No need to use different implementations for pull models between win/unix for now

pkg/inference/models/manager.go

@@ -1,6 +1,7 @@
 package models
 
 import (
+	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -235,3 +236,25 @@ func (m *Manager) GetModelPath(ref string) (string, error) {
 	}
 	return fmt.Sprintf("%s/blobs/%s/%s", m.distributionClient.GetStorePath(), parts[0], parts[1]), nil
 }
+
+// PullModel pulls a model to local storage. Any error it returns is suitable
+// for writing back to the client.
+func (m *Manager) PullModel(ctx context.Context, model string) error {
+	// Restrict model pull concurrency.
+	select {
+	case <-m.pullTokens:
+	case <-ctx.Done():
+		return context.Canceled
+	}
+	defer func() {
+		m.pullTokens <- struct{}{}
+	}()
+
+	// Pull the model using the Docker model distribution client
+	m.log.Infoln("Pulling model:", model)
+	if _, err := m.distributionClient.PullModel(ctx, model); err != nil {
+		return fmt.Errorf("error while pulling model: %w", err)
+	}
+
+	return nil
+}

pkg/inference/models/manager_unix.go

@@ -1,30 +0,0 @@
-//go:build !windows
-
-package models
-
-import (
-	"context"
-	"fmt"
-)
-
-// PullModel pulls a model to local storage. Any error it returns is suitable
-// for writing back to the client.
-func (m *Manager) PullModel(ctx context.Context, model string) error {
-	// Restrict model pull concurrency.
-	select {
-	case <-m.pullTokens:
-	case <-ctx.Done():
-		return context.Canceled
-	}
-	defer func() {
-		m.pullTokens <- struct{}{}
-	}()
-
-	// Pull the model using the Docker model distribution client
-	m.log.Infoln("Pulling model:", model)
-	if _, err := m.distributionClient.PullModel(ctx, model); err != nil {
-		return fmt.Errorf("error while pulling model: %w", err)
-	}
-
-	return nil
-}

pkg/inference/models/manager_windows.go

@@ -1,16 +0,0 @@
-package models
-
-import (
-	"context"
-	"errors"
-)
-
-// PullModel is unimplemented on Windows.
-func (m *Manager) PullModel(ctx context.Context, model string) error {
-	// TODO: The Hugging Face Hub package that we're using doesn't build on
-	// Windows due to non-portable syscall package usage. With
-	// docker/model-distribution just a few days away, it's not worth patching
-	// or reimplementing. Once we switch to docker/model-distribution, delete
-	// the _unix.go / _windows.go variants and move PullModel into manager.go.
-	return errors.New("model pulls not yet supported on Windows")
-}