HUB_TOKEN=
# HIPPIUS_HUB_USERNAME=
# HIPPIUS_HUB_PASSWORD=
LEOMA_MODEL_CACHE_DIR=/var/lib/leoma/models
LEOMA_MAX_CACHED_SNAPSHOTS=4
LEOMA_MIN_FREE_BYTES=161061273600

# Hippius S3 source-video corpus (the ground-truth continuations).
OBJECT_STORAGE_BACKEND=hippius
HIPPIUS_ENDPOINT=s3.hippius.com
HIPPIUS_REGION=decentralized
HIPPIUS_SOURCE_BUCKET=leoma-source
HIPPIUS_VIDEOS_READ_ACCESS_KEY=
HIPPIUS_VIDEOS_READ_SECRET_KEY=

# Concurrent king/challenger generation — a THROUGHPUT setting, not a consensus one
# (it is never sent by the validator and never hashed into consensus_digest). On a
# multi-GPU box, generating both duelists at the same time on separate devices
# roughly halves duel wall-clock. Falls back to today's sequential single-device
# behavior if fewer than 2 CUDA devices are visible. See leoma/eval/devices.py.
#
# ⚠️ If you enable this, re-run `leoma calibrate` with it enabled — it changes what
# noise delta_threshold has to absorb (see docs/TESTNET_RUNBOOK.md).
# LEOMA_CONCURRENT_GENERATION=1

# LEOMA_KING_DEVICE / LEOMA_CHALLENGER_DEVICE are independent of the concurrency flag
# above: they also matter with LEOMA_CONCURRENT_GENERATION unset, e.g. when running
# several eval-server PROCESSES on one multi-GPU box (one per EVAL_SERVER_URLS entry)
# and pinning each process to its own device pair so their VRAM doesn't all pile onto
# the same default GPU. Set both together, or neither — a single one is ignored.
# LEOMA_KING_DEVICE=cuda:0
# LEOMA_CHALLENGER_DEVICE=cuda:1

# Optional deterministic prompt for the I2V pipeline (same for king + challenger).
# LEOMA_DUEL_PROMPT=