ab3b02a48e
Populate the repo with the Vault/OpenBao dynamic secrets engine that mints LiteLLM virtual keys scoped by model, spending limit, and lease TTL. - Secrets backend: config, roles, creds paths and a revocable litellm_key type - LiteLLM API client (generate/update/delete/info) with master-key auth - Unit tests (mock LiteLLM) and a docker-compose e2e against both Vault and OpenBao proving the same binary works on each - Makefile, woodpecker CI (build/test/pre-commit), pre-commit config
119 lines
2.7 KiB
Go
119 lines
2.7 KiB
Go
package litellm
|
|
|
|
import (
|
|
"context"
|
|
"strings"
|
|
"sync"
|
|
|
|
"github.com/hashicorp/vault/sdk/framework"
|
|
"github.com/hashicorp/vault/sdk/logical"
|
|
)
|
|
|
|
// litellmBackend is the Vault secrets backend that manages LiteLLM virtual keys.
|
|
type litellmBackend struct {
|
|
*framework.Backend
|
|
|
|
lock sync.RWMutex
|
|
client *litellmClient
|
|
}
|
|
|
|
// Factory returns a configured LiteLLM secrets backend.
|
|
func Factory(ctx context.Context, conf *logical.BackendConfig) (logical.Backend, error) {
|
|
b := backend()
|
|
if err := b.Setup(ctx, conf); err != nil {
|
|
return nil, err
|
|
}
|
|
return b, nil
|
|
}
|
|
|
|
func backend() *litellmBackend {
|
|
b := &litellmBackend{}
|
|
|
|
b.Backend = &framework.Backend{
|
|
Help: strings.TrimSpace(backendHelp),
|
|
BackendType: logical.TypeLogical,
|
|
PathsSpecial: &logical.Paths{
|
|
LocalStorage: []string{},
|
|
SealWrapStorage: []string{
|
|
configStoragePath,
|
|
},
|
|
},
|
|
Paths: framework.PathAppend(
|
|
[]*framework.Path{
|
|
pathConfig(b),
|
|
pathRole(b),
|
|
pathRolesList(b),
|
|
pathCredentials(b),
|
|
},
|
|
),
|
|
Secrets: []*framework.Secret{
|
|
b.litellmKey(),
|
|
},
|
|
Invalidate: b.invalidate,
|
|
WALRollback: nil,
|
|
}
|
|
|
|
return b
|
|
}
|
|
|
|
// reset drops the cached LiteLLM client so it is rebuilt from storage on the
|
|
// next request. Called when the config changes.
|
|
func (b *litellmBackend) reset() {
|
|
b.lock.Lock()
|
|
defer b.lock.Unlock()
|
|
b.client = nil
|
|
}
|
|
|
|
// invalidate clears the cached client when the config is written from another
|
|
// cluster node.
|
|
func (b *litellmBackend) invalidate(_ context.Context, key string) {
|
|
if key == configStoragePath {
|
|
b.reset()
|
|
}
|
|
}
|
|
|
|
// getClient returns a cached LiteLLM client, building one from stored config if
|
|
// necessary.
|
|
func (b *litellmBackend) getClient(ctx context.Context, s logical.Storage) (*litellmClient, error) {
|
|
b.lock.RLock()
|
|
if b.client != nil {
|
|
defer b.lock.RUnlock()
|
|
return b.client, nil
|
|
}
|
|
b.lock.RUnlock()
|
|
|
|
b.lock.Lock()
|
|
defer b.lock.Unlock()
|
|
|
|
// Re-check after acquiring the write lock.
|
|
if b.client != nil {
|
|
return b.client, nil
|
|
}
|
|
|
|
config, err := getConfig(ctx, s)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
if config == nil {
|
|
return nil, errBackendNotConfigured
|
|
}
|
|
|
|
client, err := newClient(config)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
b.client = client
|
|
return b.client, nil
|
|
}
|
|
|
|
const backendHelp = `
|
|
The LiteLLM secrets backend dynamically generates LiteLLM virtual API keys
|
|
scoped to specific models, with a spending limit and a lease-bound TTL.
|
|
|
|
After mounting this backend, configure it with the connection details for the
|
|
LiteLLM proxy using the "config" path, then define one or more roles that
|
|
constrain the generated keys. Reading from the "creds/<role>" path issues a new
|
|
virtual key whose lifetime is managed by Vault: revoking the lease revokes the
|
|
key in LiteLLM.
|
|
`
|