partial decode ggml bin for more info

llm/ggml-alloc.c

@@ -0,0 +1,567 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-alloc.h"
+#include "ggml.h"
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED(x) (void)(x)
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF printf
+#define AT_PRINTF(...) ((void)0)
+
+struct hash_node {
+    struct ggml_tensor * t;
+    int n_children;
+    int n_views;
+};
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
+    size_t h = hash(t);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i].t != NULL) {
+        if (hash_table[i].t == t) {
+            return &hash_table[i];
+        }
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // hash table is full
+            GGML_ASSERT(false);
+        }
+    }
+
+    hash_table[i].t = t;
+    return &hash_table[i];
+}
+
+// TODO: GGML_PAD ?
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+struct free_block {
+    void * addr;
+    size_t size;
+};
+
+#define MAX_FREE_BLOCKS 128
+
+struct ggml_allocr {
+    void * data;
+    size_t size;
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+    size_t max_size;
+    bool measure;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct ggml_tensor * allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == NULL) {
+            alloc->allocated_tensors[i] = tensor;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == tensor ||
+            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
+            alloc->allocated_tensors[i] = NULL;
+            return;
+        }
+    }
+    printf("tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+
+static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
+void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    AT_PRINTF("block %d\n", best_fit_block);
+
+    if (best_fit_block == -1) {
+        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;
+    }
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    void * addr = block->addr;
+    block->addr = (char*)block->addr + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    tensor->data = addr;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, tensor);
+    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    if (cur_max > alloc->max_size) {
+        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i]) {
+                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            }
+        }
+        printf("\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+}
+
+// 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) {
+    void * ptr = tensor->data;
+
+    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+        // 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 = 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);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if ((char*)block->addr + block->size == ptr) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if ((char*)ptr + size == block->addr) {
+            block->addr = ptr;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+}
+
+void ggml_allocr_reset(struct ggml_allocr * alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
+}
+
+struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ data,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ false,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    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
+
+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) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ true,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+void ggml_allocr_free(struct ggml_allocr * alloc) {
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+    return alloc->measure;
+}
+
+//////////// compute graph allocator
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
+    switch (t->op) {
+        case GGML_OP_PERMUTE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+            return t->src[0];
+        case GGML_OP_CPY:
+            return t->src[1];
+        default:
+            return NULL;
+    }
+}
+
+static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
+    struct ggml_tensor * parent = t;
+    do {
+        parent = get_view_parent(parent);
+    } while (ggml_is_view(parent));
+    return parent;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_ACC:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SET:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
+    struct hash_node * ht = alloc->hash_table;
+    if (node->data == NULL) {
+        if (ggml_is_view(node)) {
+            size_t offset;
+            switch(node->op) {
+                case GGML_OP_VIEW:
+                    memcpy(&offset, node->op_params, sizeof(size_t));
+                    node->data = (char *) node->src[0]->data + offset;
+                    break;
+                case GGML_OP_PERMUTE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_TRANSPOSE:
+                    node->data = node->src[0]->data;
+                    break;
+                case GGML_OP_CPY:
+                    node->data = node->src[1]->data;
+                    break;
+                default:
+                    GGML_ASSERT(!"unknown view op");
+                    break;
+            }
+        } else {
+            // see if we can reuse a parent's buffer (inplace)
+            if (ggml_op_can_inplace(node->op)) {
+                for (int i = 0; i < GGML_MAX_SRC; i++) {
+                    struct ggml_tensor * parent = node->src[i];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    struct hash_node * p_hn = hash_get(ht, parent);
+                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
+                                // the parent's data that it will need later (same layout requirement). the problem is that then
+                                // we cannot free the tensor because the original address of the allocation is lost.
+                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
+                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
+                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                                node->data = parent->data;
+                                return;
+                            }
+                        }
+                        else {
+                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                            node->data = parent->data;
+                        }
+                        return;
+                    }
+                }
+            }
+            ggml_allocr_alloc(alloc, node);
+        }
+    }
+}
+
+static size_t ggml_allocator_alloc_graph_tensors_n(
+    struct ggml_allocr * alloc,
+    struct ggml_cgraph ** graphs, int n_graphs,
+    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+
+    // reset hash table
+    struct hash_node * ht = alloc->hash_table;
+    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+
+    // count number of children and views
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            if (ggml_is_view(node)) {
+                struct ggml_tensor * view_src = get_view_source(node);
+                hash_get(ht, view_src)->n_views += 1;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                hash_get(ht, parent)->n_children += 1;
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
+        // graph inputs are allocated first to ensure that they are not overwritten by each other
+        if (inputs != NULL && inputs[g] != NULL) {
+            for (int i = 0; inputs[g][i] != NULL; i++) {
+                struct ggml_tensor * input = inputs[g][i];
+                AT_PRINTF("input: %s\n", input->name);
+                allocate_node(alloc, input);
+            }
+        }
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                allocate_node(alloc, parent);
+            }
+
+            // allocate node
+            allocate_node(alloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
+            }
+            AT_PRINTF("\n");
+
+            // update parents
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                struct hash_node * p_hn = hash_get(ht, parent);
+                p_hn->n_children -= 1;
+
+                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = get_view_source(parent);
+                        struct hash_node * view_src_hn = hash_get(ht, view_src);
+                        view_src_hn->n_views -= 1;
+                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->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);
+                        }
+                    }
+                    else {
+                        if (parent->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, parent);
+                        }
+                    }
+                }
+            }
+            AT_PRINTF("\n");
+        }
+        // free graph outputs here that wouldn't be freed otherwise because they have no children
+        if (outputs != NULL && outputs[g] != NULL) {
+            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);
+            }
+        }
+    }
+
+    return alloc->max_size;
+}
+
+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);
+}

