fix: remove artificial model group concurrency cap

This commit is contained in:
wlx committed 2026-09-29 12:46:10 +08:00
1 parent 25114ca6b8
commit ce0a718686
13 files changed
+165 -34

No files matched your search

@@ -43,8 +43,15 @@ moyi:
redis-prefix: ${MOYI_REDIS_KEY_PREFIX:moyi:}
requests-per-minute: ${GATEWAY_REQUESTS_PER_MINUTE:60}
reads-per-minute: ${GATEWAY_READS_PER_MINUTE:300}
default-model-concurrency: ${GATEWAY_DEFAULT_MODEL_CONCURRENCY:4}
# 0 disables the MOYI-side model-group cap so channel concurrency can be
# measured directly. Set a positive value only when a provider confirms
# a shared per-model limit.
default-model-concurrency: ${GATEWAY_DEFAULT_MODEL_CONCURRENCY:0}
default-key-concurrency: ${GATEWAY_DEFAULT_KEY_CONCURRENCY:4}
# Temporary measurement switch. When true, skip MOYI model/group slots
# while retaining the independent API-Key limit, RPM, billing and the
# provider's own response/limit.
disable-model-concurrency: ${GATEWAY_DISABLE_MODEL_CONCURRENCY:false}
queue-wait-ms: ${GATEWAY_QUEUE_WAIT_MS:30000}
queue-poll-ms: ${GATEWAY_QUEUE_POLL_MS:100}
tasks:
@@ -0,0 +1,14 @@
-- The original relay catalog seeds used max_concurrency = 1 as a temporary
-- safety value. That value is a MOYI-side admission limit; it is not a
-- channel/provider quota and made every group look serial even when the
-- upstream account could process more requests.
--
-- 0 means "no MOYI model-group cap". The gateway still enforces each
-- platform API-Key limit, request-rate limit, wallet/usage accounting, and
-- the provider's own quota/rate limit. Administrators can still set an
-- explicit positive cap for a route when a channel reports one.
UPDATE model_catalog
SET max_concurrency = 0,
updated_at = now()
WHERE deleted_at IS NULL
AND max_concurrency = 1;