diff --git "a/examples/talk-llama/llama.cpp" "b/examples/talk-llama/llama.cpp" --- "a/examples/talk-llama/llama.cpp" +++ "b/examples/talk-llama/llama.cpp" @@ -1,11 +1,12 @@ #define LLAMA_API_INTERNAL +//#define LLAMA_GGML_BACKEND_CUDA_TEST // for testing only - enables ggml-cuda through ggml-backend, disables partial offloading #include "llama.h" #include "unicode.h" #include "ggml.h" - #include "ggml-alloc.h" +#include "ggml-backend.h" #ifdef GGML_USE_CUBLAS # include "ggml-cuda.h" @@ -32,6 +33,7 @@ #include #if defined(_POSIX_MAPPED_FILES) #include + #include #endif #if defined(_POSIX_MEMLOCK_RANGE) #include @@ -46,7 +48,6 @@ #endif #include #include - #include // for _fseeki64 #endif #include @@ -75,6 +76,7 @@ #include #include #include +#include #include #if defined(_MSC_VER) @@ -91,6 +93,9 @@ #define LLAMA_ATTRIBUTE_FORMAT(...) #endif +#define LLAMA_MAX_NODES 8192 +#define LLAMA_MAX_EXPERTS 8 + // // logging // @@ -190,6 +195,9 @@ enum llm_arch { LLM_ARCH_PERSIMMON, LLM_ARCH_REFACT, LLM_ARCH_BLOOM, + LLM_ARCH_STABLELM, + LLM_ARCH_QWEN, + LLM_ARCH_PHI2, LLM_ARCH_UNKNOWN, }; @@ -205,6 +213,9 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_PERSIMMON, "persimmon" }, { LLM_ARCH_REFACT, "refact" }, { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, + { LLM_ARCH_QWEN, "qwen" }, + { LLM_ARCH_PHI2, "phi2" }, }; enum llm_kv { @@ -225,6 +236,8 @@ enum llm_kv { LLM_KV_FEED_FORWARD_LENGTH, LLM_KV_USE_PARALLEL_RESIDUAL, LLM_KV_TENSOR_DATA_LAYOUT, + LLM_KV_EXPERT_COUNT, + LLM_KV_EXPERT_USED_COUNT, LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT_KV, @@ -251,6 +264,8 @@ enum llm_kv { LLM_KV_TOKENIZER_UNK_ID, LLM_KV_TOKENIZER_SEP_ID, LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_ADD_BOS, + LLM_KV_TOKENIZER_ADD_EOS, LLM_KV_TOKENIZER_HF_JSON, LLM_KV_TOKENIZER_RWKV, }; @@ -273,6 +288,8 @@ static std::map LLM_KV_NAMES = { { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, + { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, + { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, @@ -299,6 +316,8 @@ static std::map LLM_KV_NAMES = { { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, + { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, + { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, }; @@ -328,10 +347,14 @@ enum llm_tensor { LLM_TENSOR_ATTN_NORM, LLM_TENSOR_ATTN_NORM_2, LLM_TENSOR_ATTN_ROT_EMBD, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_NORM, LLM_TENSOR_FFN_GATE, LLM_TENSOR_FFN_DOWN, LLM_TENSOR_FFN_UP, - LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_DOWN_EXP, + LLM_TENSOR_FFN_GATE_EXP, + LLM_TENSOR_FFN_UP_EXP, LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_K_NORM, }; @@ -350,10 +373,14 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, }, }, { @@ -493,6 +520,54 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, }, }, + { + LLM_ARCH_STABLELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_QWEN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_PHI2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { LLM_ARCH_UNKNOWN, { @@ -540,27 +615,16 @@ struct LLM_TN { std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const { return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid) + "." + suffix; } + + std::string operator()(llm_tensor tensor, const std::string & suffix, int bid, int xid) const { + return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid, xid) + "." + suffix; + } }; // // gguf helpers // -#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \ -do { \ - const std::string skey(key); \ - const int kid = gguf_find_key(ctx, skey.c_str()); \ - if (kid >= 0) { \ - enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ - if (ktype != (type)) { \ - throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \ - } \ - (dst) = func(ctx, kid); \ - } else if (req) { \ - throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \ - } \ -} while (0) - static std::map LLAMA_ROPE_SCALING_TYPES = { { LLAMA_ROPE_SCALING_NONE, "none" }, { LLAMA_ROPE_SCALING_LINEAR, "linear" }, @@ -577,6 +641,60 @@ static int8_t llama_rope_scaling_type_from_string(const std::string & name) { return LLAMA_ROPE_SCALING_UNSPECIFIED; } +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + +static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + // // ggml helpers // @@ -596,20 +714,6 @@ static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * // llama helpers // -#ifdef GGML_USE_CUBLAS -# define llama_host_malloc(n) ggml_cuda_host_malloc(n) -# define llama_host_free(data) ggml_cuda_host_free(data) -#elif GGML_USE_METAL -# define llama_host_malloc(n) ggml_metal_host_malloc(n) -# define llama_host_free(data) ggml_metal_host_free(data) -#elif GGML_USE_CPU_HBM -# define llama_host_malloc(n) hbw_malloc(n) -# define llama_host_free(data) if (data != NULL) hbw_free(data) -#else -# define llama_host_malloc(n) malloc(n) -# define llama_host_free(data) free(data) -#endif - #if defined(_WIN32) static std::string llama_format_win_err(DWORD err) { LPSTR buf; @@ -624,40 +728,10 @@ static std::string llama_format_win_err(DWORD err) { } #endif -struct llama_buffer { - void * data = NULL; - size_t size = 0; - - // fallback to malloc / free - // useful in cases where CUDA can try to allocate PINNED memory - bool fallback = false; - - void resize(size_t n) { - llama_host_free(data); - - data = llama_host_malloc(n); - if (!data) { - fallback = true; - data = malloc(n); - } else { - fallback = false; - } - - GGML_ASSERT(data); - size = n; - } - - ~llama_buffer() { - if (data) { - if (fallback) { // NOLINT - free(data); - } else { - llama_host_free(data); - } - } - - data = NULL; - } +template +struct no_init { + T value; + no_init() { /* do nothing */ } }; struct