llama_cpp 0.5.0 → 0.5.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 715eab98a76ed825d66da6e4fcc84154dca8eed76f6cf6625d210a1ffb702958
-  data.tar.gz: 3ceafc312354d245e485b664d71450cd9c27bcd89f5faec91af6cdf1221c251f
+  metadata.gz: fd67587510fff74b8b1d55e2e5861711709dfb5d8c44cf40b3bf762276e57d5b
+  data.tar.gz: 5cb5319136e538eb2ec9a6406caaaacdabdb2dceec5cade43769eda1b02de9c5
 SHA512:
-  metadata.gz: 7ebe959d9380c9d981156606fdd8a6bcea9b88914923e693b400cfcd605b8c216bdfdcc807c0e72a21fe5fc6d7d623118fc7246524d7f59acdb8bc0064d736bc
-  data.tar.gz: c6d428234d866c09d227b5c308a573e9721454ded3f7fdd36880706e7c47c72c67e6fed119c75d6898c6a1149cde853e5dbb59e3a390ef3d370aab4f0d6be548
+  metadata.gz: c2ab28fe9bf5674976ff2e676ea4d76157bd2ebf24b92ca2f959a6cdf2c19de94fe95d76ab21ca313d9017f835387b0f9ad616cb3700024fc5394fa1e9984fda
+  data.tar.gz: 0ce0be3db250eb7d35f3784bd7a3bd54e7ab8833378745417da3504f69bc31910d4fec459d29ad28218fce2614e8321462e9873c96ed1c3793eb5f9bbe5a9eac

data/CHANGELOG.md

@@ -1,3 +1,7 @@
+## [[0.5.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.5.0...v0.5.1)] - 2023-09-08
+
+- Bump bundled llama.cpp from master-b1140 to master-b1198.
+
 ## [[0.5.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.4.0...v0.5.0)] - 2023-09-02
 
 **Breaking Changes**

data/ext/llama_cpp/src/ggml-alloc.c

@@ -1,3 +1,8 @@
+// defines MAP_ANONYMOUS
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
 #include "ggml-alloc.h"
 #include "ggml.h"
 #include <assert.h>
@@ -6,6 +11,26 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
+
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
@@ -99,19 +124,24 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(alloc);
 }
 
+// check if a tensor is allocated by this buffer
+static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
 #ifdef GGML_ALLOCATOR_DEBUG
-    GGML_ASSERT(ggml_is_view(tensor) == false); // views generally get data pointer from one of their sources
+    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
     GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
 #endif
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -135,14 +165,14 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
-            max_avail = MAX(max_avail, block->size);
         } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
-        return;
+            return;
         }
     }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
@@ -177,17 +207,17 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     void * ptr = tensor->data;
 
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+    if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
         return;
     }
 
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
 
@@ -281,17 +311,68 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// address and size of the buffer when measuring
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+// OS specific functions to allocate and free uncommitted virtual memory
+static void * alloc_vmem(size_t size) {
+#if defined(_WIN32)
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
+}
+
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 1TB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
+
     *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
-        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.data          = */ base_addr,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -311,6 +392,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
+    }
     free(alloc);
 }
 
@@ -380,8 +464,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if ((char *) parent->data < (char *) alloc->data ||
-                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                    if (ggml_allocr_is_own(alloc, parent) == false) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
@@ -415,7 +498,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocator_alloc_graph_tensors_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -493,11 +576,10 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 AT_PRINTF("\n");
             }
 
-
             // update parents
             // update immediately if there is no parse_seq
             // update only at barriers if there is parse_seq
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
+            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
                 int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
                 int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
                 for (int i = update_start; i < update_end; i++) {
@@ -521,12 +603,12 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                                 view_src_hn->n_views -= 1;
                                 AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
                                 if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, view_src);
+                                    ggml_allocr_free_tensor(alloc, view_src);
                                 }
                             }
                             else {
                                 if (parent->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, parent);
+                                    ggml_allocr_free_tensor(alloc, parent);
                                 }
                             }
                         }
@@ -543,7 +625,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             for (int i = 0; outputs[g][i] != NULL; i++) {
                 struct ggml_tensor * output = outputs[g][i];
                 AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
+                ggml_allocr_free_tensor(alloc, output);
             }
         }
     }
@@ -552,5 +634,5 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }

data/ext/llama_cpp/src/ggml-cuda.cu

@@ -81,12 +81,29 @@
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300
 
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
     const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
     const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
     const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
     return reinterpret_cast<const int&>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int&>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
 }
 
 static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
@@ -447,58 +464,91 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] / (1.0f + expf(-x[i]));
 }
 
