Fix vllm command parsing

pkg/backends/parser.go

@@ -17,8 +17,8 @@ func ParseCommand(command string, executableNames []string, subcommandNames []st
 		return fmt.Errorf("command cannot be empty")
 	}
 
-	// Extract arguments
+	// Extract arguments and positional model
-	args, err := extractArgs(command, executableNames, subcommandNames)
+	args, modelFromPositional, err := extractArgs(command, executableNames, subcommandNames)
 	if err != nil {
 		return err
 	}
@@ -29,6 +29,13 @@ func ParseCommand(command string, executableNames []string, subcommandNames []st
 		return err
 	}
 
+	// If we found a positional model and no --model flag was provided, set the model
+	if modelFromPositional != "" {
+		if _, hasModelFlag := options["model"]; !hasModelFlag {
+			options["model"] = modelFromPositional
+		}
+	}
+
 	// Convert to target struct via JSON
 	jsonData, err := json.Marshal(options)
 	if err != nil {
@@ -51,15 +58,16 @@ func normalizeCommand(command string) string {
 }
 
 // extractArgs extracts arguments from command, removing executable and subcommands
-func extractArgs(command string, executableNames []string, subcommandNames []string) ([]string, error) {
+// Returns: args, modelFromPositional, error
+func extractArgs(command string, executableNames []string, subcommandNames []string) ([]string, string, error) {
 	// Check for unterminated quotes
 	if strings.Count(command, `"`)%2 != 0 || strings.Count(command, `'`)%2 != 0 {
-		return nil, fmt.Errorf("unterminated quoted string")
+		return nil, "", fmt.Errorf("unterminated quoted string")
 	}
 
 	tokens := strings.Fields(command)
 	if len(tokens) == 0 {
-		return nil, fmt.Errorf("no tokens found")
+		return nil, "", fmt.Errorf("no tokens found")
 	}
 
 	// Skip executable
@@ -104,7 +112,16 @@ func extractArgs(command string, executableNames []string, subcommandNames []str
 		}
 	}
 
-	return tokens[start:], nil
+	args := tokens[start:]
+
+	// Extract first positional argument (model) if present and not a flag
+	var modelFromPositional string
+	if len(args) > 0 && !strings.HasPrefix(args[0], "-") {
+		modelFromPositional = args[0]
+		args = args[1:] // Remove the model from args to process remaining flags
+	}
+
+	return args, modelFromPositional, nil
 }
 
 // parseFlags parses command line flags into a map

pkg/backends/vllm/vllm_test.go

@@ -14,17 +14,22 @@ func TestParseVllmCommand(t *testing.T) {
 	}{
 		{
 			name:      "basic vllm serve command",
-			command:   "vllm serve --model microsoft/DialoGPT-medium",
+			command:   "vllm serve microsoft/DialoGPT-medium",
 			expectErr: false,
 		},
 		{
 			name:      "serve only command",
-			command:   "serve --model microsoft/DialoGPT-medium",
+			command:   "serve microsoft/DialoGPT-medium",
 			expectErr: false,
 		},
 		{
-			name:      "args only",
+			name:      "positional model with flags",
-			command:   "--model microsoft/DialoGPT-medium --tensor-parallel-size 2",
+			command:   "vllm serve microsoft/DialoGPT-medium --tensor-parallel-size 2",
+			expectErr: false,
+		},
+		{
+			name:      "model with path",
+			command:   "vllm serve /path/to/model --gpu-memory-utilization 0.8",
 			expectErr: false,
 		},
 		{
@@ -34,7 +39,7 @@ func TestParseVllmCommand(t *testing.T) {
 		},
 		{
 			name:      "unterminated quote",
-			command:   `vllm serve --model "unterminated`,
+			command:   `vllm serve "unterminated`,
 			expectErr: true,
 		},
 	}
@@ -63,7 +68,7 @@ func TestParseVllmCommand(t *testing.T) {
 }
 
 func TestParseVllmCommandValues(t *testing.T) {
-	command := "vllm serve --model test-model --tensor-parallel-size 4 --gpu-memory-utilization 0.8 --enable-log-outputs"
+	command := "vllm serve test-model --tensor-parallel-size 4 --gpu-memory-utilization 0.8 --enable-log-outputs"
 	result, err := vllm.ParseVllmCommand(command)
 
 	if err != nil {