llama_file { @@ -745,6 +819,9 @@ struct llama_mmap { #ifdef _POSIX_MAPPED_FILES static constexpr bool SUPPORTED = true; + // list of mapped fragments (first_offset, last_offset) + std::vector> mapped_fragments; + 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); @@ -752,17 +829,22 @@ struct llama_mmap { // prefetch/readahead impairs performance on NUMA systems if (numa) { prefetch = 0; } #ifdef __linux__ + // advise the kernel to read the file sequentially (increases readahead) + if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) { + LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n", + strerror(errno)); + } if (prefetch) { flags |= MAP_POPULATE; } #endif addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); - if (addr == MAP_FAILED) { + if (addr == MAP_FAILED) { // NOLINT throw std::runtime_error(format("mmap failed: %s", strerror(errno))); } if (prefetch > 0) { - // Advise the kernel to preload the mapped memory + // advise the kernel to preload the mapped memory if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) { - fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n", + LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n", strerror(errno)); } } @@ -770,14 +852,81 @@ struct llama_mmap { // advise the kernel not to use readahead // (because the next page might not belong on the same node) if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) { - fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n", + LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n", strerror(errno)); } } + + // initialize list of mapped_fragments + mapped_fragments.emplace_back(0, file->size); + } + + static void align_range(size_t * first, size_t * last, size_t page_size) { + // align first to the next page + size_t offset_in_page = *first & (page_size - 1); + size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page; + *first += offset_to_page; + + // align last to the previous page + *last = *last & ~(page_size - 1); + + if (*last <= *first) { + *last = *first; + } + } + + // partially unmap the file in the range [first, last) + void unmap_fragment(size_t first, size_t last) { + // note: this function must not be called multiple times with overlapping ranges + // otherwise, there is a risk of invalidating addresses that have been repurposed for other mappings + int page_size = sysconf(_SC_PAGESIZE); + align_range(&first, &last, page_size); + size_t len = last - first; + + if (len == 0) { + return; + } + + GGML_ASSERT(first % page_size == 0); + GGML_ASSERT(last % page_size == 0); + GGML_ASSERT(last > first); + + void * next_page_start = (uint8_t *) addr + first; + + // unmap the range + if (munmap(next_page_start, len)) { + LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno)); + } + + // update the list of mapped fragments to avoid unmapping the same range again in the destructor + std::vector> new_mapped_fragments; + for (const auto & frag : mapped_fragments) { + if (frag.first < first && frag.second > last) { + // the range is in the middle of the fragment, split it + new_mapped_fragments.emplace_back(frag.first, first); + new_mapped_fragments.emplace_back(last, frag.second); + } else if (frag.first < first && frag.second > first) { + // the range starts in the middle of the fragment + new_mapped_fragments.emplace_back(frag.first, first); + } else if (frag.first < last && frag.second > last) { + // the range ends in the middle of the fragment + new_mapped_fragments.emplace_back(last, frag.second); + } else if (frag.first >= first && frag.second <= last) { + // the range covers the entire fragment + } else { + // the range is outside the fragment + new_mapped_fragments.push_back(frag); + } + } + mapped_fragments = std::move(new_mapped_fragments); } ~llama_mmap() { - munmap(addr, size); + for (const auto & frag : mapped_fragments) { + if (munmap((char *) addr + frag.first, frag.second - frag.first)) { + LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno)); + } + } } #elif defined(_WIN32) static constexpr bool SUPPORTED = true; @@ -825,6 +974,12 @@ struct llama_mmap { } } + void unmap_fragment(size_t first, size_t last) { + // not supported + GGML_UNUSED(first); + GGML_UNUSED(last); + } + ~llama_mmap() { if (!UnmapViewOfFile(addr)) { fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n", @@ -841,6 +996,13 @@ struct llama_mmap { throw std::runtime_error(std::string("mmap not supported")); } + + void unmap(size_t offset, size_t len) { + (void) offset; + (void) len; + + throw std::runtime_error(std::string("mmap not supported")); + } #endif }; @@ -1014,11 +1176,37 @@ static std::string llama_token_to_piece(const struct llama_context * ctx, llama_ return std::string(result.data(), result.size()); } +static ggml_backend_buffer_type_t llama_default_buffer_type(int n_gpu_layers) { +#ifdef GGML_USE_METAL + if (n_gpu_layers > 0) { + return ggml_backend_metal_buffer_type(); + } +#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (n_gpu_layers > 0) { + return ggml_backend_cuda_buffer_type(0); + } +#elif defined(GGML_USE_CUBLAS) + return ggml_backend_cuda_host_buffer_type(); +#elif defined(GGML_USE_CPU_HBM) + return ggml_backend_cpu_hbm_buffer_type(); +#endif + + return ggml_backend_cpu_buffer_type(); + + GGML_UNUSED(n_gpu_layers); +} + // // globals // struct llama_state { + llama_state() { +#ifdef GGML_USE_METAL + ggml_metal_log_set_callback(log_callback, log_callback_user_data); +#endif + } + // We save the log callback globally ggml_log_callback log_callback = llama_log_callback_default; void * log_callback_user_data = nullptr; @@ -1042,9 +1230,9 @@ enum e_model { MODEL_70B, }; -static const size_t kB = 1024; -static const size_t MB = 1024*kB; -static const size_t GB = 1024*MB; +static const size_t kiB = 1024; +static const size_t MiB = 1024*kiB; +static const size_t GiB = 1024*MiB; struct llama_hparams { bool vocab_only; @@ -1056,6 +1244,8 @@ struct llama_hparams { uint32_t n_layer; uint32_t n_rot; uint32_t n_ff; + uint32_t n_expert = 0; + uint32_t n_expert_used = 0; float f_norm_eps; float f_norm_rms_eps; @@ -1070,15 +1260,18 @@ struct llama_hparams { float f_max_alibi_bias; bool operator!=(const llama_hparams & other) const { - if (this->vocab_only != other.vocab_only) return true; - if (this->n_vocab != other.n_vocab) return true; - if (this->n_ctx_train != other.n_ctx_train) return true; - if (this->n_embd != other.n_embd) return true; - if (this->n_head != other.n_head) return true; - if (this->n_head_kv != other.n_head_kv) return true; - if (this->n_layer != other.n_layer) return true; - if (this->n_rot != other.n_rot) return true; - if (this->n_ff != other.n_ff) return true; + if (this->vocab_only != other.vocab_only) return true; + if (this->n_vocab != other.n_vocab) return true; + if (this->n_ctx_train != other.n_ctx_train) return true; + if (this->n_embd != other.n_embd) return true; + if (this->n_head != other.n_head) return true; + if (this->n_head_kv != other.n_head_kv) return true; + if (this->n_layer != other.n_layer) return true; + if (this->n_rot != other.n_rot) return true; + if (this->n_ff != other.n_ff) return true; + if (this->n_expert != other.n_expert) return true; + if (this->n_expert_used != other.n_expert_used) return true; + if (this->rope_finetuned != other.rope_finetuned) return true; if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; @@ -1123,6 +1316,7 @@ struct llama_cparams { float yarn_beta_slow; bool mul_mat_q; + bool offload_kqv; }; struct llama_layer { @@ -1144,6 +1338,9 @@ struct llama_layer { struct ggml_tensor * wqkv; // attention bias + struct ggml_tensor * bq; + struct ggml_tensor * bk; + struct ggml_tensor * bv; struct ggml_tensor * bo; struct ggml_tensor * bqkv; @@ -1156,6 +1353,12 @@ struct llama_layer { struct ggml_tensor * ffn_down; // w2 struct ggml_tensor * ffn_up; // w3 + // ff MoE + struct ggml_tensor * ffn_gate_inp; + struct ggml_tensor * ffn_gate_exp[LLAMA_MAX_EXPERTS]; + struct ggml_tensor * ffn_down_exp[LLAMA_MAX_EXPERTS]; + struct ggml_tensor * ffn_up_exp [LLAMA_MAX_EXPERTS]; + // ff bias struct ggml_tensor * ffn_down_b; // b2 struct ggml_tensor * ffn_up_b; // b3 @@ -1181,28 +1384,34 @@ struct llama_kv_cache { // cannot be freely changed after a slot has been allocated. uint32_t head = 0; uint32_t size = 0; + uint32_t used = 0; // used cells (i.e. at least one seq_id) // computed before each graph build uint32_t n = 0; std::vector cells; - struct ggml_tensor * k = NULL; - struct ggml_tensor * v = NULL; + std::vector k_l; // per layer + std::vector v_l; struct ggml_context * ctx = NULL; - llama_buffer buf; + ggml_backend_buffer_t buf = NULL; ~llama_kv_cache() { +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (ggml_cublas_loaded()) { + for (size_t i = 0; i < k_l.size(); ++i) { + ggml_cuda_free_data(k_l[i]); + ggml_cuda_free_data(v_l[i]); + } + } +#endif if (ctx) { ggml_free(ctx); } -#ifdef GGML_USE_CUBLAS - ggml_cuda_free_data(k); - ggml_cuda_free_data(v); -#endif // GGML_USE_CUBLAS + ggml_backend_buffer_free(buf); } }; @@ -1233,17 +1442,20 @@ struct llama_vocab { id special_sep_id = -1; id special_pad_id = -1; + int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add. + int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add. + id linefeed_id = 13; id special_prefix_id = 32007; id special_middle_id = 32009; id special_suffix_id = 32008; id special_eot_id = 32010; - int find_bpe_rank(std::string token_left, std::string token_right) const { - GGML_ASSERT(token_left.find(" ") == std::string::npos); - GGML_ASSERT(token_left.find("\n") == std::string::npos); - GGML_ASSERT(token_right.find(" ") == std::string::npos); - GGML_ASSERT(token_right.find("\n") == std::string::npos); + int find_bpe_rank(const std::string & token_left, const std::string & token_right) const { + GGML_ASSERT(token_left.find(' ') == std::string::npos); + GGML_ASSERT(token_left.find('\n') == std::string::npos); + GGML_ASSERT(token_right.find(' ') == std::string::npos); + GGML_ASSERT(token_right.find('\n') == std::string::npos); auto it = bpe_ranks.find(std::make_pair(token_left, token_right)); if (it == bpe_ranks.end()) { @@ -1272,16 +1484,20 @@ struct llama_model { struct ggml_tensor * output_norm; struct ggml_tensor * output_norm_b; struct ggml_tensor * output; + struct ggml_tensor * output_b; std::vector layers; int n_gpu_layers; + // gguf metadata + std::unordered_map gguf_kv; + // context struct ggml_context * ctx = NULL; // the model memory buffer - llama_buffer buf; + ggml_backend_buffer_t buf = NULL; // model memory mapped file std::unique_ptr mapping; @@ -1297,38 +1513,40 @@ struct llama_model { int64_t t_start_us = 0; ~llama_model() { - if (ctx) { - ggml_free(ctx); +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (ggml_cublas_loaded()) { + for (size_t i = 0; i < tensors_by_name.size(); ++i) { + ggml_cuda_free_data(tensors_by_name[i].second); + } + ggml_cuda_free_scratch(); } +#endif -#ifdef GGML_USE_CUBLAS - for (size_t i = 0; i < tensors_by_name.size(); ++i) { - ggml_cuda_free_data(tensors_by_name[i].second); - } - ggml_cuda_free_scratch(); -#elif defined(GGML_USE_CLBLAST) +#if defined(GGML_USE_CLBLAST) for (size_t i = 0; i < tensors_by_name.size(); ++i) { ggml_cl_free_data(tensors_by_name[i].second); } #endif + if (ctx) { + ggml_free(ctx); + } + + ggml_backend_buffer_free(buf); } }; struct llama_context { llama_context(const llama_model & model) : model(model), t_start_us(model.t_start_us), t_load_us(model.t_load_us) {} ~llama_context() { -#ifdef GGML_USE_METAL - if (ctx_metal) { - ggml_metal_free(ctx_metal); - } -#endif - if (alloc) { - ggml_allocr_free(alloc); - } + ggml_allocr_free(alloc); + ggml_backend_buffer_free(buf_alloc); + ggml_backend_free(backend); } llama_cparams cparams; + ggml_backend_t backend = nullptr; + const llama_model & model; // key + value cache for the self attention @@ -1350,23 +1568,22 @@ struct llama_context { // decode output (2-dimensional array: [n_tokens][n_vocab]) std::vector logits; +#ifndef NDEBUG + // guard against access to unset logits + std::vector logits_valid; +#endif bool logits_all = false; // input embedding (1-dimensional array: [n_embd]) std::vector embedding; - // reusable buffer for `struct ggml_graph_plan.work_data` - std::vector work_buffer; - // memory buffers used to evaluate the model - llama_buffer buf_compute; - - llama_buffer buf_alloc; + std::vector buf_compute_meta; + ggml_backend_buffer_t buf_alloc = NULL; ggml_allocr * alloc = NULL; -#ifdef GGML_USE_METAL - ggml_metal_context * ctx_metal = NULL; -#endif + // temporary buffer for copying data to/from the backend + std::vector> buf_copy; #ifdef GGML_USE_MPI ggml_mpi_context * ctx_mpi = NULL; @@ -1380,61 +1597,78 @@ struct llama_context { static bool llama_kv_cache_init( const struct llama_hparams & hparams, struct llama_kv_cache & cache, - ggml_type wtype, + ggml_type ktype, + ggml_type vtype, uint32_t n_ctx, - int n_gpu_layers) { + int n_gpu_layers, + bool offload) { const uint32_t n_embd = hparams.n_embd_gqa(); const uint32_t n_layer = hparams.n_layer; - const int64_t n_mem = n_layer*n_ctx; - const int64_t n_elements = n_embd*n_mem; - cache.has_shift = false; cache.head = 0; cache.size = n_ctx; + cache.used = 0; cache.cells.clear(); cache.cells.resize(n_ctx); - cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead()); - memset(cache.buf.data, 0, cache.buf.size); - struct ggml_init_params params; - params.mem_size = cache.buf.size; - params.mem_buffer = cache.buf.data; - params.no_alloc = false; + params.mem_size = 2u*n_layer*ggml_tensor_overhead(); + params.mem_buffer = NULL; + params.no_alloc = true; cache.ctx = ggml_init(params); + size_t vram_kv_cache = 0; + if (!cache.ctx) { LLAMA_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__); return false; } - cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); - cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); - ggml_set_name(cache.k, "cache_k"); - ggml_set_name(cache.v, "cache_v"); + cache.k_l.reserve(n_layer); + cache.v_l.reserve(n_layer); + + const int i_gpu_start = (int) n_layer - n_gpu_layers; + + for (int i = 0; i < (int) n_layer; i++) { + ggml_tensor * k = ggml_new_tensor_1d(cache.ctx, ktype, n_embd*n_ctx); + ggml_tensor * v = ggml_new_tensor_1d(cache.ctx, vtype, n_embd*n_ctx); + ggml_format_name(k, "cache_k_l%d", i); + ggml_format_name(v, "cache_v_l%d", i); + cache.k_l.push_back(k); + cache.v_l.push_back(v); +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (i >= i_gpu_start) { + if (offload) { + ggml_cuda_assign_buffers_no_scratch(k); + ggml_cuda_assign_buffers_no_scratch(v); + vram_kv_cache += ggml_nbytes(k); + vram_kv_cache += ggml_nbytes(v); + // HACK: mark tensor as allocated + k->data = v->data = (void *)(uintptr_t)1; + } + } +#endif // GGML_USE_CUBLAS + } - (void) n_gpu_layers; -#ifdef GGML_USE_CUBLAS - size_t vram_kv_cache = 0; + // allocate tensors + cache.buf = ggml_backend_alloc_ctx_tensors_from_buft(cache.ctx, llama_default_buffer_type(n_gpu_layers)); - if (n_gpu_layers > (int)n_layer + 1) { - ggml_cuda_assign_buffers_no_scratch(cache.v); - LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.v); - } - if (n_gpu_layers > (int)n_layer + 2) { - ggml_cuda_assign_buffers_no_scratch(cache.k); - LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.k); + // buf may be NULL with full offload + if (cache.buf) { + // initialize the buffer to avoid NaNs in the padding + ggml_backend_buffer_clear(cache.buf, 0); } + if (vram_kv_cache > 0) { LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MB\n", __func__, vram_kv_cache / 1024.0 / 1024.0); } -#endif // GGML_USE_CUBLAS + + GGML_UNUSED(i_gpu_start); + GGML_UNUSED(offload); return true; } @@ -1491,6 +1725,8 @@ static bool llama_kv_cache_find_slot( } } + cache.used += n_tokens; + return true; } @@ -1511,6 +1747,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { cache.cells[i].seq_id.clear(); } cache.head = 0; + cache.used = 0; } static void llama_kv_cache_seq_rm( @@ -1533,6 +1770,9 @@ static void llama_kv_cache_seq_rm( continue; } if (cache.cells[i].seq_id.empty()) { + // keep count of the number of used cells + if (cache.cells[i].pos >= 0) cache.used--; + cache.cells[i].pos = -1; if (new_head == cache.size) new_head = i; } @@ -1540,7 +1780,7 @@ static void llama_kv_cache_seq_rm( } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size) cache.head = new_head; + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; } static void llama_kv_cache_seq_cp( @@ -1566,6 +1806,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id for (uint32_t i = 0; i < cache.size; ++i) { if (!cache.cells[i].has_seq_id(seq_id)) { + if (cache.cells[i].pos >= 0) cache.used--; cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; @@ -1576,7 +1817,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size) cache.head = new_head; + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; } static void llama_kv_cache_seq_shift( @@ -1597,6 +1838,7 @@ static void llama_kv_cache_seq_shift( cache.cells[i].delta += delta; if (cache.cells[i].pos < 0) { + if (!cache.cells[i].seq_id.empty()) cache.used--; cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; @@ -1647,6 +1889,169 @@ static std::string llama_format_tensor_shape(const struct ggml_tensor * t) { return buf; } +namespace GGUFMeta { + template + struct GKV_Base_Type { + static constexpr gguf_type gt = gt_; + + static T getter(const gguf_context * ctx, const int kid) { + return gfun(ctx, kid); + } + }; + + template struct GKV_Base; + + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + + template<> struct GKV_Base { + static constexpr gguf_type gt = GGUF_TYPE_STRING; + + static std::string getter(const gguf_context * ctx, const int kid) { + return gguf_get_val_str(ctx, kid); + } + }; + + struct ArrayInfo{ + const gguf_type gt; + const size_t length; + const void * data; + }; + + template<> struct GKV_Base { + public: + static constexpr gguf_type gt = GGUF_TYPE_ARRAY; + static ArrayInfo getter(const gguf_context *ctx, const int k) { + return ArrayInfo { + gguf_get_arr_type(ctx, k), + size_t(gguf_get_arr_n(ctx, k)), + gguf_get_arr_data(ctx, k), + }; + } + }; + + template + class GKV: public GKV_Base { + GKV() = delete; + + public: + static T get_kv(const gguf_context * ctx, const int k) { + const enum gguf_type kt = gguf_get_kv_type(ctx, k); + + if (kt != GKV::gt) { + throw std::runtime_error(format("key %s has wrong type %s but expected type %s", + gguf_get_key(ctx, k), gguf_type_name(kt), gguf_type_name(GKV::gt))); + } + return GKV::getter(ctx, k); + } + + static const char * override_type_to_str(const llama_model_kv_override_type ty) { + switch (ty) { + case LLAMA_KV_OVERRIDE_BOOL: return "bool"; + case LLAMA_KV_OVERRIDE_INT: return "int"; + case LLAMA_KV_OVERRIDE_FLOAT: return "float"; + } + return "unknown"; + } + + static bool validate_override(const llama_model_kv_override_type expected_type, const struct llama_model_kv_override *override) { + if (!override) { return false; } + if (override->tag == expected_type) { + LLAMA_LOG_INFO("%s: Using metadata override (%5s) '%s' = ", + __func__, override_type_to_str(override->tag), override->key); + switch (override->tag) { + case LLAMA_KV_OVERRIDE_BOOL: { + printf("%s\n", override->bool_value ? "true" : "false"); + } break; + case LLAMA_KV_OVERRIDE_INT: { + printf("%" PRId64 "\n", override->int_value); + } break; + case LLAMA_KV_OVERRIDE_FLOAT: { + printf("%.6f\n", override->float_value); + } break; + default: + // Shouldn't be possible to end up here, but just in case... + throw std::runtime_error( + format("Unsupported attempt to override %s type for metadata key %s\n", + override_type_to_str(override->tag), override->key)); + } + return true; + } + LLAMA_LOG_WARN("%s: Warning: Bad metadata override type for key '%s', expected %s but got %s\n", + __func__, override->key, override_type_to_str(expected_type), override_type_to_str(override->tag)); + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(OT & target, const struct llama_model_kv_override *override) { + if (validate_override(LLAMA_KV_OVERRIDE_BOOL, override)) { + target = override->bool_value; + return true; + } + return false; + } + + template + static typename std::enable_if::value && std::is_integral::value, bool>::type + try_override(OT & target, const struct llama_model_kv_override *override) { + if (validate_override(LLAMA_KV_OVERRIDE_INT, override)) { + target = override->int_value; + return true; + } + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(T & target, const struct llama_model_kv_override *override) { + if (validate_override(LLAMA_KV_OVERRIDE_FLOAT, override)) { + target = override->float_value; + return true; + } + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(T & target, const struct llama_model_kv_override *override) { + (void)target; + (void)override; + if (!override) { return false; } + // Currently, we should never end up here so it would be a bug if we do. + throw std::runtime_error(format("Unsupported attempt to override string type for metadata key %s\n", + override ? override->key : "NULL")); + } + + static bool set(const gguf_context * ctx, const int k, T & target, const struct llama_model_kv_override *override = nullptr) { + if (try_override(target, override)) { + return true; + } + if (k < 0) { return false; } + target = get_kv(ctx, k); + return true; + } + + static bool set(const gguf_context * ctx, const char * key, T & target, const struct llama_model_kv_override *override = nullptr) { + return set(ctx, gguf_find_key(ctx, key), target, override); + } + + static bool set(const gguf_context * ctx, const std::string & key, T & target, const struct llama_model_kv_override *override = nullptr) { + return set(ctx, key.c_str(), target, override); + } + }; +} + struct llama_model_loader { int n_kv = 0; int n_tensors = 0; @@ -1662,21 +2067,34 @@ struct llama_model_loader { llama_fver fver; std::unique_ptr mapping; + std::unordered_map kv_overrides; struct gguf_context * ctx_gguf = NULL; struct ggml_context * ctx_meta = NULL; - llama_model_loader(const std::string & fname, bool use_mmap) : file(fname.c_str(), "rb") { + std::string arch_name; + LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); + + llama_model_loader(const std::string & fname, bool use_mmap, const struct llama_model_kv_override * param_overrides_p) : file(fname.c_str(), "rb") { struct gguf_init_params params = { /*.no_alloc = */ true, /*.ctx = */ &ctx_meta, }; + if (param_overrides_p != nullptr) { + for (const struct llama_model_kv_override *p = param_overrides_p; p->key[0] != 0; p++) { + kv_overrides.insert({std::string(p->key), *p}); + } + } + ctx_gguf = gguf_init_from_file(fname.c_str(), params); if (!ctx_gguf) { throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str())); } + get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false); + llm_kv = LLM_KV(llm_arch_from_string(arch_name)); + n_kv = gguf_get_n_kv(ctx_gguf); n_tensors = gguf_get_n_tensors(ctx_gguf); @@ -1701,17 +2119,16 @@ struct llama_model_loader { enum ggml_type type_max = GGML_TYPE_F32; for (int i = 0; i < n_tensors; i++) { - const char * name = gguf_get_tensor_name(ctx_gguf, i); - struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, name); + enum ggml_type type = gguf_get_tensor_type(ctx_gguf, i); - n_type[meta->type]++; + n_type[type]++; - if (n_type_max < n_type[meta->type]) { - n_type_max = n_type[meta->type]; - type_max = meta->type; + if (n_type_max < n_type[type]) { + n_type_max = n_type[type]; + type_max = type; } - LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str()); + // LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str()); } switch (type_max) { @@ -1728,10 +2145,10 @@ struct llama_model_loader { case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; default: - { - LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); - ftype = LLAMA_FTYPE_ALL_F32; - } break; + { + LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); + ftype = LLAMA_FTYPE_ALL_F32; + } break; } // this is a way to mark that we have "guessed" the file type @@ -1744,11 +2161,23 @@ struct llama_model_loader { } } + LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(ctx_gguf, i); - const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const char * name = gguf_get_key(ctx_gguf, i); + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx_gguf, i)), gguf_get_arr_n(ctx_gguf, i)) + : gguf_type_name(type); + + std::string value = gguf_kv_to_str(ctx_gguf, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); - LLAMA_LOG_INFO("%s: - kv %3d: %42s %-8s\n", __func__, i, name, gguf_type_name(type)); + LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); } // print type counts @@ -1778,62 +2207,90 @@ struct llama_model_loader { } } - std::string get_arch_name() const { - const auto kv = LLM_KV(LLM_ARCH_UNKNOWN); + template + typename std::enable_if::value, bool>::type + get_arr_n(const std::string & key, T & result, const bool required = true) { + const int kid = gguf_find_key(ctx_gguf, key.c_str()); - std::string arch_name; - GGUF_GET_KEY(ctx_gguf, arch_name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_ARCHITECTURE)); + if (kid < 0) { + if (required) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + return false; + } - return arch_name; - } + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(ctx_gguf, kid); - enum llm_arch get_arch() const { - const std::string arch_name = get_arch_name(); - return llm_arch_from_string(arch_name); + result = arr_info.length; + return true; } - const char * get_tensor_name(int i) const { - return gguf_get_tensor_name(ctx_gguf, i); + template + typename std::enable_if::value, bool>::type + get_arr_n(const enum llm_kv kid, T & result, const bool required = true) { + return get_arr_n(llm_kv(kid), result, required); } - struct ggml_tensor * get_tensor_meta(int i) const { - return ggml_get_tensor(ctx_meta, get_tensor_name(i)); - } + template + bool get_key(const std::string & key, T & result, const bool required = true) { + auto it = kv_overrides.find(key); - void calc_sizes(size_t & ctx_size_p, size_t & mmapped_size_p) const { - ctx_size_p = 0; - mmapped_size_p = 0; + const struct llama_model_kv_override * override = + it != kv_overrides.end() ? &it->second : nullptr; - for (int i = 0; i < n_tensors; i++) { - struct ggml_tensor * meta = get_tensor_meta(i); - ctx_size_p += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE; - (use_mmap ? mmapped_size_p : ctx_size_p) += ggml_nbytes_pad(meta); + const bool found = GGUFMeta::GKV::set(ctx_gguf, key, result, override); + + if (required && !found) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); } + + return found; } - struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend_type backend) { - if (backend != GGML_BACKEND_CPU) { - ggml_set_no_alloc(ctx, true); - } + template + bool get_key(const enum llm_kv kid, T & result, const bool required = true) { + return get_key(llm_kv(kid), result, required); + } + + std::string get_arch_name() const { + return arch_name; + } + + enum llm_arch get_arch() const { + return llm_kv.arch; + } + + const char * get_tensor_name(int i) const { + return gguf_get_tensor_name(ctx_gguf, i); + } + + struct ggml_tensor * get_tensor_meta(const char * name) const { + return ggml_get_tensor(ctx_meta, name); + } + struct ggml_tensor * get_tensor_meta(int i) const { + return get_tensor_meta(get_tensor_name(i)); + } + + struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend_type backend) { struct ggml_tensor * tensor = ggml_dup_tensor(ctx, meta); tensor->backend = backend; // TODO: ggml_set_backend ggml_set_name(tensor, ggml_get_name(meta)); - if (backend != GGML_BACKEND_CPU) { - ggml_set_no_alloc(ctx, use_mmap); - } - n_created++; return tensor; } - struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, ggml_backend_type backend) { + struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, ggml_backend_type backend, bool required = true) { struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str()); if (cur == NULL) { + if (!required) { + return NULL; + } throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str())); } @@ -1879,91 +2336,144 @@ struct llama_model_loader { return gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, idx); } + void init_mapping(bool prefetch = true) { + /* + // prefetch only CPU tensors + if (use_mmap) { + size_t size_pref = 0; // prefetch + + for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { + struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i)); + if (cur->backend == GGML_BACKEND_CPU) { + size_t tensor_end = gguf_get_tensor_offset(ctx_gguf, i) + ggml_nbytes(cur); + size_pref = std::max(size_pref, tensor_end); + } + } + mapping.reset(new