llama_cpp 0.3.4 → 0.3.5

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

@@ -61,6 +61,13 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 // get data from the device into host memory
 void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
 
+// try to find operations that can be run concurrently in the graph
+// you should run it again if the topology of your graph changes
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+// if the graph has been optimized for concurrently dispatch
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

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

@@ -36,12 +36,16 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
+    int concur_list[GGML_MAX_NODES];
+    int concur_list_len;
+
     // custom kernels
 #define GGML_METAL_DECL_KERNEL(name) \
     id<MTLFunction>             function_##name; \
     id<MTLComputePipelineState> pipeline_##name
 
     GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
@@ -97,6 +101,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
+    ctx->concur_list_len = 0;
 
     // determine if we can use MPS
     if (MPSSupportsMTLDevice(ctx->device)) {
@@ -157,6 +162,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
 
         GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(add_row);
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
@@ -215,6 +221,13 @@ void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
     ctx->n_cb = n_cb;
 }
 
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+    if (ctx->concur_list_len) {
+        return true;
+    }
+    return false;
+}
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -353,11 +366,98 @@ void ggml_metal_get_tensor(
     memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
 }
 
+void ggml_metal_graph_find_concurrency(
+        struct ggml_metal_context * ctx,
+        struct ggml_cgraph * gf) {
+    int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
+    int nodes_unused[GGML_MAX_NODES];
+
+    for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;}
+    for (int i = 0; i < gf->n_nodes; i++) {nodes_unused[i] = 1;}
+    ctx->concur_list_len = 0;
+
+    int n_left = gf->n_nodes;
+    int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0;  // at ctx->concur_list, the last layer (level) ends at level_pos
+
+    while (n_left > 0) {
+        // number of nodes at a layer (that can be issued concurrently)
+        int concurrency = 0;
+        for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
+            if (nodes_unused[i]) {
+                // if the requirements for gf->nodes[i] are satisfied
+                int exe_flag=1;
+                // scan all srcs
+                for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
+                    struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
+                    if (src_cur) {
+                        // if is leaf nodes it's satisfied.
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;}
+
+                        // otherwise this src should be the output from previous nodes.
+                        int is_found = 0;
+                        // scan 2*search_depth back because we inserted barrier.
+                        for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                            if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;}
+                        }
+                        if (is_found == 0) {exe_flag = 0; break;}
+                    }
+                }
+                if (exe_flag) {
+                    // check if nodes[i]'s data will be overwritten by a node before nodes[i].
+                    // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
+                    int64_t data_start = (int64_t) gf->nodes[i]->data;
+                    int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    for (int j = n_start; j < i; j++) {
+                        if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
+                                            && gf->nodes[j]->op != GGML_OP_VIEW \
+                                            && gf->nodes[j]->op != GGML_OP_TRANSPOSE \
+                                            && gf->nodes[j]->op != GGML_OP_PERMUTE) {
+                            if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
+                                ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
+                                continue;
+                            } else {
+                                exe_flag = 0;
+                            }
+                        }
+                    }
+                }
+                if (exe_flag) {
+                    ctx->concur_list[level_pos + concurrency] = i;
+                    nodes_unused[i] = 0;
+                    concurrency++;
+                    ctx->concur_list_len++;
+                }
+            }
+        }
+        n_left -= concurrency;
+        // adding a barrier different layer
+        ctx->concur_list[level_pos + concurrency] = -1;
+        ctx->concur_list_len++;
+        // jump all sorted nodes at nodes_bak
+        while (!nodes_unused[n_start]) {n_start++;}
+        level_pos += concurrency + 1;
+    }
+
+    if (ctx->concur_list_len > GGML_MAX_NODES) {
+        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+    }
+}
+
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
     metal_printf("%s: evaluating graph\n", __func__);
 
+    // if there is ctx->concur_list, dispatch concurrently
+    // else fallback to serial dispatch
+    MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
+
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES;
+
+    const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
+    edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
+
     // create multiple command buffers and enqueue them
     // then, we encode the graph into the command buffers in parallel
 
@@ -376,7 +476,7 @@ void ggml_metal_graph_compute(
     dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
 
     for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
+        const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
 
         dispatch_async(queue, ^{
             size_t offs_src0 = 0;
@@ -387,10 +487,21 @@ void ggml_metal_graph_compute(
 
             id<MTLComputeCommandEncoder> encoder = nil;
 
-            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+
+            for (int ind = node_start; ind < node_end; ++ind) {
+                const int i = has_concur ? ctx->concur_list[ind] : ind;
+
+                if (i == -1) {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                        continue;
+                    }
+                    [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
+                    continue;
+                }
 
-            for (int i = node_start; i < node_end; ++i) {
                 metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
@@ -461,13 +572,19 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ADD:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_add];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
 