llm/ggml-alloc.h

@@ -0,0 +1,48 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+
+GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+
+
+#ifdef  __cplusplus
+}
+#endif

llm/ggml-cuda.h

@@ -0,0 +1,63 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+void   ggml_init_cublas(void);
+void   ggml_cuda_set_tensor_split(const float * tensor_split);
+
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+// TODO: export these with GGML_API
+void * ggml_cuda_host_malloc(size_t size);
+void   ggml_cuda_host_free(void * ptr);
+
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
+void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
+bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif

llm/ggml-metal.h

@@ -0,0 +1,106 @@
+//go:build darwin
+
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+// number of command buffers to use
+struct ggml_metal_context * ggml_metal_init(int n_cb);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// set the number of command buffers to use
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size,
+                           size_t   max_size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// 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);
+
+#ifdef __cplusplus
+}
+#endif
+

llm/ggml-mpi.c

@@ -0,0 +1,244 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-mpi.h"
+
+#include "ggml.h"
+
+#include <mpi.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+
+#define UNUSED GGML_UNUSED
+
+struct ggml_mpi_context {
+    int rank;
+    int size;
+};
+
+void ggml_mpi_backend_init(void) {
+    MPI_Init(NULL, NULL);
+}
+
+void ggml_mpi_backend_free(void) {
+    MPI_Finalize();
+}
+
+struct ggml_mpi_context * ggml_mpi_init(void) {
+    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
+
+    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
+
+    return ctx;
+}
+
+void ggml_mpi_free(struct ggml_mpi_context * ctx) {
+    free(ctx);
+}
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
+    return ctx->rank;
+}
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads) {
+    UNUSED(ctx_mpi);
+
+    // synchronize the worker node parameters with the root node
+    MPI_Barrier(MPI_COMM_WORLD);
+
+    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
+}
+
+static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
+    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
+    if (t == NULL) {
+        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
+        return -1;
+    }
+
+    for (int i = 0; i < gf->n_nodes; i++) {
+        if (gf->nodes[i] == t) {
+            return i;
+        }
+    }
+
+    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
+    return -1;
+}
+
+static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    MPI_Status status; UNUSED(status);
+
+    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+// TODO: there are many improvements that can be done to this implementation
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
+    if (inp_tokens == NULL) {
+        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
+        return;
+    }
+
+    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
+    if (inp0 == NULL) {
+        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
+        return;
+    }
+
+    GGML_ASSERT(inp0 == gf->nodes[0]);
+
+    // distribute the compute graph into slices across the MPI nodes
+    //
+    // the main node (0) processes the last layers + the remainder of the compute graph
+    // and is responsible to pass the input tokens to the first node (1)
+    //
+    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
+    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
+    // ...
+    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
+    // node 0:   [(n-1) * n_per_node,            n_nodes)
+    //
+    if (mpi_rank > 0) {
+        if (mpi_rank == 1) {
+            // the first node (1) receives the input tokens from the main node (0)
+            ggml_mpi_tensor_recv(inp_tokens, 0);
+        } else {
+            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
+            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
+        }
+    } else if (mpi_size > 1) {
+        // node 0 sends the input tokens to node 1
+        ggml_mpi_tensor_send(inp_tokens, 1);
+
+        // recv the output data from the last node
+        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
+    }
+
+    {
+        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
+
+        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
+
+        const int il0 =               (mpi_idx + 0) * n_per_node;
+        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
+
+        char name_l0[GGML_MAX_NAME];
+        char name_l1[GGML_MAX_NAME];
+
+        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
+        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
+
+        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
+        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
+
+        if (idx_l0 < 0 || idx_l1 < 0) {
+            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
+            return;
+        }
+
+        // attach the input data to all nodes that need it
+        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
+        for (int i = idx_l0; i < idx_l1; i++) {
+            if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[0] =  inp0;
+            }
+            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[1] =  inp0;
+            }
+        }
+
+        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
+        for (int i = 1; i < idx_l1 - idx_l0; i++) {
+            gf->nodes[i] = gf->nodes[idx_l0 + i];
+            gf->grads[i] = gf->grads[idx_l0 + i];
+        }
+
+        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
+        if (mpi_idx != 0) {
+            gf->nodes[0]->op = GGML_OP_NONE;
+        }
+
+        gf->n_nodes = idx_l1 - idx_l0;
+
+        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
+    }
+}
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    UNUSED(n_layers);
+
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    // send the output data to the next node
+    if (mpi_rank > 0) {
+        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
+    }
+}

llm/ggml-mpi.h

@@ -0,0 +1,67 @@
+//go:build mpi
+
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+struct ggml_context;
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_mpi_context;
+
+void ggml_mpi_backend_init(void);
+void ggml_mpi_backend_free(void);
+
+struct ggml_mpi_context * ggml_mpi_init(void);
+void ggml_mpi_free(struct ggml_mpi_context * ctx);
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx);
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads);
+
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+#ifdef __cplusplus
+}
+#endif

