fix(vllm): update model path handling to use directory for safetensors

Author: ilopezluna

pkg/inference/backends/vllm/vllm_config.go

@@ -2,6 +2,7 @@ package vllm
 
 import (
 	"fmt"
+	"path/filepath"
 	"strconv"
 
 	"github.com/docker/model-runner/pkg/distribution/types"
@@ -27,7 +28,7 @@ func (c *Config) GetArgs(bundle types.ModelBundle, socket string, mode inference
 	args := append([]string{}, c.Args...)
 
 	// Add the serve command and model path (use directory for safetensors)
-	modelPath := bundle.SafetensorsPath()
+	modelPath := filepath.Dir(bundle.SafetensorsPath())
 	if modelPath != "" {
 		// vLLM expects the directory containing the safetensors files
 		args = append(args, "serve", modelPath)

pkg/inference/backends/vllm/vllm_config_test.go

@@ -51,7 +51,7 @@ func TestGetArgs(t *testing.T) {
 			config: nil,
 			expected: []string{
 				"serve",
-				"/path/to/model",
+				"/path/to",
 				"--uds",
 				"/tmp/socket",
 			},
@@ -66,7 +66,7 @@ func TestGetArgs(t *testing.T) {
 			},
 			expected: []string{
 				"serve",
-				"/path/to/model",
+				"/path/to",
 				"--uds",
 				"/tmp/socket",
 				"--max-model-len",
@@ -83,7 +83,7 @@ func TestGetArgs(t *testing.T) {
 			},
 			expected: []string{
 				"serve",
-				"/path/to/model",
+				"/path/to",
 				"--uds",
 				"/tmp/socket",
 				"--gpu-memory-utilization",
@@ -103,7 +103,7 @@ func TestGetArgs(t *testing.T) {
 			},
 			expected: []string{
 				"serve",
-				"/path/to/model",
+				"/path/to",
 				"--uds",
 				"/tmp/socket",
 				"--max-model-len",