server/internal/client/ollama: hold DiskCache on Registry (#9463)

Previously, using a Registry required a DiskCache to be passed in for
use in various methods. This was a bit cumbersome, as the DiskCache is
required for most operations, and the DefaultCache is used in most of
those cases. This change makes the DiskCache an optional field on the
Registry struct.

This also changes DefaultCache to initialize on first use. This is to
not burden clients with the cost of creating a new cache per use, or
having to hold onto a cache for the lifetime of the Registry.

Also, slip in some minor docs updates for Trace.
This commit is contained in:
Blake Mizerany
2025-03-02 20:55:44 -08:00
committed by GitHub
parent e41c4cbea7
commit 3519dd1c6e
8 changed files with 132 additions and 105 deletions

View File

@@ -27,6 +27,7 @@ import (
"slices"
"strconv"
"strings"
"sync"
"sync/atomic"
"time"
@@ -73,19 +74,22 @@ const (
DefaultMaxChunkSize = 8 << 20
)
// DefaultCache returns a new disk cache for storing models. If the
// OLLAMA_MODELS environment variable is set, it uses that directory;
// otherwise, it uses $HOME/.ollama/models.
func DefaultCache() (*blob.DiskCache, error) {
var defaultCache = sync.OnceValues(func() (*blob.DiskCache, error) {
dir := os.Getenv("OLLAMA_MODELS")
if dir == "" {
home, err := os.UserHomeDir()
if err != nil {
return nil, err
}
home, _ := os.UserHomeDir()
home = cmp.Or(home, ".")
dir = filepath.Join(home, ".ollama", "models")
}
return blob.Open(dir)
})
// DefaultCache returns the default cache used by the registry. It is
// configured from the OLLAMA_MODELS environment variable, or defaults to
// $HOME/.ollama/models, or, if an error occurs obtaining the home directory,
// it uses the current working directory.
func DefaultCache() (*blob.DiskCache, error) {
return defaultCache()
}
// Error is the standard error returned by Ollama APIs. It can represent a
@@ -168,6 +172,10 @@ func CompleteName(name string) string {
// Registry is a client for performing push and pull operations against an
// Ollama registry.
type Registry struct {
// Cache is the cache used to store models. If nil, [DefaultCache] is
// used.
Cache *blob.DiskCache
// UserAgent is the User-Agent header to send with requests to the
// registry. If empty, the User-Agent is determined by HTTPClient.
UserAgent string
@@ -206,12 +214,18 @@ type Registry struct {
// It is only used when a layer is larger than [MaxChunkingThreshold].
MaxChunkSize int64
// Mask, if set, is the name used to convert non-fully qualified
// names to fully qualified names. If empty, the default mask
// ("registry.ollama.ai/library/_:latest") is used.
// Mask, if set, is the name used to convert non-fully qualified names
// to fully qualified names. If empty, [DefaultMask] is used.
Mask string
}
func (r *Registry) cache() (*blob.DiskCache, error) {
if r.Cache != nil {
return r.Cache, nil
}
return defaultCache()
}
func (r *Registry) parseName(name string) (names.Name, error) {
mask := defaultMask
if r.Mask != "" {
@@ -282,12 +296,17 @@ type PushParams struct {
}
// Push pushes the model with the name in the cache to the remote registry.
func (r *Registry) Push(ctx context.Context, c *blob.DiskCache, name string, p *PushParams) error {
func (r *Registry) Push(ctx context.Context, name string, p *PushParams) error {
if p == nil {
p = &PushParams{}
}
m, err := r.ResolveLocal(c, cmp.Or(p.From, name))
c, err := r.cache()
if err != nil {
return err
}
m, err := r.ResolveLocal(cmp.Or(p.From, name))
if err != nil {
return err
}
@@ -403,7 +422,7 @@ func canRetry(err error) bool {
// chunks of the specified size, and then reassembled and verified. This is
// typically slower than splitting the model up across layers, and is mostly
// utilized for layers of type equal to "application/vnd.ollama.image".
func (r *Registry) Pull(ctx context.Context, c *blob.DiskCache, name string) error {
func (r *Registry) Pull(ctx context.Context, name string) error {
scheme, n, _, err := r.parseNameExtended(name)
if err != nil {
return err
@@ -417,6 +436,11 @@ func (r *Registry) Pull(ctx context.Context, c *blob.DiskCache, name string) err
return fmt.Errorf("%w: no layers", ErrManifestInvalid)
}
c, err := r.cache()
if err != nil {
return err
}
exists := func(l *Layer) bool {
info, err := c.Get(l.Digest)
return err == nil && info.Size == l.Size
@@ -554,11 +578,15 @@ func (r *Registry) Pull(ctx context.Context, c *blob.DiskCache, name string) err
// Unlink is like [blob.DiskCache.Unlink], but makes name fully qualified
// before attempting to unlink the model.
func (r *Registry) Unlink(c *blob.DiskCache, name string) (ok bool, _ error) {
func (r *Registry) Unlink(name string) (ok bool, _ error) {
n, err := r.parseName(name)
if err != nil {
return false, err
}
c, err := r.cache()
if err != nil {
return false, err
}
return c.Unlink(n.String())
}
@@ -631,12 +659,17 @@ type Layer struct {
}
// ResolveLocal resolves a name to a Manifest in the local cache.
func (r *Registry) ResolveLocal(c *blob.DiskCache, name string) (*Manifest, error) {
func (r *Registry) ResolveLocal(name string) (*Manifest, error) {
_, n, d, err := r.parseNameExtended(name)
if err != nil {
return nil, err
}
c, err := r.cache()
if err != nil {
return nil, err
}
if !d.IsValid() {
// No digest, so resolve the manifest by name.
d, err = c.Resolve(n.String())
if err != nil {
return nil, err