llm/ggml-opencl.h

@@ -0,0 +1,53 @@
+//go:build opencl
+
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void ggml_cl_init(void);
+
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
+
+void ggml_cl_free_data(const struct ggml_tensor* tensor);
+
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+#ifdef  __cplusplus
+}
+#endif

llm/ggml.go

@@ -0,0 +1,180 @@
+package llm
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+type ModelFamily string
+
+const ModelFamilyLlama ModelFamily = "llama"
+
+type ModelType uint32
+
+const (
+	ModelType3B  ModelType = 26
+	ModelType7B  ModelType = 32
+	ModelType13B ModelType = 40
+	ModelType30B ModelType = 60
+	ModelType65B ModelType = 80
+)
+
+type FileType uint32
+
+const (
+	FileTypeF32 FileType = iota
+	FileTypeF16
+	FileTypeQ4_0
+	FileTypeQ4_1
+	FileTypeQ4_1_F16
+	FileTypeQ8_0 = iota + 3
+	FileTypeQ5_0
+	FileTypeQ5_1
+	FileTypeQ2_K
+	FileTypeQ3_K
+	FileTypeQ4_K
+	FileTypeQ5_K
+	FileTypeQ6_K
+	FileTypeUnknown = -1
+)
+
+type GGML struct {
+	ModelFamily
+	ModelType
+
+	magic uint32
+	container
+
+	llamaHyperparameters
+}
+
+type container interface {
+	Name() string
+	Decode(io.Reader) error
+}
+
+type containerGGML struct {
+}
+
+func (c *containerGGML) Name() string {
+	return "ggml"
+}
+
+func (c *containerGGML) Decode(r io.Reader) error {
+	return nil
+}
+
+type containerGGMF struct {
+	version uint32
+}
+
+func (c *containerGGMF) Name() string {
+	return "ggmf"
+}
+
+func (c *containerGGMF) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerGGJT struct {
+	version uint32
+}
+
+func (c *containerGGJT) Name() string {
+	return "ggjt"
+}
+
+func (c *containerGGJT) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1, 2, 3:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+type containerLORA struct {
+	version uint32
+}
+
+func (c *containerLORA) Name() string {
+	return "ggla"
+}
+
+func (c *containerLORA) Decode(r io.Reader) error {
+	var version uint32
+	binary.Read(r, binary.LittleEndian, &version)
+
+	switch version {
+	case 1:
+	default:
+		return errors.New("invalid version")
+	}
+
+	c.version = version
+	return nil
+}
+
+const (
+	// / Magic constant for `ggml` files (unversioned).
+	FILE_MAGIC_GGML = 0x67676d6c
+	// / Magic constant for `ggml` files (versioned, ggmf).
+	FILE_MAGIC_GGMF = 0x67676d66
+	// / Magic constant for `ggml` files (versioned, ggjt).
+	FILE_MAGIC_GGJT = 0x67676a74
+	// / Magic constant for `ggla` files (LoRA adapter).
+	FILE_MAGIC_GGLA = 0x67676C61
+)
+
+func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
+	var ggml GGML
+	binary.Read(r, binary.LittleEndian, &ggml.magic)
+
+	switch ggml.magic {
+	case FILE_MAGIC_GGML:
+		ggml.container = &containerGGML{}
+	case FILE_MAGIC_GGMF:
+		ggml.container = &containerGGMF{}
+	case FILE_MAGIC_GGJT:
+		ggml.container = &containerGGJT{}
+	case FILE_MAGIC_GGLA:
+		ggml.container = &containerLORA{}
+	default:
+		return nil, errors.New("invalid file magic")
+	}
+
+	if err := ggml.Decode(r); err != nil {
+		return nil, err
+	}
+
+	// different model types may have different layouts for hyperparameters
+	switch hint {
+	case ModelFamilyLlama:
+		binary.Read(r, binary.LittleEndian, &ggml.llamaHyperparameters)
+		// TODO: sanity check hyperparameters
+	default:
+		return nil, fmt.Errorf("unsupported model type: %s", hint)
+	}
+
+	// final model type
+	ggml.ModelFamily = hint
+	ggml.ModelType = ModelType(ggml.NumLayer)
+	return &ggml, nil
+}

llm/k_quants.h

@@ -0,0 +1,191 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+

llm/llama-util.h

@@ -0,0 +1,530 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+#include <stdexcept>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, len, 1, fp);
+        if (ferror(fp)) {
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        }
+        if (ret != 1) {
+            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, len, 1, fp);
+        if (ret != 1) {
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+        // prefetch/readahead impairs performance on NUMA systems
+        if (numa) { prefetch = 0; }
+#ifdef __linux__
+        if (prefetch) { flags |= MAP_POPULATE; }
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+        if (numa) {
+            // advise the kernel not to use readahead
+            // (because the next page might not belong on the same node)
+            if (madvise(addr, file->size, MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+
+        if (hMapping == NULL) {
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
+        }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
+        throw std::runtime_error(std::string("mmap not supported"));
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            struct rlimit lock_limit;
+            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
+                suggest = false;
+            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
+                suggest = false;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * ptr, size_t len) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(ptr, len)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                    len, size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = len + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    void raw_unlock(const void * addr, size_t len) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    llama_buffer() = default;
+
+    void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
+        delete[] addr;
+        addr = new uint8_t[len];
+#endif
+        size = len;
+    }
+
+    ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
+        delete[] addr;
+#endif
+        addr = NULL;
+    }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
+};
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    bool is_cuda;
+    size_t size = 0;
+
+    llama_ctx_buffer() = default;
+
+    void resize(size_t size) {
+        free();
+
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        if (addr) {
+            is_cuda = true;
+        }
+        else {
+            // fall back to pageable memory
+            addr = new uint8_t[size];
+            is_cuda = false;
+        }
+        this->size = size;
+    }
+
+    void free() {
+        if (addr) {
+            if (is_cuda) {
+                ggml_cuda_host_free(addr);
+            }
+            else {
+                delete[] addr;
+            }
+        }
+        addr = NULL;
+    }
+
+    ~llama_ctx_buffer() {
+        free();
+    }
+
+    // disable copy and move
+    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
+    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
+#endif

llm/llama.go

@@ -0,0 +1,469 @@
+package llm
+
+/*
+#cgo CPPFLAGS: -O3 -Wall -Wextra -Wno-unused-function -Wno-unused-variable -DNDEBUG -DGGML_USE_K_QUANTS
+#cgo CXXFLAGS: -std=gnu++11
+#cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
+#cgo darwin,arm64 CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+#cgo darwin LDFLAGS: -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+#include <stdlib.h>
+#include "llama.h"
+
+struct llama_sample_options
+{
+	float repeat_penalty;
+	float frequency_penalty;
+	float presence_penalty;
+	float temperature;
+	int32_t top_k;
+	float top_p;
+	float tfs_z;
+	float typical_p;
+	int mirostat;
+	float mirostat_tau;
+	float mirostat_eta;
+	bool penalize_newline;
+};
+
+llama_token llama_sample(
+		struct llama_context *ctx,
+		struct llama_token_data *candidates,
+		size_t n_candidates,
+		const llama_token *last_tokens,
+		size_t n_last_tokens,
+		struct llama_sample_options *opts)
+{
+	llama_token_data_array candidates_p = {
+		candidates,
+		n_candidates,
+		false,
+	};
+
+	struct llama_token_data newline = candidates_p.data[llama_token_nl()];
+
+	llama_sample_repetition_penalty(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->repeat_penalty);
+
+	llama_sample_frequency_and_presence_penalties(
+		ctx, &candidates_p,
+		last_tokens, n_last_tokens,
+		opts->frequency_penalty, opts->presence_penalty);
+
+	if (!opts->penalize_newline) {
+		candidates_p.data[llama_token_nl()] = newline;
+	}
+
+	if (opts->temperature <= 0) {
+		return llama_sample_token_greedy(ctx, &candidates_p);
+	}
+
+	if (opts->mirostat == 1) {
+		int mirostat_m = 100;
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			mirostat_m, &mirostat_mu);
+	} else if (opts->mirostat == 2) {
+		float mirostat_mu = 2.0f * opts->mirostat_tau;
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token_mirostat_v2(
+			ctx, &candidates_p,
+			opts->mirostat_tau, opts->mirostat_eta,
+			&mirostat_mu);
+	} else {
+		llama_sample_top_k(ctx, &candidates_p, opts->top_k, 1);
+		llama_sample_tail_free(ctx, &candidates_p, opts->tfs_z, 1);
+		llama_sample_typical(ctx, &candidates_p, opts->typical_p, 1);
+		llama_sample_top_p(ctx, &candidates_p, opts->top_p, 1);
+		llama_sample_temperature(ctx, &candidates_p, opts->temperature);
+		return llama_sample_token(ctx, &candidates_p);
+	}
+}
+*/
+import "C"
+
+import (
+	"bytes"
+	"embed"
+	"errors"
+	"fmt"
+	"io"
+	"log"
+	"os"
+	"strings"
+	"sync"
+	"unicode/utf8"
+	"unsafe"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+//go:embed ggml-metal.metal
+var fs embed.FS
+
+type llama struct {
+	params *C.struct_llama_context_params
+	model  *C.struct_llama_model
+	ctx    *C.struct_llama_context
+
+	last   []C.llama_token
+	embd   []C.llama_token
+	cursor int
+
+	mu sync.Mutex
+	gc bool
+
+	api.Options
+}
+
+type llamaHyperparameters struct {
+	// NumVocab is the size of the model's vocabulary.
+	NumVocab uint32
+
+	// NumEmbd is the size of the model's embedding layer.
+	NumEmbd uint32
+	NumMult uint32
+	NumHead uint32
+
+	// NumLayer is the number of layers in the model.
+	NumLayer uint32
+	NumRot   uint32
+	// FileType describes the quantization level of the model, e.g. Q4_0, Q5_K, etc.
+	FileType
+}
+
+func newLlama(model string, opts api.Options) (*llama, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	llm := llama{Options: opts}
+
+	C.llama_backend_init(C.bool(llm.UseNUMA))
+
+	params := C.llama_context_default_params()
+	params.seed = C.uint(llm.Seed)
+	params.n_ctx = C.int(llm.NumCtx)
+	params.n_batch = C.int(llm.NumBatch)
+	params.n_gqa = C.int(llm.NumGQA)
+	params.n_gpu_layers = C.int(llm.NumGPU)
+	params.main_gpu = C.int(llm.MainGPU)
+	params.low_vram = C.bool(llm.LowVRAM)
+	params.f16_kv = C.bool(llm.F16KV)
+	params.logits_all = C.bool(llm.LogitsAll)
+	params.vocab_only = C.bool(llm.VocabOnly)
+	params.use_mmap = C.bool(llm.UseMMap)
+	params.use_mlock = C.bool(llm.UseMLock)
+	params.embedding = C.bool(llm.EmbeddingOnly)
+	params.rope_freq_base = C.float(llm.RopeFrequencyBase)
+	params.rope_freq_scale = C.float(llm.RopeFrequencyScale)
+	llm.params = &params
+
+	cModel := C.CString(model)
+	defer C.free(unsafe.Pointer(cModel))
+
+	llm.model = C.llama_load_model_from_file(cModel, params)
+	if llm.model == nil {
+		return nil, errors.New("failed to load model")
+	}
+
+	llm.ctx = C.llama_new_context_with_model(llm.model, params)
+	if llm.ctx == nil {
+		return nil, errors.New("failed to create context")
+	}
+
+	// warm up the model
+	bos := []C.llama_token{C.llama_token_bos()}
+	C.llama_eval(llm.ctx, unsafe.SliceData(bos), C.int(len(bos)), 0, C.int(opts.NumThread))
+	C.llama_reset_timings(llm.ctx)
+
+	return &llm, nil
+}
+
+func (llm *llama) Close() {
+	llm.gc = true
+
+	llm.mu.Lock()
+	defer llm.mu.Unlock()
+
+	defer C.llama_free_model(llm.model)
+	defer C.llama_free(llm.ctx)
+
+	C.llama_print_timings(llm.ctx)
+}
+
+func (llm *llama) SetOptions(opts api.Options) {
+	llm.Options = opts
+}
+
+var errNeedMoreData = errors.New("need more data")
+
+func (llm *llama) Predict(ctx []int, prompt string, fn func(api.GenerateResponse)) error {
+	C.llama_reset_timings(llm.ctx)
+
+	llm.marshalPrompt(ctx, prompt)
+
+	C.llama_set_rng_seed(llm.ctx, C.uint(llm.Seed))
+
+	var b bytes.Buffer
+	for {
+		token, err := llm.next()
+		if llm.gc {
+			return nil
+		} else if errors.Is(err, io.EOF) {
+			break
+		} else if err != nil {
+			return err
+		}
+
+		b.WriteString(llm.Decode(int(token)))
+
+		if err := llm.checkStopConditions(b); err != nil {
+			if errors.Is(err, io.EOF) {
+				break
+			} else if errors.Is(err, errNeedMoreData) {
+				continue
+			}
+
+			return err
+		}
+
+		if utf8.Valid(b.Bytes()) || b.Len() >= utf8.UTFMax {
+			fn(api.GenerateResponse{Response: b.String()})
+			b.Reset()
+		}
+	}
+
+	embd := make([]int, len(llm.embd))
+	for i := range llm.embd {
+		embd[i] = int(llm.embd[i])
+	}
+
+	timings := C.llama_get_timings(llm.ctx)
+	fn(api.GenerateResponse{
+		Done:               true,
+		Context:            embd,
+		SampleCount:        int(timings.n_sample),
+		SampleDuration:     parseDurationMs(float64(timings.t_sample_ms)),
+		PromptEvalCount:    int(timings.n_p_eval),
+		PromptEvalDuration: parseDurationMs(float64(timings.t_p_eval_ms)),
+		EvalCount:          int(timings.n_eval),
+		EvalDuration:       parseDurationMs(float64(timings.t_eval_ms)),
+	})
+
+	return nil
+}
+
+func (llm *llama) checkStopConditions(b bytes.Buffer) error {
+	for _, stopCondition := range llm.Stop {
+		if stopCondition == strings.TrimSpace(b.String()) {
+			return io.EOF
+		} else if strings.HasPrefix(stopCondition, strings.TrimSpace(b.String())) {
+			return errNeedMoreData
+		}
+	}
+
+	return nil
+}
+
+func (llm *llama) marshalPrompt(ctx []int, prompt string) []C.llama_token {
+	tokens := append(ctx, llm.Encode(prompt)...)
+	if llm.NumKeep < 0 {
+		llm.NumKeep = len(tokens)
+	}
+
+	cTokens := make([]C.llama_token, len(tokens))
+	for i := range tokens {
+		cTokens[i] = C.llama_token(tokens[i])
+	}
+
+	// min(llm.NumCtx - 4, llm.NumKeep)
+	if llm.NumCtx-4 < llm.NumKeep {
+		llm.NumKeep = llm.NumCtx - 4
+	}
+
+	if len(tokens) >= llm.NumCtx {
+		// truncate input
+		numLeft := (llm.NumCtx - llm.NumKeep) / 2
+		truncated := cTokens[:llm.NumKeep]
+		erasedBlocks := (len(cTokens) - llm.NumKeep - numLeft - 1) / numLeft
+		truncated = append(truncated, cTokens[llm.NumKeep+erasedBlocks*numLeft:]...)
+		copy(llm.last, cTokens[len(cTokens)-llm.NumCtx:])
+
+		cTokens = truncated
+		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated))
+	} else {
+		llm.last = make([]C.llama_token, llm.NumCtx-len(cTokens))
+		llm.last = append(llm.last, cTokens...)
+	}
+
+	var i int
+	for i = 0; i < len(llm.embd) && i < len(cTokens) && llm.embd[i] == cTokens[i]; i++ {
+		// noop
+	}
+
+	llm.embd = cTokens
+	if i == len(cTokens) {
+		// evaluate at least one token to generate logits
+		i--
+	}
+
+	llm.cursor = i
+
+	log.Printf("prompt: num_past=%d cached=%v eval=%v", i, len(llm.embd[:i]), len(llm.embd[i:]))
+	return cTokens
+}
+
+func (llm *llama) Encode(prompt string) []int {
+	cPrompt := C.CString(prompt)
+	defer C.free(unsafe.Pointer(cPrompt))
+
+	cTokens := make([]C.llama_token, len(prompt)+1)
+	if n := C.llama_tokenize(llm.ctx, cPrompt, unsafe.SliceData(cTokens), C.int(len(cTokens)), true); n > 0 {
+		tokens := make([]int, n)
+		for i := range cTokens[:n] {
+			tokens[i] = int(cTokens[i])
+		}
+
+		return tokens
+	}
+
+	return nil
+}
+
+func (llm *llama) Decode(tokens ...int) string {
+	var sb strings.Builder
+	for _, token := range tokens {
+		sb.WriteString(C.GoString(C.llama_token_to_str(llm.ctx, C.llama_token(token))))
+	}
+
+	return sb.String()
+}
+
+func (llm *llama) next() (C.llama_token, error) {
+	llm.mu.Lock()
+	defer llm.mu.Unlock()
+
+	if len(llm.embd) >= llm.NumCtx {
+		numLeft := (llm.NumCtx - llm.NumKeep) / 2
+		truncated := llm.embd[:llm.NumKeep]
+		truncated = append(truncated, llm.embd[len(llm.embd)-numLeft:]...)
+
+		llm.embd = truncated
+		llm.cursor = llm.NumKeep
+		log.Printf("input truncated: num_ctx=%d num_keep=%d num_left=%d num_tokens=%d cursor=%d", llm.NumCtx, llm.NumKeep, numLeft, len(truncated), llm.cursor)
+	}
+
+	for {
+		if llm.gc {
+			return 0, io.EOF
+		}
+
+		if llm.cursor >= len(llm.embd) {
+			break
+		}
+
+		numEval := len(llm.embd) - llm.cursor
+		if numEval > llm.NumBatch {
+			numEval = llm.NumBatch
+		}
+
+		if retval := C.llama_eval(llm.ctx, unsafe.SliceData(llm.embd[llm.cursor:]), C.int(numEval), C.int(llm.cursor), C.int(llm.NumThread)); retval != 0 {
+			return 0, fmt.Errorf("llama_eval: %d", retval)
+		}
+
+		llm.cursor += numEval
+	}
+
+	var sampleOpts C.struct_llama_sample_options
+	sampleOpts.repeat_penalty = C.float(llm.RepeatPenalty)
+	sampleOpts.frequency_penalty = C.float(llm.FrequencyPenalty)
+	sampleOpts.presence_penalty = C.float(llm.PresencePenalty)
+	sampleOpts.temperature = C.float(llm.Temperature)
+	sampleOpts.top_k = C.int(llm.TopK)
+	sampleOpts.top_p = C.float(llm.TopP)
+	sampleOpts.tfs_z = C.float(llm.TFSZ)
+	sampleOpts.typical_p = C.float(llm.TypicalP)
+	sampleOpts.mirostat = C.int(llm.Mirostat)
+	sampleOpts.mirostat_tau = C.float(llm.MirostatTau)
+	sampleOpts.mirostat_eta = C.float(llm.MirostatEta)
+	sampleOpts.penalize_newline = C.bool(llm.PenalizeNewline)
+
+	numVocab := C.llama_n_vocab(llm.ctx)
+	logits := unsafe.Slice(C.llama_get_logits(llm.ctx), numVocab)
+
+	// TODO: logit bias
+
+	candidates := make([]C.llama_token_data, numVocab)
+	for i := range logits {
+		candidates[i] = C.llama_token_data{
+			id:    C.int(i),
+			logit: logits[i],
+			p:     0,
+		}
+	}
+
+	repeatLastN := llm.RepeatLastN
+	if len(llm.last) < repeatLastN {
+		repeatLastN = len(llm.last)
+	}
+
+	if llm.NumCtx < repeatLastN {
+		repeatLastN = llm.NumCtx
+	}
+
+	lastN := llm.last[len(llm.last)-repeatLastN:]
+
+	token := C.llama_sample(
+		llm.ctx,
+		unsafe.SliceData(candidates), C.size_t(len(candidates)),
+		unsafe.SliceData(lastN), C.size_t(len(lastN)),
+		&sampleOpts,
+	)
+
+	llm.last = append(llm.last, token)
+	llm.embd = append(llm.embd, token)
+
+	if token == C.llama_token_eos() {
+		return 0, io.EOF
+	}
+
+	return token, nil
+}
+
+func (llm *llama) Embedding(input string) ([]float64, error) {
+	if !llm.EmbeddingOnly {
+		return nil, errors.New("llama: embedding not enabled")
+	}
+
+	tokens := llm.Encode(input)
+	if tokens == nil {
+		return nil, errors.New("llama: tokenize embedding")
+	}
+
+	cTokens := make([]C.llama_token, len(tokens))
+	for i := range tokens {
+		cTokens[i] = C.llama_token(tokens[i])
+	}
+
+	retval := C.llama_eval(llm.ctx, unsafe.SliceData(cTokens), C.int(len(tokens)), 0, C.int(llm.NumThread))
+	if retval != 0 {
+		return nil, errors.New("llama: eval")
+	}
+
+	n := C.llama_n_embd(llm.ctx)
+	if n <= 0 {
+		return nil, errors.New("llama: no embeddings generated")
+	}
+	cEmbeddings := unsafe.Slice(C.llama_get_embeddings(llm.ctx), n)
+
+	embeddings := make([]float64, len(cEmbeddings))
+	for i, v := range cEmbeddings {
+		embeddings[i] = float64(v)
+	}
+	return embeddings, nil
+}

