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

@@ -33,12 +33,11 @@ func TestLlamaCppBuildCommandArgs_BooleanFields(t *testing.T) {
 		{
 			name: "multiple booleans",
 			options: backends.LlamaServerOptions{
-				Verbose:   true,
-				FlashAttn: true,
-				Mlock:     false,
-				NoMmap:    true,
+				Verbose: true,
+				Mlock:   false,
+				NoMmap:  true,
 			},
-			expected: []string{"--verbose", "--flash-attn", "--no-mmap"},
+			expected: []string{"--verbose", "--no-mmap"},
 			excluded: []string{"--mlock"},
 		},
 	}