llama_cpp 0.1.3 → 0.1.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: ed569e816938dfca79c345228cf592eb81827c04acfeae3c8e26e0635bbc518b
-  data.tar.gz: f4a899df0cf450370d7dc75e486a17617f1af0cbcacd9d9a8c7d3bde10016441
+  metadata.gz: f08992d10701b3ac0ab87c32d8a28d7f81101a4896e3300c461d7015f6486814
+  data.tar.gz: 38fd72fe9cdd596f7878ef902ddf8ec8e36954bbac50388fe8ef8437a93bfe29
 SHA512:
-  metadata.gz: 0f3d38eed6628e8d68efc741fe00024fb0c5199fb2e1a33d6f04d9299e1c59deb969e3eafe36190ade84522e70ddca50956fbee9b6406edc5d613f654889a83a
-  data.tar.gz: 0b1705a8d70564a59ad6472b03dc0241727766d4121e26a2e9c3c0d4725ddf2ccf65cb8f4a862688661ea9fa2b1c8858cd6e5e722821e6c2c30c91401475ef74
+  metadata.gz: e3d024db508be6cbe7e644c4a9295da97742b45f921c9c5d64a7f4b4eb6be624e79f4c63d39c226566dbb9676215ae3b986828095a185cb2069547a12cf651a0
+  data.tar.gz: d884334f2d77a7204f0bc96c037fa86ee6fdf3f2879c5c5bd721be336dc743a0034733fc566c114bc6f22e620e5d79ccd4c67b6ded7d929d1949315b31445701

data/CHANGELOG.md

@@ -1,25 +1,27 @@
-## [Unreleased]
+## [[0.1.4](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.3...v0.1.4)] - 2023-06-03
+
+- Bump bundled llama.cpp from master-66874d4 to master-ffb06a3.
 
 ## [[0.1.3](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.2...v0.1.3)] - 2023-05-27
 
-- Bump bundled llama.cpp from master-265db98 to master-66874d4
+- Bump bundled llama.cpp from master-265db98 to master-66874d4.
 
 ## [[0.1.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.1...v0.1.2)] - 2023-05-22
 
 **Breaking Changes**
 
-- Bump bundled llama.cpp from master-6986c78 to master-265db98
-  - bump LLAMA_FILE_VERSION to 3
+- Bump bundled llama.cpp from master-6986c78 to master-265db98.
+  - bump LLAMA_FILE_VERSION to 3.
 
 ## [[0.1.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.0...v0.1.1)] - 2023-05-21
 
-- Add load_session_file method to Context
-- Add save_session_file method to Context
+- Add load_session_file method to Context.
+- Add save_session_file method to Context.
 
 **Breaking Changes**
 
-- Bump bundled llama.cpp from master-173d0e6 to master-6986c78
-  - bump LLAMA_FILE_VERSION to 2
+- Bump bundled llama.cpp from master-173d0e6 to master-6986c78.
+  - bump LLAMA_FILE_VERSION to 2.
 
 ## [[0.1.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.0.7...v0.1.0)] - 2023-05-20
 

data/ext/llama_cpp/src/ggml-opencl.cpp

@@ -469,16 +469,11 @@ void ggml_cl_init(void) {
 
     size_t ext_str_size;
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
-    char* ext_buffer = (char*) malloc(sizeof(char) * ext_str_size);
+    char *ext_buffer = (char *)alloca(ext_str_size + 1);
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
+    ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
     // Check if ext_buffer contains cl_khr_fp16
-    for (size_t i = 0; i < ext_str_size - 12; i++) {
-        if (memcmp(ext_buffer + i, "cl_khr_fp16", 11) == 0) {
-            fp16_support = true;
-            break;
-        }
-    }
-    free(ext_buffer);
+    fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
     fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
 
     cl_context_properties properties[] = {
@@ -672,7 +667,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_CL) {
-        d_X = *(cl_mem*) src0->data;
+        d_X = (cl_mem) src0->data;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
     }
@@ -748,7 +743,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_CL) {
-        d_X = *(cl_mem*) src0->data;
+        d_X = (cl_mem) src0->data;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
     }
@@ -873,7 +868,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
             if (src0->backend == GGML_BACKEND_CPU) {
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, NULL));
             } else if (src0->backend == GGML_BACKEND_CL) {
-                d_Q = *(cl_mem*) src0->data;
+                d_Q = (cl_mem) src0->data;
             } else {
                 GGML_ASSERT(false);
             }
@@ -1016,14 +1011,13 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
 
     size_t q_size;
-    cl_mem* dst = (cl_mem*) malloc(sizeof(cl_mem));
-    *dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
+    cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
 
