@novastera-oss/llamarn 0.2.5 → 0.2.7

package/cpp/llama.cpp/ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,98 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

package/cpp/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h

@@ -944,10 +944,8 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     for (int i = 0; i < offset; ++i) {              \
         x[i] = vec_add(x[i], x[offset + i]);        \
     }                                               \
-    res = vec_extract(x[0], 0) +                    \
-          vec_extract(x[0], 1) +                    \
-          vec_extract(x[0], 2) +                    \
-          vec_extract(x[0], 3);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4

package/cpp/llama.cpp/ggml/src/ggml-cpu/{ggml-cpu-traits.cpp → traits.cpp}

@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"

package/cpp/llama.cpp/ggml/src/ggml-cuda/common.cuh

@@ -207,9 +207,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -262,11 +262,11 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]
@@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
@@ -635,6 +632,7 @@ struct ggml_cuda_device_info {
         int     nsm;                // number of streaming multiprocessors
         size_t  smpb;               // max. shared memory per block
         size_t  smpbo;              // max. shared memory per block (with opt-in)
+        bool    integrated;         // Device is integrated as opposed to discrete
         bool    vmm;                // virtual memory support
         size_t  vmm_granularity;    // granularity of virtual memory
         size_t  total_vram;

package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
             KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
         }
@@ -1246,7 +1249,7 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
-#endif __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
     static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 

package/cpp/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -243,10 +243,10 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-
-        info.devices[id].nsm       = prop.multiProcessorCount;
-        info.devices[id].smpb      = prop.sharedMemPerBlock;
-        info.devices[id].warp_size = prop.warpSize;
+        info.devices[id].integrated = prop.integrated;
+        info.devices[id].nsm        = prop.multiProcessorCount;
+        info.devices[id].smpb       = prop.sharedMemPerBlock;
+        info.devices[id].warp_size  = prop.warpSize;
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -615,9 +615,8 @@ static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemsetAsync(ctx->dev_ptr, value, buffer->size, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
@@ -1065,6 +1064,10 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+}
+
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
@@ -1140,7 +1143,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
     const char * src_ptr = (const char *) src->data;
     char       * dst_ptr = (char       *) dst;
 
@@ -1423,8 +1425,6 @@ static void ggml_cuda_op_mul_mat(
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
     ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
     ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
 
@@ -1746,7 +1746,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
@@ -2641,6 +2641,8 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
+    // flag used to determine whether it is an integrated_gpu
+    const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@@ -2659,10 +2661,12 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     if (node->src[j] != nullptr) {
                         assert(node->src[j]->buffer);
                         assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) ||
-                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft));
+                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                     }
                 }
-#endif
+#else
+                GGML_UNUSED(integrated);
+#endif // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
@@ -2994,9 +2998,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
-                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
-                // this avoids some edge cases (and the performance would not be good anyways)
                 if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    if (a->ne[2] > 1 || a->ne[3] > 1) {
+                        return false;
+                    }
+                    // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                    // this avoids some edge cases (and the performance would not be good anyways)
                     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
                     int64_t row_low;
                     int64_t row_high;
@@ -3263,7 +3270,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 }
 
 static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return (ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+    const bool integrated = ggml_cuda_info().devices[dev_ctx->device].integrated;
+    return (((ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev) || (integrated && ggml_backend_buft_is_cuda_host(buft)));
 }
 
 static int64_t get_op_batch_size(const ggml_tensor * op) {

package/cpp/llama.cpp/ggml/src/ggml-cuda/ssm-scan.cu

@@ -10,6 +10,8 @@ __global__ void __launch_bounds__(splitD, 2)
                  float * __restrict__ dst, const int64_t L) {
     GGML_UNUSED(src1_nb0);
     GGML_UNUSED(src2_nb0);
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     const int bidx = blockIdx.x;  // split along B
     const int bidy = blockIdx.y;  // split along D
     const int tid  = threadIdx.x;
@@ -44,16 +46,16 @@ __global__ void __launch_bounds__(splitD, 2)
     if (N == 16) {
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
+            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
             // todo: bank conflict
             // I am always confused with how to use the swizzling method to solve
             // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
-            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
+            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 

package/cpp/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt

@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
+    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

package/cpp/llama.cpp/ggml/src/ggml-impl.h

@@ -32,6 +32,8 @@
 extern "C" {
 #endif
 
+void ggml_print_backtrace(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif

package/cpp/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt

@@ -44,21 +44,22 @@ if (GGML_METAL_EMBED_LIBRARY)
     set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
-        OUTPUT ${METALLIB_EMBED_ASM}
+        OUTPUT "${METALLIB_EMBED_ASM}"
         COMMAND echo "Embedding Metal library"
-        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
-        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
+        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
         DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
         COMMENT "Generate assembly for embedded Metal library"
+        VERBATIM
     )
 
-    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
+    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
 else()
     if (GGML_METAL_SHADER_DEBUG)
         # custom command to do the following:

package/cpp/llama.cpp/ggml/src/ggml-metal/ggml-metal.m

@@ -498,6 +498,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+    GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
     GGML_METAL_KERNEL_TYPE_ARGMAX,
@@ -1454,6 +1455,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
@@ -1653,6 +1655,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_LOG:
             return false; // TODO: implement
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
@@ -2400,11 +2403,30 @@ static bool ggml_metal_encode_node(
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             {
                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (dst->op) {
+                    case GGML_OP_SUM_ROWS:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                        break;
+                    case GGML_OP_MEAN:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
+                        break;
+                    default:
+                        GGML_ABORT("fatal error");
+                }
+
+                int nth = 32; // SIMD width
+
+                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
 
+                nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {
                    /*.ne00 =*/ ne00,
@@ -2434,11 +2456,12 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBytes:&args length:sizeof(args) atIndex:2];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {
@@ -4766,6 +4789,8 @@ static bool ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    const int is_q = ggml_is_quantized(src1->type) ? 1 : 0;
+
                     // 2*(2*ncpsg + nqptg)*(nsg)
                     // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
                     //
@@ -4773,7 +4798,7 @@ static bool ggml_metal_encode_node(
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(2*ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -4810,9 +4835,9 @@ static bool ggml_metal_encode_node(
                     // and store the soft_max values and the mask
                     //
                     // ne00*(nsg)
-                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
+                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {