diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
index c21fd9717..b202f19ba 100644
--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal.metal
@@ -3247,12 +3247,14 @@ template [[host_name("kernel_flash_attn_ext_f16_h112")]]  kernel flash_attn_ext_
 template [[host_name("kernel_flash_attn_ext_f16_h128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  128>;
 template [[host_name("kernel_flash_attn_ext_f16_h256")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256>;
 
+#if !defined(GGML_METAL_NO_BFLOAT)
 template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64>;
 template [[host_name("kernel_flash_attn_ext_bf16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80>;
 template [[host_name("kernel_flash_attn_ext_bf16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96>;
 template [[host_name("kernel_flash_attn_ext_bf16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112>;
 template [[host_name("kernel_flash_attn_ext_bf16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256>;
+#endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80>;
@@ -3630,7 +3632,9 @@ kernel void kernel_flash_attn_ext_vec(
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64>) flash_attn_ext_vec_t;
 
 template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,     1, dequantize_f16,  128>;
+#if !defined(GGML_METAL_NO_BFLOAT)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,   1, dequantize_bf16, 128>;
+#endif
 template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0,  2, dequantize_q4_0, 128>;
 template [[host_name("kernel_flash_attn_ext_vec_q4_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1,  2, dequantize_q4_1, 128>;
 template [[host_name("kernel_flash_attn_ext_vec_q5_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0,  2, dequantize_q5_0, 128>;
@@ -3638,7 +3642,9 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0,  2, dequantize_q8_0, 128>;
 
 template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,     1, dequantize_f16,  256>;
+#if !defined(GGML_METAL_NO_BFLOAT)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,   1, dequantize_bf16, 256>;
+#endif
 template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0,  2, dequantize_q4_0, 256>;
 template [[host_name("kernel_flash_attn_ext_vec_q4_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1,  2, dequantize_q4_1, 256>;
 template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0,  2, dequantize_q5_0, 256>;