feat: skip re-downloading models on shared storage

When multiple nodes mount the same filesystem (e.g., GPFS/NFS at
/storage/models), the model-agent on each node would independently
re-download from HuggingFace or OCI, causing rate-limiting and hours
of unnecessary I/O.

Add isModelAlreadyDownloaded() that checks:
1. config.json exists
2. If model.safetensors.index.json exists, ALL expected shards present
3. Otherwise, at least one weight file (.safetensors/.bin/.pt/.gguf)

Only applies to fresh Download tasks (not DownloadOverride) so spec
updates and failed retries still re-evaluate.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Kangyan-Zhou

pkg/modelagent/gopher.go

@@ -2,6 +2,7 @@ package modelagent
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -324,6 +325,25 @@ func (s *Gopher) processTask(task *GopherTask) error {
 				s.logger.Errorf("Failed to get target directory path for model %s: %v", modelInfo, err)
 				return err
 			}
+
+			// Check if the model is already present on shared storage.
+			// Only for fresh Download tasks — DownloadOverride indicates a spec change
+			// or failed retry that must re-evaluate the model files.
+			if task.TaskType == Download && s.isModelAlreadyDownloaded(destPath) {
+				s.logger.Infof("Model %s already exists at %s (shared storage), skipping OCI download", modelInfo, destPath)
+				var baseModel *v1beta1.BaseModel
+				var clusterBaseModel *v1beta1.ClusterBaseModel
+				if task.BaseModel != nil {
+					baseModel = task.BaseModel
+				} else if task.ClusterBaseModel != nil {
+					clusterBaseModel = task.ClusterBaseModel
+				}
+				if err := s.safeParseAndUpdateModelConfig(destPath, baseModel, clusterBaseModel, nil); err != nil {
+					s.logger.Errorf("Failed to parse and update model config for pre-existing model: %v", err)
+				}
+				break
+			}
+
 			err = utils.Retry(s.downloadRetry, 100*time.Millisecond, func() error {
 				downloadErr := s.downloadModel(ctx, osUri, destPath, task)
 				if downloadErr != nil {
@@ -971,6 +991,29 @@ func (s *Gopher) processHuggingFaceModel(ctx context.Context, task *GopherTask,
 	// Create destination path
 	destPath := getDestPath(&baseModelSpec, s.modelRootDir)
 
+	// Check if the model is already present on shared storage (e.g., another node
+	// already downloaded it to the same NFS/shared filesystem path). When storage is
+	// shared across nodes, each model-agent would otherwise independently re-download
+	// from HuggingFace, causing rate-limiting and hours of unnecessary I/O.
+	// Only for fresh Download tasks — DownloadOverride indicates a spec change
+	// or failed retry that must re-evaluate the model files.
+	if task.TaskType == Download && s.isModelAlreadyDownloaded(destPath) {
+		s.logger.Infof("Model %s already exists at %s (shared storage), skipping HuggingFace download", modelInfo, destPath)
+
+		var baseModel *v1beta1.BaseModel
+		var clusterBaseModel *v1beta1.ClusterBaseModel
+		if task.BaseModel != nil {
+			baseModel = task.BaseModel
+		} else if task.ClusterBaseModel != nil {
+			clusterBaseModel = task.ClusterBaseModel
+		}
+
+		if err := s.safeParseAndUpdateModelConfig(destPath, baseModel, clusterBaseModel, nil); err != nil {
+			s.logger.Errorf("Failed to parse and update model config for pre-existing model: %v", err)
+		}
+		return nil
+	}
+
 	// fetch sha value based on model ID from Huggingface model API
 	shaStr, isShaAvailable := s.fetchSha(ctx, hfComponents.ModelID, name)
 	isReuseEligible, matchedModelTypeAndModeName, parentPath := s.isEligibleForOptimization(ctx, task, baseModelSpec, modelType, namespace, isShaAvailable, shaStr, name)
@@ -1440,3 +1483,96 @@ func (s *Gopher) isRemoveParentArtifactDirectory(ctx context.Context, hasChildre
 	s.logger.Infof("parent entry %s:%s exists on node configmap: %v", parentName, parentDir, exists)
 	return !exists
 }
+
+// isModelAlreadyDownloaded checks whether the model files are already present at
+// destPath. This handles the shared-storage case: when multiple nodes mount the
+// same filesystem (e.g., NFS at /storage/models), the first node that finishes an
+// HF download writes the files once. Subsequent nodes should detect the existing
+// files and skip re-downloading.
+//
+// The check is deliberately conservative — it requires:
+//  1. The directory exists
+//  2. A config.json file exists
+//  3. At least one model weight file exists (.safetensors, .bin, .pt, .gguf)
+//  4. If model.safetensors.index.json exists, ALL expected shards must be present
+//
+// This avoids false positives from partially-downloaded directories.
+func (s *Gopher) isModelAlreadyDownloaded(destPath string) bool {
+	// Check if directory exists
+	info, err := os.Stat(destPath)
+	if err != nil || !info.IsDir() {
+		return false
+	}
+
+	// Check for config.json (primary indicator of a complete HF download).
+	// Use err != nil (not os.IsNotExist) so that NFS I/O errors and permission
+	// errors are treated conservatively as "not present" rather than silently
+	// falling through as "exists".
+	configPath := filepath.Join(destPath, "config.json")
+	if _, err := os.Stat(configPath); err != nil {
+		return false
+	}
+
+	// Read directory entries once for all subsequent checks
+	entries, err := os.ReadDir(destPath)
+	if err != nil {
+		return false
+	}
+
+	// Build a set of filenames for fast lookup
+	fileSet := make(map[string]bool, len(entries))
+	for _, entry := range entries {
+		if !entry.IsDir() {
+			fileSet[entry.Name()] = true
+		}
+	}
+
+	// If model.safetensors.index.json exists, verify ALL expected shards are present.
+	// This is the strongest completeness check — it catches partial downloads.
+	indexPath := filepath.Join(destPath, "model.safetensors.index.json")
+	if _, err := os.Stat(indexPath); err == nil {
+		indexData, err := os.ReadFile(indexPath)
+		if err != nil {
+			s.logger.Warnf("Failed to read model index file %s: %v", indexPath, err)
+			return false
+		}
+
+		var index struct {
+			WeightMap map[string]string `json:"weight_map"`
+		}
+		if err := json.Unmarshal(indexData, &index); err != nil {
+			s.logger.Warnf("Failed to parse model index file %s: %v", indexPath, err)
+			return false
+		}
+
+		// Collect unique shard filenames from weight_map values
+		expectedShards := make(map[string]bool)
+		for _, shard := range index.WeightMap {
+			expectedShards[shard] = true
+		}
+
+		// Verify every expected shard exists on disk
+		for shard := range expectedShards {
+			if !fileSet[shard] {
+				s.logger.Infof("Model at %s is missing shard %s (expected %d shards), not treating as complete",
+					destPath, shard, len(expectedShards))
+				return false
+			}
+		}
+
+		s.logger.Infof("Model at %s has all %d expected shards from index", destPath, len(expectedShards))
+		return true
+	}
+
+	// No index file — fall back to checking for at least one weight file
+	weightExtensions := []string{".safetensors", ".bin", ".pt", ".gguf"}
+	for name := range fileSet {
+		for _, ext := range weightExtensions {
+			if strings.HasSuffix(name, ext) {
+				return true
+			}
+		}
+	}
+
+	return false
+}

pkg/modelagent/gopher_test.go

@@ -5,6 +5,8 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"os"
+	"path/filepath"
 	"testing"
 
 	"k8s.io/apimachinery/pkg/runtime/schema"
@@ -961,6 +963,100 @@ func TestIsEligibleForOptimization_NoMatch(t *testing.T) {
 	assert.Empty(t, parent)
 }
 
+func TestIsModelAlreadyDownloaded(t *testing.T) {
+	logger, _ := zap.NewDevelopment()
+	sugaredLogger := logger.Sugar()
+	defer func() { _ = sugaredLogger.Sync() }()
+
+	gopher := &Gopher{logger: sugaredLogger}
+
+	t.Run("nonexistent directory returns false", func(t *testing.T) {
+		assert.False(t, gopher.isModelAlreadyDownloaded("/nonexistent/path/that/does/not/exist"))
+	})
+
+	t.Run("empty directory returns false", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.False(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with only config.json returns false (no weights)", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.False(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with only weights returns false (no config.json)", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00001-of-00002.safetensors"), []byte("weight data"), 0644))
+		assert.False(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with config.json and safetensors returns true", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00001-of-00002.safetensors"), []byte("weight data"), 0644))
+		assert.True(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with config.json and .bin returns true", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "pytorch_model.bin"), []byte("weight data"), 0644))
+		assert.True(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with config.json and .pt returns true", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model.pt"), []byte("weight data"), 0644))
+		assert.True(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("directory with config.json and .gguf returns true", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model.gguf"), []byte("weight data"), 0644))
+		assert.True(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("file path instead of directory returns false", func(t *testing.T) {
+		dir := t.TempDir()
+		filePath := filepath.Join(dir, "somefile")
+		assert.NoError(t, os.WriteFile(filePath, []byte("data"), 0644))
+		assert.False(t, gopher.isModelAlreadyDownloaded(filePath))
+	})
+
+	// Shard completeness tests using model.safetensors.index.json
+
+	t.Run("index with all shards present returns true", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		index := `{"metadata":{"total_size":100},"weight_map":{"w1":"model-00001-of-00002.safetensors","w2":"model-00002-of-00002.safetensors"}}`
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model.safetensors.index.json"), []byte(index), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00001-of-00002.safetensors"), []byte("data"), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00002-of-00002.safetensors"), []byte("data"), 0644))
+		assert.True(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("index with missing shard returns false", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		index := `{"metadata":{"total_size":100},"weight_map":{"w1":"model-00001-of-00002.safetensors","w2":"model-00002-of-00002.safetensors"}}`
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model.safetensors.index.json"), []byte(index), 0644))
+		// Only write shard 1, shard 2 is missing
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00001-of-00002.safetensors"), []byte("data"), 0644))
+		assert.False(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+
+	t.Run("malformed index file returns false", func(t *testing.T) {
+		dir := t.TempDir()
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "config.json"), []byte(`{"model_type":"llama"}`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model.safetensors.index.json"), []byte(`{invalid json`), 0644))
+		assert.NoError(t, os.WriteFile(filepath.Join(dir, "model-00001-of-00001.safetensors"), []byte("data"), 0644))
+		assert.False(t, gopher.isModelAlreadyDownloaded(dir))
+	})
+}
+
 func TestIsEligibleForOptimization_AlwaysDownloadNotEligible(t *testing.T) {
 	nodeName := "node-1"
 	sha := "123abc"