                             const int64_t n = ggml_nelements(dst);
 
@@ -476,7 +593,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             if (ggml_nelements(src1) == ne10) {
@@ -497,7 +614,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_SCALE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float scale = *(const float *) src1->data;
@@ -511,52 +628,60 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
-                    case GGML_OP_SILU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_silu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_RELU:
-                        {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_relu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    case GGML_OP_UNARY:
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_RELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_relu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_GELU:
+                                {
+                                    if (encoder == nil) {
+                                        encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
+                                    }
+
+                                    [encoder setComputePipelineState:ctx->pipeline_gelu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ASSERT(false);
+                                }
                         } break;
-                    case GGML_OP_GELU:
-                    {
-                            if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_gelu];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                    } break;
                     case GGML_OP_SOFT_MAX:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -574,10 +699,10 @@ void ggml_metal_graph_compute(
                     case GGML_OP_DIAG_MASK_INF:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *)(dst->op_params))[0];
 
                             [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -637,7 +762,7 @@ void ggml_metal_graph_compute(
                                 }
                             } else {
                                 if (encoder == nil) {
-                                    encoder = [command_buffer computeCommandEncoder];
+                                    encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                                 }
 
                                 int nth0 = 32;
@@ -764,7 +889,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_GET_ROWS:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             switch (src0->type) {
@@ -793,10 +918,11 @@ void ggml_metal_graph_compute(
                     case GGML_OP_RMS_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const float eps = 1e-6f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
                             const int nth = 512;
 
@@ -815,7 +941,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_NORM:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const float eps = 1e-5f;
@@ -837,14 +963,15 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ALIBI:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
-                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
-                            const int   n_head   = ((int32_t *) src1->data)[1];
-                            const float max_bias = ((float *)   src1->data)[2];
+                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            const int n_head = ((int32_t *) dst->op_params)[1];
+                            float max_bias;
+                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
                             if (__builtin_popcount(n_head) != 1) {
                                 GGML_ASSERT(false && "only power-of-two n_head implemented");
@@ -879,18 +1006,17 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ROPE:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
-                            const int n_dims = ((int32_t *) src1->data)[1];
-                            const int mode   = ((int32_t *) src1->data)[2];
-
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *) dst->op_params)[0];
+                            const int n_dims = ((int32_t *) dst->op_params)[1];
+                            const int mode   = ((int32_t *) dst->op_params)[2];
 
                             float freq_base;
                             float freq_scale;
-                            memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
+                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -919,10 +1045,12 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
+                    case GGML_OP_DUP:
                     case GGML_OP_CPY:
+                    case GGML_OP_CONT:
                         {
                             if (encoder == nil) {
-                                encoder = [command_buffer computeCommandEncoder];
+                                encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
                             }
 
                             const int nth = 32;
@@ -969,8 +1097,10 @@ void ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     default:
-                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                        GGML_ASSERT(false);
+                        {
+                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_ASSERT(false);
+                        }
                 }
             }
 

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

@@ -67,6 +67,17 @@ kernel void kernel_add(
     dst[tpig] = src0[tpig] + src1[tpig];
 }
 
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        device const float * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] + src1[tpig % ne00];
+}
+
 kernel void kernel_mul(
         device const float * src0,
         device const float * src1,
@@ -376,87 +387,90 @@ kernel void kernel_rms_norm(
     }
 }
 
-// function for calculate inner product between a q4_0 block and 32 floats (yl), sumy is SUM(yl[i])
-float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl) {
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
-    float4 acc = 0.f;
-    device uint16_t * qs = ((device uint16_t *)qb_curr + 1);
-    for (int i = 0; i < 16; i+=2) {
-        acc[0] += yl[i]      * (qs[i / 2] & 0x000F);
-        acc[1] += yl[i + 16] * (qs[i / 2] & 0x00F0);
-        acc[2] += yl[i +  1] * (qs[i / 2] & 0x0F00);
-        acc[3] += yl[i + 17] * (qs[i / 2] & 0xF000);
+    float2 acc = 0.f;
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
     }
-    return d * (sumy * -8.f + acc[0] + acc[1]/16.f + acc[2]/256.f + acc[3]/4096.f);
+    return d * (sumy * -8.f + acc[0] + acc[1]);
 }
 
-// function for calculate inner product between a q4_1 block and 32 floats (yl), sumy is SUM(yl[i])
-float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl) {
+// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
     float m = qb_curr->m;
-    float4 acc = 0.f;
-    device uint16_t * qs = ((device uint16_t *)qb_curr + 2);
-    for (int i = 0; i < 16; i+=2) {
-        acc[0] += yl[i]      * (qs[i / 2] & 0x000F);
-        acc[1] += yl[i + 16] * (qs[i / 2] & 0x00F0);
-        acc[2] += yl[i +  1] * (qs[i / 2] & 0x0F00);
-        acc[3] += yl[i + 17] * (qs[i / 2] & 0xF000);
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+    float2 acc = 0.f;
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
     }
-    return d * (acc[0] + acc[1]/16.f + acc[2]/256.f + acc[3]/4096.f) + sumy * m;
+    return d * (acc[0] + acc[1]) + sumy * m;
 }
 
 // putting them in the kernel cause a significant performance penalty
 #define N_DST 4 // each SIMD group works on 4 rows
 #define N_SIMDGROUP 2 // number of SIMD groups in a thread group
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-template<typename block_q_type>
+//Note: This is a template, but strictly speaking it only applies to
+//      quantizations where the block size is 32. It also does not
+//      giard against the number of rows not being divisible by
+//      N_DST, so this is another explicit assumption of the implementation.
+template<typename block_q_type, int nr, int nsg, int nw>
 void mul_vec_q_n_f32(device const void * src0, device const float * src1, device float * dst,
                     int64_t ne00, int64_t ne10, int64_t ne0, int64_t ne01,
                     uint2 tgpig, uint tiisg, uint sgitg) {
     const int nb = ne00/QK4_0;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
-    device const block_q_type * x = (device const block_q_type *) src0 + (r0 * N_SIMDGROUP + sgitg) * N_DST * nb;
+    const int first_row = (r0 * nsg + sgitg) * nr;
+    device const block_q_type * x = (device const block_q_type *) src0 + first_row * nb;
     device const float      * y = (device const float      *) src1 + r1*ne10;
-    float4 y_curr[8];       // src1 vector cache
-    float sumf[N_DST]={0.f}, all_sum;
-    thread float * yl=(thread float *)y_curr;
+    float yl[16];       // src1 vector cache
+    float sumf[nr]={0.f};
 
-    // each thread in a SIMD group deals with 1 block.
-    for (int column = 0; column < nb / N_SIMDWIDTH; column++) {
-        float sumy = 0;
-        for (int i = 0; i < QK4_0 / 4; i++) {
-            y_curr[i] = *((device float4  *)(y + N_SIMDWIDTH * (tiisg + column * QK4_0)) + i);
-            sumy += y_curr[i][0] + y_curr[i][1] + y_curr[i][2] + y_curr[i][3];
-        }
+    const int ix = tiisg/2;
+    const int il = 8*(tiisg%2);
 
-        for (int row = 0; row < N_DST; row++) {
-            sumf[row] += block_q_n_dot_y(x+(tiisg + row * nb + column * N_SIMDWIDTH), sumy, yl);
-        }
-    }
+    device const float * yb = y + ix * QK4_0 + il;
 
-    // from now loads two rows every time and 16 blocks per row
-    int ir = tiisg / (N_SIMDWIDTH / 2);
-    int ib = tiisg % (N_SIMDWIDTH / 2);
-    for (int ind = 0; ind < (nb % N_SIMDWIDTH + N_SIMDWIDTH / 2 - 1)/(N_SIMDWIDTH / 2); ind++) {
-        int nb_start = (nb / N_SIMDWIDTH) * N_SIMDWIDTH + ind * (N_SIMDWIDTH / 2); //where the left blocks start
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += nw/2) {
         float sumy = 0;
-        for (int i = 0; i < QK4_0 / 4; i++) {
-            y_curr[i] = *((device float4 *)(y + (nb_start + ib) * QK4_0) + i);
-            sumy += y_curr[i][0] + y_curr[i][1] + y_curr[i][2] + y_curr[i][3];
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
         }
 
-        for (int row = 0; row < N_DST; row+=2) {
-            if (nb_start + ib < nb) {
-                sumf[row + ir] += block_q_n_dot_y(x + (nb_start + ib + (row + ir) * nb), sumy, yl);
-            }
+        for (int row = 0; row < nr; row++) {
+            sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
         }
+
+        yb += QK4_0 * 16;
     }
 
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0 && ((r0 * N_SIMDGROUP + sgitg) * N_DST + row) < ne01) {
-            dst[r1*ne0 + (r0 * N_SIMDGROUP + sgitg) * N_DST + row] = all_sum;
+    for (int row = 0; row < nr; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + first_row + row] = tot;
         }
     }
 }
@@ -472,7 +486,7 @@ kernel void kernel_mul_mat_q4_0_f32(
         uint2 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32<block_q4_0>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mat_q4_1_f32(
@@ -486,7 +500,7 @@ kernel void kernel_mul_mat_q4_1_f32(
         uint2 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32<block_q4_1>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
+     mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne10,ne0,ne01,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mat_f16_f32(