flash attn: add auto mode for llama engine (#13052)

* flash attn: add auto mode for llama engine

If the user does not specify fa in the environment, use auto-mode.

* review comments

* ensure kv cache quantized types have FA explicitly enabled

additional review comments

ml/device.go

@@ -492,6 +492,32 @@ func FlashAttentionSupported(l []DeviceInfo) bool {
 	return true
 }
 
+type FlashAttentionType int32
+
+const (
+	// Aligned with llama_flash_attn_type
+	FlashAttentionAuto     FlashAttentionType = -1
+	FlashAttentionDisabled FlashAttentionType = 0
+	FlashAttentionEnabled  FlashAttentionType = 1
+)
+
+func (f FlashAttentionType) LogValue() slog.Value {
+	return slog.AnyValue(f.String())
+}
+
+func (f FlashAttentionType) String() string {
+	switch f {
+	case FlashAttentionAuto:
+		return "Auto"
+	case FlashAttentionDisabled:
+		return "Disabled"
+	case FlashAttentionEnabled:
+		return "Enabled"
+	default:
+		return "unknown"
+	}
+}
+
 // Given the list of GPUs this instantiation is targeted for,
 // figure out the visible devices environment variables
 // Set mustFilter true to enable filtering of CUDA devices