diff --git a/README.md b/README.md
index c2f3342f0ff42..aa41b9e760c9b 100644
--- a/README.md
+++ b/README.md
@@ -17,6 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
+- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
 - Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
diff --git a/ggml-metal.m b/ggml-metal.m
index 54d1f1d5a3cee..416b24532f0d0 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -1392,6 +1392,14 @@ static enum ggml_status ggml_metal_graph_compute(
                             (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
                             //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
+                            // some Metal matrix data types require aligned pointers
+                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
+                                case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8  == 0); break;
+                                default: break;
+                            }
+
                             id<MTLComputePipelineState> pipeline = nil;
 
                             switch (src0->type) {
@@ -1706,6 +1714,14 @@ static enum ggml_status ggml_metal_graph_compute(
                             ne20 % 32 == 0 && ne20 >= 64 &&
                             ne11 > ne11_mm_min) {
 
+                            // some Metal matrix data types require aligned pointers
+                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
+                                case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8  == 0); break;
+                                default: break;
+                            }
+
                             id<MTLComputePipelineState> pipeline = nil;
 
                             switch (src2->type) {
diff --git a/ggml-metal.metal b/ggml-metal.metal
index 5552c8b28834a..748f0acef27b1 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -5760,18 +5760,18 @@ typedef void (mat_mm_t)(
         threadgroup   uchar *,
         uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mat_mm_t kernel_mul_mm<float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mat_mm_t kernel_mul_mm<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q6_K,    QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
@@ -5820,18 +5820,18 @@ typedef void (mat_mm_id_t)(
         threadgroup    uchar *,
         uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_id_f32_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_id_f16_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K,    QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;