     // copy tensor to device
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             int i = i3*ne2 + i2;
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, *dst, i*ne0*ne1, tensor, i3, i2, NULL));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
         }
     }
 

data/ext/llama_cpp/src/ggml.c

@@ -186,10 +186,12 @@ typedef double ggml_float;
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>
 #else
+#if !defined(__riscv)
 #include <immintrin.h>
 #endif
 #endif
 #endif
+#endif
 
 #ifdef __F16C__
 
@@ -3494,7 +3496,7 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = {
 };
 static_assert(GGML_TYPE_COUNT == 13, "GGML_IS_QUANTIZED is outdated");
 
-static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
+static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "NONE",
 
     "DUP",
@@ -3749,6 +3751,9 @@ const char * ggml_type_name(enum ggml_type type) {
     return GGML_TYPE_NAME[type];
 }
 
+const char * ggml_op_name(enum ggml_op op) {
+    return GGML_OP_NAME[op];
+}
 
 size_t ggml_element_size(const struct ggml_tensor * tensor) {
     return GGML_TYPE_SIZE[tensor->type];
@@ -3805,6 +3810,10 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
     return wtype;
 }
 
+size_t ggml_tensor_overhead(void) {
+    return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE + 16;
+}
+
 static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
@@ -4017,6 +4026,18 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch)
     return result;
 }
 
+void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc) {
+    ctx->no_alloc = no_alloc;
+}
+
+void * ggml_get_mem_buffer(struct ggml_context * ctx) {
+    return ctx->mem_buffer;
+}
+
+size_t ggml_get_mem_size(struct ggml_context * ctx) {
+    return ctx->mem_size;
+}
+
 // IMPORTANT:
 // when creating "opt" tensors, always save and load the scratch buffer
 // this is an error prone process, but it is necessary to support inplace
@@ -4061,7 +4082,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
     if (ctx->scratch.data == NULL || data != NULL) {
-        size_needed += sizeof(struct ggml_tensor);
+        size_needed += GGML_TENSOR_SIZE;
 
         if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
             GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
@@ -4077,14 +4098,15 @@ struct ggml_tensor * ggml_new_tensor_impl(
         };
     } else {
         if (ctx->scratch.offs + size_needed > ctx->scratch.size) {
-            GGML_PRINT("%s: not enough space in the scratch memory\n", __func__);
+            GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
+                    __func__, ctx->scratch.offs + size_needed, ctx->scratch.size);
             assert(false);
             return NULL;
         }
 
-        if (cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE > ctx->mem_size) {
+        if (cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE > ctx->mem_size) {
             GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
-                    __func__, cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE, ctx->mem_size);
+                    __func__, cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE, ctx->mem_size);
             assert(false);
             return NULL;
         }
@@ -4093,7 +4115,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
 
         *obj_new = (struct ggml_object) {
             .offs = cur_end + GGML_OBJECT_SIZE,
-            .size = sizeof(struct ggml_tensor),
+            .size = GGML_TENSOR_SIZE,
             .next = NULL,
         };
 
@@ -4509,6 +4531,23 @@ struct ggml_tensor * ggml_view_tensor(
     return result;
 }
 
+struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name) {
+    struct ggml_object * obj = ctx->objects_begin;
+
+    char * const mem_buffer = ctx->mem_buffer;
+
+    while (obj != NULL) {
+        struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs);
+        if (strcmp(cur->name, name) == 0) {
+            return cur;
+        }
+
+        obj = obj->next;
+    }
+
+    return NULL;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 // ggml_dup
@@ -6303,7 +6342,7 @@ struct ggml_tensor * ggml_alibi(
 
     ggml_scratch_save(ctx);
 