llm/llama.h

@@ -0,0 +1,495 @@
+/**
+ * llama.cpp - git 8183159cf3def112f6d1fe94815fce70e1bffa12
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Georgi Gerganov
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include "ggml.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#else
+#define LLAMA_MAX_DEVICES 1
+#endif // GGML_USE_CUBLAS
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#ifdef __GNUC__
+#    define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define DEPRECATED(func, hint) func
+#endif
+
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
+#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
+#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
+#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
+
+#define LLAMA_FILE_VERSION           3
+#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
+#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
+#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
+#define LLAMA_SESSION_VERSION        1
+
+#define LLAMA_DEFAULT_SEED           0xFFFFFFFF
+
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
+
+#ifndef LLAMA_DEFAULT_RMS_EPS
+#define LLAMA_DEFAULT_RMS_EPS 5e-6f
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_model;
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id; // token id
+        float logit;    // log-odds of the token
+        float p;        // probability of the token
+    } llama_token_data;
+
+    typedef struct llama_token_data_array {
+        llama_token_data * data;
+        size_t size;
+        bool sorted;
+    } llama_token_data_array;
+
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
+
+   struct llama_context_params {
+        uint32_t seed;         // RNG seed, -1 for random
+        int32_t  n_ctx;        // text context
+        int32_t  n_batch;      // prompt processing batch size
+        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
+        float    rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
+        int32_t  n_gpu_layers; // number of layers to store in VRAM
+        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float    rope_freq_base;  // RoPE base frequency
+        float    rope_freq_scale; // RoPE frequency scaling factor
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+    };
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32              = 0,
+        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+    };
+
+    // model quantization parameters
+    typedef struct llama_model_quantize_params {
+        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype   ftype;    // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+    } llama_model_quantize_params;
+
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
+    // performance timing information
+    struct llama_timings {
+        double t_start_ms;
+        double t_end_ms;
+        double t_load_ms;
+        double t_sample_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_sample;
+        int32_t n_p_eval;
+        int32_t n_eval;
+    };
+
+    LLAMA_API int llama_max_devices();
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
+
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
+    // TODO: not great API - very likely to change
+    // Initialize the llama + ggml backend
+    // If numa is true, use NUMA optimizations
+    // Call once at the start of the program
+    LLAMA_API void llama_backend_init(bool numa);
+    // Call once at the end of the program - currently only used for MPI
+    LLAMA_API void llama_backend_free();
+
+    LLAMA_API int64_t llama_time_us();
+
+    LLAMA_API struct llama_model * llama_load_model_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    LLAMA_API void llama_free_model(struct llama_model * model);
+
+    LLAMA_API struct llama_context * llama_new_context_with_model(
+                     struct llama_model * model,
+            struct llama_context_params   params);
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API DEPRECATED(struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params),
+            "please use llama_load_model_from_file combined with llama_new_context_with_model instead");
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+            const llama_model_quantize_params * params);
+
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads),
+            "please use llama_model_apply_lora_from_file instead");
+
+    LLAMA_API int llama_model_apply_lora_from_file(
+            const struct llama_model * model,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+
+    // Returns the maximum size in bytes of the state (rng, logits, embedding
+    // and kv_cache) - will often be smaller after compacting tokens
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
+
+    // Copies the state to the specified destination address.
+    // Destination needs to have allocated enough memory.
+    // Returns the number of bytes copied
+    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+
+    // Set the state reading from the specified address
+    // Returns the number of bytes read
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+
+    // Save/load session file
+    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
+    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Same as llama_eval, but use float matrix input directly.
+    LLAMA_API int llama_eval_embd(
+            struct llama_context * ctx,
+                     const float * embd,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Export a static computation graph for context of 511 and batch size of 1
+    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
+    //       parameters here to keep things simple
+    // IMPORTANT: do not use for anything else other than debugging and testing!
+    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_tokenize_with_model(
+        const struct llama_model * model,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
+
+    LLAMA_API int llama_n_vocab_from_model(const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_from_model  (const struct llama_model * model);
+    LLAMA_API int llama_n_embd_from_model (const struct llama_model * model);
+
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
+    LLAMA_API int llama_get_vocab_from_model(
+              const struct llama_model * model,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(
+            const struct llama_context * ctx,
+                           llama_token   token);
+
+    LLAMA_API const char * llama_token_to_str_with_model(
+              const struct llama_model * model,
+                           llama_token   token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
+    LLAMA_API llama_token llama_token_nl();   // next-line
+
+    // Grammar
+    //
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
+    // Sampling functions
+
+    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
+    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+
+    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+
+    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
+    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_classifier_free_guidance(
+              struct llama_context * ctx,
+            llama_token_data_array * candidates,
+              struct llama_context * guidance_ctx,
+                             float   scale);
+
+    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
+    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+
+    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+
+    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
+    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+
+    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
+    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+
+    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+
+    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+
+    /// @details Selects the token with the highest probability.
+    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Randomly selects a token from the candidates based on their probabilities.
+    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+
+    // Performance information
+    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
+#ifdef LLAMA_API_INTERNAL
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif
+
+#endif // LLAMA_H

llm/llama_darwin.go

@@ -0,0 +1,80 @@
+package llm
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"errors"
+	"io"
+	"log"
+	"os"
+	"path/filepath"
+)
+
+func init() {
+	if err := initBackend(); err != nil {
+		log.Printf("WARNING: GPU could not be initialized correctly: %v", err)
+		log.Printf("WARNING: falling back to CPU")
+	}
+}
+
+func initBackend() error {
+	exec, err := os.Executable()
+	if err != nil {
+		return err
+	}
+
+	exec, err = filepath.EvalSymlinks(exec)
+	if err != nil {
+		return err
+	}
+
+	metal := filepath.Join(filepath.Dir(exec), "ggml-metal.metal")
+	fi, err := os.Stat(metal)
+	if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return err
+	}
+
+	if fi != nil {
+		actual, err := os.Open(metal)
+		if err != nil {
+			return err
+		}
+
+		actualSum := sha256.New()
+		if _, err := io.Copy(actualSum, actual); err != nil {
+			return err
+		}
+
+		expect, err := fs.Open("ggml-metal.metal")
+		if err != nil {
+			return err
+		}
+
+		expectSum := sha256.New()
+		if _, err := io.Copy(expectSum, expect); err != nil {
+			return err
+		}
+
+		if bytes.Equal(actualSum.Sum(nil), expectSum.Sum(nil)) {
+			return nil
+		}
+	}
+
+	dst, err := os.Create(filepath.Join(filepath.Dir(exec), "ggml-metal.metal"))
+	if err != nil {
+		return err
+	}
+	defer dst.Close()
+
+	src, err := fs.Open("ggml-metal.metal")
+	if err != nil {
+		return err
+	}
+	defer src.Close()
+
+	if _, err := io.Copy(dst, src); err != nil {
+		return err
+	}
+
+	return nil
+}

llm/llm.go

@@ -0,0 +1,40 @@
+package llm
+
+import (
+	"fmt"
+	"os"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+type LLM interface {
+	Predict([]int, string, func(api.GenerateResponse)) error
+	Embedding(string) ([]float64, error)
+	Encode(string) []int
+	Decode(...int) string
+	SetOptions(api.Options)
+	Close()
+}
+
+func New(model string, opts api.Options) (LLM, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	f, err := os.Open(model)
+	if err != nil {
+		return nil, err
+	}
+
+	ggml, err := DecodeGGML(f, ModelFamilyLlama)
+	if err != nil {
+		return nil, err
+	}
+
+	switch ggml.ModelFamily {
+	case ModelFamilyLlama:
+		return newLlama(model, opts)
+	default:
+		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily)
+	}
+}

llm/update-llama-cpp.sh

@@ -0,0 +1,70 @@
+#!/bin/sh
+
+set -eu
+
+
+status() { echo >&2 ">>> $*"; }
+error() { status "ERROR $*"; }
+usage() {
+    echo "usage: $(basename $0) /path/to/repo"
+    exit 1
+}
+
+OUT=$(dirname $0)
+while getopts "hC:" OPTION; do
+    case $OPTION in
+        C) OUT=$OPTARG ;;
+        *) usage ;;
+    esac
+done
+
+shift $(( $OPTIND - 1 ))
+[ $# -eq 1 ] || usage
+
+status "updating source..."
+cp -a "$1"/*.{c,h,cpp,m,metal,cu} "$OUT"
+
+status "removing incompatible files..."
+rm -f "$OUT"/build-info.h
+
+SHA1=$(git -C $1 rev-parse @)
+
+LICENSE=$(mktemp)
+cleanup() {
+    rm -f $LICENSE
+}
+trap cleanup 0
+
+cat <<EOF | sed 's/ *$//' >$LICENSE
+/**
+ * llama.cpp - git $SHA1
+ *
+$(sed 's/^/ * /' <$1/LICENSE)
+ */
+
+EOF
+
+for IN in $OUT/*.{c,h,cpp,m,metal,cu}; do
+    TMP=$(mktemp)
+    status "updating license $IN"
+    cat $LICENSE $IN >$TMP
+    mv $TMP $IN
+done
+
+touchup() {
+    local CONSTRAINT=$1 && shift
+
+    for IN in $*; do
+        status "touching up $IN..."
+        TMP=$(mktemp)
+        {
+            echo "//go:build $CONSTRAINT"
+            echo
+        } | cat - $IN >$TMP
+        mv $TMP $IN
+    done
+}
+
+touchup darwin $OUT/ggml-metal.*
+touchup mpi $OUT/ggml-mpi.*
+touchup opencl $OUT/ggml-opencl.*

llm/utils.go

@@ -0,0 +1,15 @@
+package llm
+
+import (
+	"fmt"
+	"time"
+)
+
+func parseDurationMs(ms float64) time.Duration {
+	dur, err := time.ParseDuration(fmt.Sprintf("%fms", ms))
+	if err != nil {
+		panic(err)
+	}
+
+	return dur
+}