Spaces:
Running
Running
File size: 3,020 Bytes
125f9c0 | 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 | # DAYDREAM — deploy + COST CONTROL surface.
#
# The golden rule after the $200 incident: GPU apps are ALWAYS deployed
# scale-to-zero (min_containers=0, max_containers=1). They cost nothing idle and
# cannot be left running. "Warm for a demo" is a time-boxed LEASE driven by the
# cloud guardian — forget to tear down and it expires on its own, laptop or not.
#
# make status what's running + lease state (run this whenever unsure)
# make stop PANIC BUTTON — stop every app now
# make deploy (re)deploy all endpoints scale-to-zero (safe, $0/hr idle)
# make demo-up lease a 30-min warm window for a live demo
# make demo-down end the lease now (GPUs scale to zero within ~5 min)
#
# Cost lives in the ACTIVE Modal workspace — check it with `make whoami`.
# Activate venv AND load .env (so MODAL_WORKSPACE is exported for the guardian
# deploy, which bakes it into its image to derive endpoint URLs).
PY := . .venv/bin/activate && set -a && [ -f .env ] && . ./.env ; set +a ;
APPS := small-hack-vllm small-hack-llamacpp small-hack-flux
LEASE_MIN ?= 30
LEASE_APPS ?= vllm,llamacpp,flux
.PHONY: whoami status stop deploy deploy-guardian demo-up demo-down
whoami:
@$(PY) echo "Modal workspace: $$(modal profile current)"
status:
@$(PY) echo "=== Modal workspace: $$(modal profile current) ===" ; \
modal app list 2>/dev/null | grep -iE "small-hack|App ID" || true ; \
echo "=== spend ===" ; \
python scripts/spend.py 2>/dev/null || echo "(spend unavailable)" ; \
echo "=== warm lease ===" ; \
modal run modal_app/guardian.py::lease_status 2>/dev/null || echo "(guardian not deployed)"
# PANIC BUTTON: clear the lease AND stop every app. Use any time spend looks wrong.
stop:
@$(PY) modal run modal_app/guardian.py::lease_clear 2>/dev/null || true ; \
for a in $(APPS) ; do echo "stopping $$a" ; modal app stop $$a -y 2>/dev/null || true ; done ; \
echo "ALL STOPPED — \$$0/hr."
# Deploy everything scale-to-zero + the always-on guardian. Safe to leave
# deployed: GPU idle cost is $0, the guardian is a cheap CPU cron.
deploy: deploy-guardian
@$(PY) modal deploy modal_app/vllm_server.py ; \
modal deploy modal_app/llamacpp_server.py ; \
modal deploy modal_app/flux_server.py
# The guardian bakes $MODAL_WORKSPACE into its image to derive endpoint URLs.
deploy-guardian:
@$(PY) test -n "$$MODAL_WORKSPACE" || { echo "ERROR: set MODAL_WORKSPACE in .env first"; exit 1; } ; \
echo "deploying guardian for workspace: $$MODAL_WORKSPACE" ; \
modal deploy modal_app/guardian.py
# Lease a warm window for a live demo. Auto-expires — this is the safety net.
demo-up:
@$(PY) modal run modal_app/guardian.py::lease_set --minutes $(LEASE_MIN) --apps $(LEASE_APPS)
demo-down:
@$(PY) modal run modal_app/guardian.py::lease_clear
# Watch the guardian's scheduled keepalive runs (confirm the cron is firing and
# whether it's warming or idle). Ctrl-C to stop; dashboard also shows run history.
logs:
@$(PY) modal app logs small-hack-guardian
|