-    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3);
 
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_head;
@@ -13792,11 +13831,19 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
         // reached a leaf node, not part of the gradient graph (e.g. a constant)
         GGML_ASSERT(cgraph->n_leafs < GGML_MAX_NODES);
 
+        if (strlen(node->name) == 0) {
+            snprintf(node->name, sizeof(node->name), "leaf_%d", cgraph->n_leafs);
+        }
+
         cgraph->leafs[cgraph->n_leafs] = node;
         cgraph->n_leafs++;
     } else {
         GGML_ASSERT(cgraph->n_nodes < GGML_MAX_NODES);
 
+        if (strlen(node->name) == 0) {
+            snprintf(node->name, sizeof(node->name), "node_%d", cgraph->n_nodes);
+        }
+
         cgraph->nodes[cgraph->n_nodes] = node;
         cgraph->grads[cgraph->n_nodes] = node->grad;
         cgraph->n_nodes++;
@@ -14510,6 +14557,481 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
     }
 }
 
+struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) {
+    for (int i = 0; i < cgraph->n_leafs; i++) {
+        struct ggml_tensor * leaf = cgraph->leafs[i];
+
+        if (strcmp(leaf->name, name) == 0) {
+            return leaf;
+        }
+    }
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        if (strcmp(node->name, name) == 0) {
+            return node;
+        }
+    }
+
+    return NULL;
+}
+
+static void ggml_graph_export_leaf(const struct ggml_tensor * tensor, FILE * fout) {
+    const int64_t * ne = tensor->ne;
+    const size_t  * nb = tensor->nb;
+
+    fprintf(fout, "%-6s %-12s %8d %8lld %8lld %8lld %8lld %16zu %16zu %16zu %16zu %16p %16s\n",
+            ggml_type_name(tensor->type),
+            ggml_op_name  (tensor->op),
+            tensor->n_dims,
+            ne[0], ne[1], ne[2], ne[3],
+            nb[0], nb[1], nb[2], nb[3],
+            tensor->data,
+            tensor->name);
+}
+
+static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char * arg, FILE * fout) {
+    const int64_t * ne = tensor->ne;
+    const size_t  * nb = tensor->nb;
+
+    fprintf(fout, "%-6s %-6s %-12s %8d %8lld %8lld %8lld %8lld %16zu %16zu %16zu %16zu %8d %16p %16s\n",
+            arg,
+            ggml_type_name(tensor->type),
+            ggml_op_name  (tensor->op),
+            tensor->n_dims,
+            ne[0], ne[1], ne[2], ne[3],
+            nb[0], nb[1], nb[2], nb[3],
+            tensor->n_tasks,
+            tensor->data,
+            tensor->name);
+}
+
+void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
+    assert(cgraph->work      == NULL);
+    assert(cgraph->work_size == 0);
+
+    uint64_t size_eval = 0;
+
+    // compute size of intermediate results
+    // TODO: does not take into account scratch buffers !!!!
+    for (int i = 0; i < cgraph->n_nodes; ++i) {
+        size_eval += ggml_nbytes(cgraph->nodes[i]);
+    }
+
+    // print
+    {
+        FILE * fout = stdout;
+
+        fprintf(fout, "\n");
+        fprintf(fout, "%-16s %8x\n",   "magic",   GGML_FILE_MAGIC);
+        fprintf(fout, "%-16s %8d\n",   "version", GGML_FILE_VERSION);
+        fprintf(fout, "%-16s %8d\n",   "leafs",   cgraph->n_leafs);
+        fprintf(fout, "%-16s %8d\n",   "nodes",   cgraph->n_nodes);
+        fprintf(fout, "%-16s %8llu\n", "eval",    size_eval);
+
+        // header
+        fprintf(fout, "\n");
+        fprintf(fout, "%-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %16s %16s\n",
+                "TYPE", "OP", "NDIMS", "NE0", "NE1", "NE2", "NE3", "NB0", "NB1", "NB2", "NB3", "DATA", "NAME");
+
+        for (int i = 0; i < cgraph->n_leafs; ++i) {
+            ggml_graph_export_leaf(cgraph->leafs[i], fout);
+
+            GGML_ASSERT(cgraph->leafs[i]->op   == GGML_OP_NONE);
+            GGML_ASSERT(cgraph->leafs[i]->src0 == NULL);
+            GGML_ASSERT(cgraph->leafs[i]->src1 == NULL);
+        }
+
+        // header
+        fprintf(fout, "\n");
+        fprintf(fout, "%-6s %-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %8s %16s %16s\n",
+                "ARG", "TYPE", "OP", "NDIMS", "NE0", "NE1", "NE2", "NE3", "NB0", "NB1", "NB2", "NB3", "NTASKS", "DATA", "NAME");
+
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            ggml_graph_export_node(cgraph->nodes[i], "DST", fout);
+
+            if (cgraph->nodes[i]->src0) {
+                ggml_graph_export_node(cgraph->nodes[i]->src0, "SRC0", fout);
+            }
+
+            if (cgraph->nodes[i]->src1) {
+                ggml_graph_export_node(cgraph->nodes[i]->src1, "SRC1", fout);
+            }
+
+            for (int j = 0; j < GGML_MAX_OPT; ++j) {
+                if (cgraph->nodes[i]->opt[j]) {
+                    ggml_graph_export_node(cgraph->nodes[i]->opt[j], "OPT", fout);
+                }
+            }
+
+            fprintf(fout, "\n");
+        }
+
+        fprintf(fout, "\n");
+    }
+
+    // write binary data
+    {
+        FILE * fout = fopen(fname, "wb");
+
+        if (!fout) {
+            fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
+            return;
+        }
+
+        // header
+        {
+            const uint32_t magic   = GGML_FILE_MAGIC;
+            const uint32_t version = GGML_FILE_VERSION;
+            const uint32_t n_leafs = cgraph->n_leafs;
+            const uint32_t nodes   = cgraph->n_nodes;
+
+            fwrite(&magic,     sizeof(uint32_t), 1, fout);
+            fwrite(&version,   sizeof(uint32_t), 1, fout);
+            fwrite(&n_leafs,   sizeof(uint32_t), 1, fout);
+            fwrite(&nodes,     sizeof(uint32_t), 1, fout);
+            fwrite(&size_eval, sizeof(uint64_t), 1, fout);
+        }
+
+        // leafs
+        {
+            for (int i = 0; i < cgraph->n_leafs; ++i) {
+                const struct ggml_tensor * tensor = cgraph->leafs[i];
+
+                const uint32_t type   = tensor->type;
+                const uint32_t op     = tensor->op;
+                const uint32_t n_dims = tensor->n_dims;
+
+                fwrite(&type,   sizeof(uint32_t), 1, fout);
+                fwrite(&op,     sizeof(uint32_t), 1, fout);
+                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    const uint64_t ne = tensor->ne[j];
+                    const uint64_t nb = tensor->nb[j];
+
+                    fwrite(&ne, sizeof(uint64_t), 1, fout);
+                    fwrite(&nb, sizeof(uint64_t), 1, fout);
+                }
+
+                // store the pointer address
+                {
+                    const uint64_t ptr = (uint64_t) tensor->data;
+
+                    fwrite(&ptr, sizeof(uint64_t), 1, fout);
+                }
+
+                fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout);
+
+                // dump the data
+                // TODO: pad this to 32 byte boundary
+                {
+                    const size_t size = ggml_nbytes(tensor);
+
+                    fwrite(tensor->data, sizeof(char), size, fout);
+                }
+            }
+        }
+
+        // nodes
+        {
+            for (int i = 0; i < cgraph->n_nodes; ++i) {
+                const struct ggml_tensor * tensor = cgraph->nodes[i];
+
+                const uint32_t type   = tensor->type;
+                const uint32_t op     = tensor->op;
+                const uint32_t n_dims = tensor->n_dims;
+
+                fwrite(&type,   sizeof(uint32_t), 1, fout);
+                fwrite(&op,     sizeof(uint32_t), 1, fout);
+                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    const uint64_t ne = tensor->ne[j];
+                    const uint64_t nb = tensor->nb[j];
+
+                    fwrite(&ne, sizeof(uint64_t), 1, fout);
+                    fwrite(&nb, sizeof(uint64_t), 1, fout);
+                }
+
+                // store the pointer address
+                {
+                    const uint64_t ptr = (uint64_t) tensor->data;
+
+                    fwrite(&ptr, sizeof(uint64_t), 1, fout);
+                }
+
+                fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout);
+
+                // output the op arguments
+                {
+                    struct ggml_tensor * args[2 + GGML_MAX_OPT] = { NULL };
+
+                    args[0] = tensor->src0;
+                    args[1] = tensor->src1;
+
+                    for (int j = 0; j < GGML_MAX_OPT; ++j) {
+                        args[2 + j] = tensor->opt[j];
+                    }
+
+                    for (int j = 0; j < 2 + GGML_MAX_OPT; ++j) {
+                        if (args[j]) {
+                            int32_t idx = -1;
+
+                            // check if leaf
+                            {
+                                for (int k = 0; k < cgraph->n_leafs; ++k) {
+                                    if (args[j] == cgraph->leafs[k]) {
+                                        idx = k;
+                                        break;
+                                    }
+                                }
+                            }
+
+                            // check if node
+                            if (idx == -1) {
+                                for (int k = 0; k < cgraph->n_nodes; ++k) {
+                                    if (args[j] == cgraph->nodes[k]) {
+                                        idx = GGML_MAX_NODES + k;
+                                        break;
+                                    }
+                                }
+                            }
+
+                            if (idx == -1) {
+                                fprintf(stderr, "%s: failed to find tensor, arg = %d, node = %d\n", __func__, j, i);
+                                return;
+                            }
+
+                            fwrite(&idx, sizeof(int32_t), 1, fout);
+                        } else {
+                            const int32_t nul = -1;
+
+                            fwrite(&nul, sizeof(int32_t), 1, fout);
+                        }
+                    }
+                }
+            }
+        }
+
+        fclose(fout);
+    }
+}
+
+struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) {
+    assert(*ctx_data == NULL);
+    assert(*ctx_eval == NULL);
+
+    struct ggml_cgraph result = { 0 };
+
+    struct ggml_tensor * data = NULL;
+
+    // read file into data
+    {
+        FILE * fin = fopen(fname, "rb");
+
+        if (!fin) {
+            fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
+            return result;
+        }
+
+        size_t fsize = 0;
+
+        fseek(fin, 0, SEEK_END);
+        fsize = ftell(fin);
+        fseek(fin, 0, SEEK_SET);
+
+        // create the data context
+        {
+            const size_t overhead = 1*ggml_tensor_overhead();
+
+            struct ggml_init_params params = {
+                .mem_size   = fsize + overhead,
+                .mem_buffer = NULL,
+                .no_alloc   = false,
+            };
+
+            *ctx_data = ggml_init(params);
+
+            if (!*ctx_data) {
+                fprintf(stderr, "%s: failed to create ggml context\n", __func__);
+                return result;
+            }
+        }
+
+        data = ggml_new_tensor_1d(*ctx_data, GGML_TYPE_I8, fsize);
+
+        fread(data->data, sizeof(char), fsize, fin);
+
+        fclose(fin);
+    }
+
+    // populate result
+    {
+        char * ptr = (char *) data->data;
+
+        const uint32_t magic = *(const uint32_t *) ptr; ptr += sizeof(magic);
+
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid magic number, got %08x\n", __func__, magic);
+            return result;
+        }
+
+        const uint32_t version = *(const uint32_t *) ptr; ptr += sizeof(version);
+
+        if (version != GGML_FILE_VERSION) {
+            fprintf(stderr, "%s: invalid version number\n", __func__);
+            return result;
+        }
+
+        const uint32_t n_leafs   = *(const uint32_t *) ptr; ptr += sizeof(n_leafs);
+        const uint32_t n_nodes   = *(const uint32_t *) ptr; ptr += sizeof(n_nodes);
+        const uint64_t size_eval = *(const uint64_t *) ptr; ptr += sizeof(size_eval);
+
+        result.n_leafs = n_leafs;
+        result.n_nodes = n_nodes;
+
+        // create the data context
+        {
+            const size_t overhead = (n_leafs + n_nodes)*ggml_tensor_overhead();
+
+            struct ggml_init_params params = {
+                .mem_size   = size_eval + overhead,
+                .mem_buffer = NULL,
+                .no_alloc   = true,
+            };
+
+            *ctx_eval = ggml_init(params);
+
+            if (!*ctx_eval) {
+                fprintf(stderr, "%s: failed to create ggml context\n", __func__);
+                return result;
+            }
+        }
+
+        // leafs
+        {
+            uint32_t type;
+            uint32_t op;
+            uint32_t n_dims;
+
+            for (uint32_t i = 0; i < n_leafs; ++i) {
+                type   = *(const uint32_t *) ptr; ptr += sizeof(type);
+                op     = *(const uint32_t *) ptr; ptr += sizeof(op);
+                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
+
+                int64_t ne[GGML_MAX_DIMS];
+                size_t  nb[GGML_MAX_DIMS];
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    uint64_t ne_cur;
+                    uint64_t nb_cur;
+
+                    ne_cur = *(const uint64_t *) ptr; ptr += sizeof(ne_cur);
+                    nb_cur = *(const uint64_t *) ptr; ptr += sizeof(nb_cur);
+
+                    ne[j] = ne_cur;
+                    nb[j] = nb_cur;
+                }
+
+                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+
+                tensor->op = (enum ggml_op) op;
+
+                uint64_t ptr_cur = *(const uint64_t *) ptr; ptr += sizeof(ptr_cur);
+
+                memcpy(tensor->name, ptr, GGML_MAX_NAME); ptr += GGML_MAX_NAME;
+
+                tensor->data = (void *) ptr;
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    tensor->nb[j] = nb[j];
+                }
+
+                result.leafs[i] = tensor;
+
+                ptr += ggml_nbytes(tensor);
+
+                fprintf(stderr, "%s: loaded leaf %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+            }
+        }
+
+        ggml_set_no_alloc(*ctx_eval, false);
+
+        // nodes
+        {
+            uint32_t type;
+            uint32_t op;
+            uint32_t n_dims;
+
+            for (uint32_t i = 0; i < n_nodes; ++i) {
+                type   = *(const uint32_t *) ptr; ptr += sizeof(type);
+                op     = *(const uint32_t *) ptr; ptr += sizeof(op);
+                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
+
+                int64_t ne[GGML_MAX_DIMS];
+                size_t  nb[GGML_MAX_DIMS];
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    uint64_t ne_cur;
+                    uint64_t nb_cur;
+
+                    ne_cur = *(const uint64_t *) ptr; ptr += sizeof(ne_cur);
+                    nb_cur = *(const uint64_t *) ptr; ptr += sizeof(nb_cur);
+
+                    ne[j] = ne_cur;
+                    nb[j] = nb_cur;
+                }
+
+                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+
+                tensor->op = (enum ggml_op) op;
+
+                uint64_t ptr_cur = *(const uint64_t *) ptr; ptr += sizeof(ptr_cur);
+
+                memcpy(tensor->name, ptr, GGML_MAX_NAME); ptr += GGML_MAX_NAME;
+
+                for (int j = 0; j < GGML_MAX_DIMS; ++j) {
+                    tensor->nb[j] = nb[j];
+                }
+
+                // parse args
+                {
+                    struct ggml_tensor ** args[2 + GGML_MAX_OPT] = {
+                        &tensor->src0,
+                        &tensor->src1,
+                    };
+
+                    for (int j = 0; j < GGML_MAX_OPT; ++j) {
+                        args[2 + j] = &tensor->opt[j];
+                    }
+
+                    for (int j = 0; j < 2 + GGML_MAX_OPT; ++j) {
+                        const int32_t arg_idx = *(const int32_t *) ptr; ptr += sizeof(arg_idx);
+
+                        if (arg_idx == -1) {
+                            continue;
+                        }
+
+                        if (arg_idx < GGML_MAX_NODES) {
+                            *args[j] = result.leafs[arg_idx];
+                        } else {
+                            *args[j] = result.nodes[arg_idx - GGML_MAX_NODES];
+                        }
+                    }
+                }
+
+                result.nodes[i] = tensor;
+
+                fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+            }
+        }
+    }
+
+    return result;
+}
+
 void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     int64_t perf_total_per_op_us[GGML_OP_COUNT] = {0};
 