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+template <int block_size>
 static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
     const float eps = 1e-5f;
 
-    float mean = 0.0f;
-    float var = 0.0f;
+    float2 mean_var = make_float2(0.f, 0.f);
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[row*ncols + col];
-        mean += xi;
-        var += xi * xi;
+        mean_var.x += xi;
+        mean_var.y += xi * xi;
     }
 
     // sum up partial sums
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        mean += __shfl_xor_sync(0xffffffff, mean, mask, 32);
-        var += __shfl_xor_sync(0xffffffff, var, mask, 32);
+    mean_var = warp_reduce_sum(mean_var);
+    if (block_size > WARP_SIZE) {
+        __shared__ float2 s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        __syncthreads();
+        mean_var = s_sum[lane_id];
+        mean_var = warp_reduce_sum(mean_var);
     }
 
-    mean /= ncols;
-    var = var / ncols - mean * mean;
-    const float inv_var = rsqrtf(var + eps);
+    const float mean = mean_var.x / ncols;
+    const float var = mean_var.y / ncols - mean * mean;
+    const float inv_std = rsqrtf(var + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
+    }
+}
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
-        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_var;
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
     }
+    return x;
 }
 
+template <int block_size>
 static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
     float tmp = 0.0f; // partial sum for thread in warp
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[row*ncols + col];
         tmp += xi * xi;
     }
 
     // sum up partial sums
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
     }
 
     const float mean = tmp / ncols;
     const float scale = rsqrtf(mean + eps);
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         dst[row*ncols + col] = scale * x[row*ncols + col];
     }
 }
@@ -4186,14 +4236,24 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
 
 static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    }
 }
 
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
 }
 
 static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) {

data/ext/llama_cpp/src/ggml-metal.m

@@ -76,6 +76,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
@@ -116,10 +117,24 @@ static NSString * const msl_library_source = @"see metal.metal";
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
     metal_printf("%s: allocating\n", __func__);
 
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
+    id <MTLDevice> device;
+    NSString * s;
+    for (device in devices) {
+        s = [device name];
+        metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
+    }
 
+    // Pick and show default Metal device
+    device = MTLCreateSystemDefaultDevice();
+    s = [device name];
+    metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]);
+
+    // Configure context
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
-    ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
     ctx->concur_list_len = 0;
@@ -205,6 +220,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
@@ -270,6 +286,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(norm);
     GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
@@ -310,7 +327,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
 
 void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
-    const int result = posix_memalign((void **) &data, getpagesize(), n);
+    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
     if (result != 0) {
         metal_printf("%s: error: posix_memalign failed\n", __func__);
         return NULL;
@@ -384,7 +401,7 @@ bool ggml_metal_add_buffer(
             }
         }
 
-        const size_t size_page = getpagesize();
+        const size_t size_page = sysconf(_SC_PAGESIZE);
 
         size_t size_aligned = size;
         if ((size_aligned % size_page) != 0) {
@@ -854,7 +871,11 @@ void ggml_metal_graph_compute(
                                         {
                                             nth0 = 32;
                                             nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            if (ne11 * ne12 < 4) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            } else {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            }
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
@@ -906,8 +927,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 2;
-                                            nth1 = 32;
+                                            nth0 = 4; //1;
+                                            nth1 = 8; //32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -955,9 +976,12 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
-                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
+                                    src0t == GGML_TYPE_Q2_K) {// || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -971,8 +995,8 @@ void ggml_metal_graph_compute(
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    int64_t ny = (ne11 + 3)/4;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;
@@ -1117,7 +1141,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
                             [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
                         } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY: