| 1 | #include "arg.h" |
| 2 | #include "common.h" |
| 3 | #include "log.h" |
| 4 | #include "llama.h" |
| 5 | #include "ggml.h" |
| 6 | |
| 7 | #include <cstdio> |
| 8 | #include <string> |
| 9 | #include <vector> |
| 10 | #include <numeric> |
| 11 | |
| 12 | /** |
| 13 | * This the arbitrary data which will be passed to each callback. |
| 14 | * Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor. |
| 15 | */ |
| 16 | struct callback_data { |
| 17 | std::vector<uint8_t> data; |
| 18 | }; |
| 19 | |
| 20 | static std::string ggml_ne_string(const ggml_tensor * t) { |
| 21 | std::string str; |
| 22 | for (int i = 0; i < GGML_MAX_DIMS; ++i) { |
| 23 | str += std::to_string(val: t->ne[i]); |
| 24 | if (i + 1 < GGML_MAX_DIMS) { |
| 25 | str += ", " ; |
| 26 | } |
| 27 | } |
| 28 | return str; |
| 29 | } |
| 30 | |
| 31 | static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) { |
| 32 | union { |
| 33 | float f; |
| 34 | uint32_t i; |
| 35 | } u; |
| 36 | u.i = (uint32_t)h.bits << 16; |
| 37 | return u.f; |
| 38 | } |
| 39 | |
| 40 | static float ggml_get_float_value(uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) { |
| 41 | size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0]; |
| 42 | float v; |
| 43 | if (type == GGML_TYPE_F16) { |
| 44 | v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]); |
| 45 | } else if (type == GGML_TYPE_F32) { |
| 46 | v = *(float *) &data[i]; |
| 47 | } else if (type == GGML_TYPE_I64) { |
| 48 | v = (float) *(int64_t *) &data[i]; |
| 49 | } else if (type == GGML_TYPE_I32) { |
| 50 | v = (float) *(int32_t *) &data[i]; |
| 51 | } else if (type == GGML_TYPE_I16) { |
| 52 | v = (float) *(int16_t *) &data[i]; |
| 53 | } else if (type == GGML_TYPE_I8) { |
| 54 | v = (float) *(int8_t *) &data[i]; |
| 55 | } else if (type == GGML_TYPE_BF16) { |
| 56 | v = ggml_compute_bf16_to_fp32(h: *(ggml_bf16_t *) &data[i]); |
| 57 | } else { |
| 58 | GGML_ABORT("fatal error" ); |
| 59 | } |
| 60 | return v; |
| 61 | } |
| 62 | |
| 63 | static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) { |
| 64 | GGML_ASSERT(n > 0); |
| 65 | float sum = 0; |
| 66 | for (int64_t i3 = 0; i3 < ne[3]; i3++) { |
| 67 | for (int64_t i2 = 0; i2 < ne[2]; i2++) { |
| 68 | for (int64_t i1 = 0; i1 < ne[1]; i1++) { |
| 69 | for (int64_t i0 = 0; i0 < ne[0]; i0++) { |
| 70 | const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3); |
| 71 | sum += v; |
| 72 | } |
| 73 | } |
| 74 | } |
| 75 | } |
| 76 | for (int64_t i3 = 0; i3 < ne[3]; i3++) { |
| 77 | LOG(" [\n" ); |
| 78 | for (int64_t i2 = 0; i2 < ne[2]; i2++) { |
| 79 | if (i2 == n && ne[2] > 2*n) { |
| 80 | LOG(" ..., \n" ); |
| 81 | i2 = ne[2] - n; |
| 82 | } |
| 83 | LOG(" [\n" ); |
| 84 | for (int64_t i1 = 0; i1 < ne[1]; i1++) { |
| 85 | if (i1 == n && ne[1] > 2*n) { |
| 86 | LOG(" ..., \n" ); |
| 87 | i1 = ne[1] - n; |
| 88 | } |
| 89 | LOG(" [" ); |
| 90 | for (int64_t i0 = 0; i0 < ne[0]; i0++) { |
| 91 | if (i0 == n && ne[0] > 2*n) { |
| 92 | LOG("..., " ); |
| 93 | i0 = ne[0] - n; |
| 94 | } |
| 95 | const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3); |
| 96 | LOG("%12.4f" , v); |
| 97 | if (i0 < ne[0] - 1) LOG(", " ); |
| 98 | } |
| 99 | LOG("],\n" ); |
| 100 | } |
| 101 | LOG(" ],\n" ); |
| 102 | } |
| 103 | LOG(" ]\n" ); |
| 104 | LOG(" sum = %f\n" , sum); |
| 105 | } |
| 106 | |
| 107 | // TODO: make this abort configurable/optional? |
| 108 | if (std::isnan(x: sum)) { |
| 109 | LOG_ERR("encountered NaN - aborting\n" ); |
| 110 | exit(status: 0); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | /** |
| 115 | * GGML operations callback during the graph execution. |
| 116 | * |
| 117 | * @param t current tensor |
| 118 | * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor |
| 119 | * if we return true, a follow-up call will be made with ask=false in which we can do the actual collection. |
| 120 | * see ggml_backend_sched_eval_callback |
| 121 | * @param user_data user data to pass at each call back |
| 122 | * @return true to receive data or continue the graph, false otherwise |
| 123 | */ |
| 124 | static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) { |
| 125 | auto * cb_data = (callback_data *) user_data; |
| 126 | |
| 127 | const struct ggml_tensor * src0 = t->src[0]; |
| 128 | const struct ggml_tensor * src1 = t->src[1]; |
| 129 | |
| 130 | if (ask) { |
| 131 | return true; // Always retrieve data |
| 132 | } |
| 133 | |
| 134 | char src1_str[128] = {0}; |
| 135 | if (src1) { |
| 136 | snprintf(s: src1_str, maxlen: sizeof(src1_str), format: "%s{%s}" , src1->name, ggml_ne_string(t: src1).c_str()); |
| 137 | } |
| 138 | |
| 139 | LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n" , __func__, |
| 140 | t->name, ggml_type_name(t->type), ggml_op_desc(t), |
| 141 | src0->name, ggml_ne_string(src0).c_str(), |
| 142 | src1 ? src1_str : "" , |
| 143 | ggml_ne_string(t).c_str()); |
| 144 | |
| 145 | |
| 146 | // copy the data from the GPU memory if needed |
| 147 | const bool is_host = ggml_backend_buffer_is_host(buffer: t->buffer); |
| 148 | |
| 149 | if (!is_host) { |
| 150 | auto n_bytes = ggml_nbytes(tensor: t); |
| 151 | cb_data->data.resize(new_size: n_bytes); |
| 152 | ggml_backend_tensor_get(tensor: t, data: cb_data->data.data(), offset: 0, size: n_bytes); |
| 153 | } |
| 154 | |
| 155 | if (!ggml_is_quantized(type: t->type)) { |
| 156 | uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data(); |
| 157 | ggml_print_tensor(data, type: t->type, ne: t->ne, nb: t->nb, n: 3); |
| 158 | } |
| 159 | |
| 160 | return true; |
| 161 | } |
| 162 | |
| 163 | static bool run(llama_context * ctx, const common_params & params) { |
| 164 | const llama_model * model = llama_get_model(ctx); |
| 165 | const llama_vocab * vocab = llama_model_get_vocab(model); |
| 166 | |
| 167 | const bool add_bos = llama_vocab_get_add_bos(vocab); |
| 168 | |
| 169 | std::vector<llama_token> tokens = common_tokenize(ctx, text: params.prompt, add_special: add_bos); |
| 170 | |
| 171 | if (tokens.empty()) { |
| 172 | LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n" , __func__); |
| 173 | return false; |
| 174 | } |
| 175 | |
| 176 | if (llama_decode(ctx, batch: llama_batch_get_one(tokens: tokens.data(), n_tokens: tokens.size()))) { |
| 177 | LOG_ERR("%s : failed to eval\n" , __func__); |
| 178 | return false; |
| 179 | } |
| 180 | |
| 181 | return true; |
| 182 | } |
| 183 | |
| 184 | int main(int argc, char ** argv) { |
| 185 | callback_data cb_data; |
| 186 | |
| 187 | common_params params; |
| 188 | |
| 189 | if (!common_params_parse(argc, argv, params, ex: LLAMA_EXAMPLE_COMMON)) { |
| 190 | return 1; |
| 191 | } |
| 192 | |
| 193 | common_init(); |
| 194 | |
| 195 | llama_backend_init(); |
| 196 | llama_numa_init(numa: params.numa); |
| 197 | |
| 198 | // pass the callback to the backend scheduler |
| 199 | // it will be executed for each node during the graph computation |
| 200 | params.cb_eval = ggml_debug; |
| 201 | params.cb_eval_user_data = &cb_data; |
| 202 | params.warmup = false; |
| 203 | |
| 204 | // init |
| 205 | common_init_result llama_init = common_init_from_params(params); |
| 206 | |
| 207 | llama_model * model = llama_init.model.get(); |
| 208 | llama_context * ctx = llama_init.context.get(); |
| 209 | |
| 210 | if (model == nullptr || ctx == nullptr) { |
| 211 | LOG_ERR("%s : failed to init\n" , __func__); |
| 212 | return 1; |
| 213 | } |
| 214 | |
| 215 | // print system information |
| 216 | { |
| 217 | LOG_INF("\n" ); |
| 218 | LOG_INF("%s\n" , common_params_get_system_info(params).c_str()); |
| 219 | LOG_INF("\n" ); |
| 220 | } |
| 221 | |
| 222 | bool OK = run(ctx, params); |
| 223 | if (!OK) { |
| 224 | return 1; |
| 225 | } |
| 226 | |
| 227 | LOG("\n" ); |
| 228 | llama_perf_context_print(ctx); |
| 229 | |
| 230 | llama_backend_free(); |
| 231 | |
| 232 | return 0; |
| 233 | } |
| 234 | |