@@ -14527,7 +15049,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
         GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
-                GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
+                GGML_OP_NAME[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
                 (double) node->perf_cycles  / (double) ggml_cycles_per_ms(),
                 (double) node->perf_cycles  / (double) ggml_cycles_per_ms() / (double) node->perf_runs,
                 (double) node->perf_time_us / 1000.0,
@@ -14541,7 +15063,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
         GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n",
                 i,
                 node->ne[0], node->ne[1],
-                GGML_OP_LABEL[node->op]);
+                GGML_OP_NAME[node->op]);
     }
 
     for (int i = 0; i < GGML_OP_COUNT; i++) {
@@ -14549,7 +15071,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
             continue;
         }
 
-        GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_LABEL[i], (double) perf_total_per_op_us[i] / 1000.0);
+        GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_NAME[i], (double) perf_total_per_op_us[i] / 1000.0);
     }
 
     GGML_PRINT("========================================\n");

data/ext/llama_cpp/src/ggml.h

@@ -198,6 +198,7 @@
 #define GGML_MAX_PARAMS        256
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_OPT           4
+#define GGML_MAX_NAME          32
 #define GGML_DEFAULT_N_THREADS 4
 
 #define GGML_ASSERT(x) \
@@ -372,11 +373,13 @@ extern "C" {
 
         void * data;
 
-        char name[32];
+        char name[GGML_MAX_NAME];
 
         char padding[16];
     };
 
+    static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -429,6 +432,7 @@ extern "C" {
     GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
+    GGML_API const char * ggml_op_name  (enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -437,6 +441,9 @@ extern "C" {
     // TODO: temporary until model loading of ggml examples is refactored
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
+    // use this to compute the memory overhead of a tensor
+    GGML_API size_t ggml_tensor_overhead(void);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -444,7 +451,11 @@ extern "C" {
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
+
+    GGML_API void *  ggml_get_mem_buffer(struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size  (struct ggml_context * ctx);
 
     GGML_API struct ggml_tensor * ggml_new_tensor(
             struct ggml_context * ctx,
@@ -484,6 +495,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
 
+    GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
+
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
     GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
@@ -970,6 +983,11 @@ extern "C" {
     GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
     GGML_API void ggml_graph_reset  (struct ggml_cgraph * cgraph);
 
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+
+    GGML_API void               ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 

data/ext/llama_cpp/src/llama.cpp

@@ -42,6 +42,7 @@
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_3B,
     MODEL_7B,
     MODEL_13B,
     MODEL_30B,
@@ -58,6 +59,7 @@ static const size_t MB = 1024*1024;
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    128ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
@@ -69,6 +71,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    128ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
@@ -81,6 +84,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
 static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    682ull * MB },
         { MODEL_7B,   1026ull * MB },
         { MODEL_13B,  1608ull * MB },
         { MODEL_30B,  3124ull * MB },
@@ -94,6 +98,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
 static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,   512ull * MB },
         { MODEL_7B,   768ull * MB },
         { MODEL_13B, 1024ull * MB },
         { MODEL_30B, 1280ull * MB },
@@ -899,6 +904,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
 
 static const char *llama_model_type_name(e_model type) {
     switch (type) {
+        case MODEL_3B: return "3B";
         case MODEL_7B: return "7B";
         case MODEL_13B: return "13B";
         case MODEL_30B: return "30B";
@@ -932,6 +938,7 @@ static void llama_model_load_internal(
 
     {
         switch (hparams.n_layer) {
+            case 26: model.type = e_model::MODEL_3B; break;
             case 32: model.type = e_model::MODEL_7B; break;
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;

data/ext/llama_cpp/src/llama.h

@@ -31,6 +31,11 @@
 #define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.1.3'
+  VERSION = '0.1.4'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'master-66874d4'
+  LLAMA_CPP_VERSION = 'master-ffb06a3'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.1.3
+  version: 0.1.4
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-05-26 00:00:00.000000000 Z
+date: 2023-06-03 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email: