GPT-Load 2.x is a greenfield rewrite and is not data-compatible with 1.x.
Do not open a 1.x database with this release; rebuild configuration and cut over manually.
New 2.x deployments must start with an empty dedicated database and a matching new
encryption key. Existing supported 2.x installations must preserve their matching key
files and follow the version-specific release notes before upgrading.
This release provides OpenAI, Anthropic, and Gemini client protocol gateways backed
by channel presets for official, compatible, Azure OpenAI, AWS Bedrock, and Google
Vertex targets. It includes an embedded management UI, scheduling and health controls,
request logs, usage estimates, and model price management. It is single-instance and
supports SQLite, MySQL, and PostgreSQL through one unified DATABASE_DSN contract.
GPT-Load remains the routing owner: Groups select a code-defined channel and hold
encrypted credentials, AccessKeys authorize data-plane access, and GPT-Load performs
candidate selection, retry policy, health tracking, logging, and pricing. The embedded
provider SDK executes one selected logical attempt and performs supported protocol
conversion; its name is not part of persisted domain identities.
Docker runs as a fixed non-root user with a named volume. Verified Beta images move
the shared v2beta channel while retaining the exact version tag; stable releases
move only their major/minor aliases. No 2.x release moves latest. Stop cleanly,
then back up the selected database using its engine-native procedure and keep
auth/encryption key files with the recovery set.
Managed SQLite also requires its -wal/-shm sidecars. Usage and cost are best-effort
estimates, not a billing ledger or provider invoice.
Follow the public operations baseline
for deployment, backup/restore, and 1.x cutover and rollback.
This release contains five raw binaries, SHA256SUMS, a CycloneDX SBOM,
and the project and third-party license notices.
If publication stops between channels, inspect the reconciliation inventory and
recover manually; the workflow does not claim cross-system atomic rollback.
Usage quality can be missing, partial, or unpriced; a compatible upstream may
never provide authoritative terminal usage. Cost remains an estimate, not a provider
bill. The current 2.x line does not support encryption key rotation.
The automatic model catalog combines an embedded official catalog with Models.dev;
embedded official records take precedence while other Models.dev records remain
available. Prices resolve by exact channel_id + model_id; explicit user overrides
remain available, and all costs remain estimates.
GPT-Load keeps a unified data/control dual-plane architecture: Groups select a
channel and credentials, AccessKeys authorize data-plane access, and model discovery
imports available models from the selected channel.
Full Changelog: v2.0.0-beta.6...v2.0.0-beta.7