Change FlashAttn field type to string in LlamaServerOptions

pkg/backends/llama.go

@@ -41,7 +41,7 @@ type LlamaServerOptions struct {
 	BatchSize               int      `json:"batch_size,omitempty"`
 	UBatchSize              int      `json:"ubatch_size,omitempty"`
 	Keep                    int      `json:"keep,omitempty"`
-	FlashAttn               bool     `json:"flash_attn,omitempty"`
+	FlashAttn               string   `json:"flash_attn,omitempty"`
 	NoPerf                  bool     `json:"no_perf,omitempty"`
 	Escape                  bool     `json:"escape,omitempty"`
 	NoEscape                bool     `json:"no_escape,omitempty"`

pkg/backends/llama_test.go

@@ -34,11 +34,10 @@ func TestLlamaCppBuildCommandArgs_BooleanFields(t *testing.T) {
 			name: "multiple booleans",
 			options: backends.LlamaServerOptions{
 				Verbose: true,
-				FlashAttn: true,
 				Mlock:   false,
 				NoMmap:  true,
 			},
-			expected: []string{"--verbose", "--flash-attn", "--no-mmap"},
+			expected: []string{"--verbose", "--no-mmap"},
 			excluded: []string{"--mlock"},
 		},
 	}

pkg/validation/validation_test.go

@@ -239,25 +239,3 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 		})
 	}
 }
-
-func TestValidateInstanceOptions_NonStringFields(t *testing.T) {
-	// Test that non-string fields don't interfere with validation
-	options := backends.Options{
-		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &backends.LlamaServerOptions{
-			Port:        8080,
-			GPULayers:   32,
-			CtxSize:     4096,
-			Temperature: 0.7,
-			TopK:        40,
-			TopP:        0.9,
-			Verbose:     true,
-			FlashAttn:   false,
-		},
-	}
-
-	err := options.ValidateInstanceOptions()
-	if err != nil {
-		t.Errorf("ValidateInstanceOptions with non-string fields should not error, got: %v", err)
-	}
-}