llama_mmap(&file, gguf_get_data_offset(ctx_gguf) + size_pref, ggml_is_numa())); + } + */ + // prefetch the whole file - all the data is needed anyway + if (use_mmap) { + mapping.reset(new llama_mmap(&file, prefetch ? -1 : 0, ggml_is_numa())); + } + } + + // for backwards compatibility, does not support ggml-backend void load_data_for(struct ggml_tensor * cur) const { const size_t offs = file_offset(ggml_get_name(cur)); - if (use_mmap) { - cur->data = (uint8_t *) mapping->addr + offs; + if (use_mmap && mapping) { + GGML_ASSERT(cur->data == nullptr); + cur->data = (uint8_t *)mapping->addr + offs; } else { + GGML_ASSERT(cur->data != nullptr); file.seek(offs, SEEK_SET); file.read_raw(cur->data, ggml_nbytes(cur)); } } - void load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) { + // Returns false if cancelled by progress_callback + bool load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const { size_t size_data = 0; - size_t size_lock = 0; - size_t size_pref = 0; // prefetch for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i)); size_data += ggml_nbytes(cur); - if (cur->backend == GGML_BACKEND_CPU) { - size_pref += ggml_nbytes(cur); - } } - if (use_mmap) { - mapping.reset(new llama_mmap(&file, size_pref, ggml_is_numa())); + if (use_mmap && buf_mmap) { if (lmlock) { lmlock->init(mapping->addr); } } - size_t done_size = 0; +#if (defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)) || defined(GGML_USE_CLBLAST) + const bool legacy_offload = true; +#else + const bool legacy_offload = false; +#endif + + std::vector> read_buf; + + size_t size_done = 0; + + size_t mmap_first = -1; + size_t mmap_last = 0; + for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i)); GGML_ASSERT(cur); // unused tensors should have been caught by load_data already if (progress_callback) { - progress_callback((float) done_size / size_data, progress_callback_user_data); - } - - // allocate temp buffer if not using mmap - if (!use_mmap && cur->data == NULL) { - GGML_ASSERT(cur->backend != GGML_BACKEND_CPU); - #ifdef GGML_USE_CPU_HBM - cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur)); - #else - cur->data = (uint8_t*)malloc(ggml_nbytes(cur)); - #endif + if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { + return false; + } } - load_data_for(cur); + const size_t offs = file_offset(ggml_get_name(cur)); - switch (cur->backend) { - case GGML_BACKEND_CPU: - if (use_mmap && lmlock) { - size_lock += ggml_nbytes(cur); - lmlock->grow_to(size_lock); + if (!legacy_offload || cur->backend == GGML_BACKEND_CPU) { + if (use_mmap && mapping) { + if (buf_mmap) { + ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + offs); + if (lmlock) { + lmlock->grow_to(offs + ggml_nbytes(cur)); + } + mmap_first = std::min(mmap_first, offs); + mmap_last = std::max(mmap_last, offs + ggml_nbytes(cur)); + } else { + ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + offs, 0, ggml_nbytes(cur)); } - break; -#ifdef GGML_USE_CUBLAS - case GGML_BACKEND_GPU: - case GGML_BACKEND_GPU_SPLIT: - // old code: - //ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor); - - // TODO: test if this works !! - ggml_cuda_transform_tensor(cur->data, cur); - if (!use_mmap) { - free(cur->data); + } else { + if (ggml_backend_buffer_is_host(cur->buffer)) { + file.seek(offs, SEEK_SET); + file.read_raw(cur->data, ggml_nbytes(cur)); + } else { + read_buf.resize(ggml_nbytes(cur)); + file.seek(offs, SEEK_SET); + file.read_raw(read_buf.data(), ggml_nbytes(cur)); + ggml_backend_tensor_set(cur, read_buf.data(), 0, ggml_nbytes(cur)); } - break; + } + } else { + // HACK: mark tensor as allocated + cur->data = (void *)(uintptr_t)1; + void * data; + if (use_mmap && mapping) { + data = (uint8_t *) mapping->addr + offs; + } else { + read_buf.resize(ggml_nbytes(cur)); + file.seek(offs, SEEK_SET); + file.read_raw(read_buf.data(), ggml_nbytes(cur)); + data = read_buf.data(); + } + +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + ggml_cuda_transform_tensor(data, cur); #elif defined(GGML_USE_CLBLAST) - case GGML_BACKEND_GPU: - ggml_cl_transform_tensor(cur->data, cur); - if (!use_mmap) { - free(cur->data); - } - break; + GGML_ASSERT(cur->backend == GGML_BACKEND_GPU); + ggml_cl_transform_tensor(data, cur); +#else + GGML_ASSERT(!"GPU tensor without a GPU backend"); + GGML_UNUSED(data); #endif - default: - continue; } - done_size += ggml_nbytes(cur); + size_done += ggml_nbytes(cur); + } + + // unmap offloaded tensors and metadata + if (use_mmap && mapping) { + mapping->unmap_fragment(0, mmap_first); + mapping->unmap_fragment(mmap_last, mapping->size); + } + + if (progress_callback) { + // Even though the model is done loading, we still honor + // cancellation since we need to free allocations. + return progress_callback(1.0f, progress_callback_user_data); } + return true; } }; @@ -1986,25 +2496,25 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { switch (ftype) { case LLAMA_FTYPE_ALL_F32: return "all F32"; - case LLAMA_FTYPE_MOSTLY_F16: return "mostly F16"; - case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0"; - case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1"; + case LLAMA_FTYPE_MOSTLY_F16: return "F16"; + case LLAMA_FTYPE_MOSTLY_Q4_0: return "Q4_0"; + case LLAMA_FTYPE_MOSTLY_Q4_1: return "Q4_1"; case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: - return "mostly Q4_1, some F16"; - case LLAMA_FTYPE_MOSTLY_Q5_0: return "mostly Q5_0"; - case LLAMA_FTYPE_MOSTLY_Q5_1: return "mostly Q5_1"; - case LLAMA_FTYPE_MOSTLY_Q8_0: return "mostly Q8_0"; + return "Q4_1, some F16"; + case LLAMA_FTYPE_MOSTLY_Q5_0: return "Q5_0"; + case LLAMA_FTYPE_MOSTLY_Q5_1: return "Q5_1"; + case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0"; // K-quants - case LLAMA_FTYPE_MOSTLY_Q2_K: return "mostly Q2_K"; - case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "mostly Q3_K - Small"; - case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "mostly Q3_K - Medium"; - case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "mostly Q3_K - Large"; - case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "mostly Q4_K - Small"; - case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "mostly Q4_K - Medium"; - case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "mostly Q5_K - Small"; - case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "mostly Q5_K - Medium"; - case LLAMA_FTYPE_MOSTLY_Q6_K: return "mostly Q6_K"; + case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K"; + case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large"; + case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K"; default: return "unknown, may not work"; } @@ -2037,49 +2547,66 @@ static void llm_load_arch(llama_model_loader & ml, llama_model & model) { static void llm_load_hparams( llama_model_loader & ml, llama_model & model) { - struct gguf_context * ctx = ml.ctx_gguf; - - const auto kv = LLM_KV(model.arch); - auto & hparams = model.hparams; + const gguf_context * ctx = ml.ctx_gguf; + + // get metadata as string + for (int i = 0; i < gguf_get_n_kv(ctx); i++) { + enum gguf_type type = gguf_get_kv_type(ctx, i); + if (type == GGUF_TYPE_ARRAY) { + continue; + } + const char * name = gguf_get_key(ctx, i); + const std::string value = gguf_kv_to_str(ctx, i); + model.gguf_kv.emplace(name, value); + } // get general kv - GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME)); + ml.get_key(LLM_KV_GENERAL_NAME, model.name, false); // get hparams kv - GGUF_GET_KEY(ctx, hparams.n_vocab, gguf_get_arr_n, GGUF_TYPE_ARRAY, true, kv(LLM_KV_TOKENIZER_LIST)); - GGUF_GET_KEY(ctx, hparams.n_ctx_train, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_CONTEXT_LENGTH)); - GGUF_GET_KEY(ctx, hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH)); - GGUF_GET_KEY(ctx, hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH)); - GGUF_GET_KEY(ctx, hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT)); - GGUF_GET_KEY(ctx, hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT)); + ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + ml.get_key (LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); + ml.get_key (LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); + ml.get_key (LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff); + ml.get_key (LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head); + ml.get_key (LLM_KV_BLOCK_COUNT, hparams.n_layer); + ml.get_key (LLM_KV_EXPERT_COUNT, hparams.n_expert, false); + ml.get_key (LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + + GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS); + GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert); + if (hparams.n_expert > 0) { + GGML_ASSERT(hparams.n_expert_used > 0); + } else { + GGML_ASSERT(hparams.n_expert_used == 0); + } // n_head_kv is optional, default to n_head hparams.n_head_kv = hparams.n_head; - GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); + ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv, false); - hparams.rope_finetuned = false; - GGUF_GET_KEY(ctx, hparams.rope_finetuned, gguf_get_val_bool, GGUF_TYPE_BOOL, false, - kv(LLM_KV_ROPE_SCALING_FINETUNED)); + bool rope_finetuned = false; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; hparams.n_yarn_orig_ctx = hparams.n_ctx_train; - GGUF_GET_KEY(ctx, hparams.n_yarn_orig_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, - kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN)); + ml.get_key(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_yarn_orig_ctx, false); // rope_freq_base (optional) hparams.rope_freq_base_train = 10000.0f; - GGUF_GET_KEY(ctx, hparams.rope_freq_base_train, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); + ml.get_key(LLM_KV_ROPE_FREQ_BASE, hparams.rope_freq_base_train, false); std::string rope_scaling("linear"); - GGUF_GET_KEY(ctx, rope_scaling, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_ROPE_SCALING_TYPE)); + ml.get_key(LLM_KV_ROPE_SCALING_TYPE, rope_scaling, false); hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling); GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_UNSPECIFIED); // rope_freq_scale (inverse of the kv) is optional float ropescale = 0.0f; - GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALING_FACTOR)); - if (ropescale == 0.0f) { // try the old key name - GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); + if (!ml.get_key(LLM_KV_ROPE_SCALING_FACTOR, ropescale, false)) { + // try the old key name + ml.get_key(LLM_KV_ROPE_SCALE_LINEAR, ropescale, false); } hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; @@ -2087,7 +2614,7 @@ static void llm_load_hparams( { hparams.n_rot = hparams.n_embd / hparams.n_head; - GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT)); + ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) { if (hparams.n_rot != hparams.n_embd / hparams.n_head) { @@ -2102,9 +2629,10 @@ static void llm_load_hparams( switch (model.arch) { case LLM_ARCH_LLAMA: { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { + case 22: model.type = e_model::MODEL_1B; break; 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; @@ -2116,7 +2644,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_FALCON: { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { case 32: model.type = e_model::MODEL_7B; break; @@ -2126,7 +2654,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_BAICHUAN: { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { case 32: model.type = e_model::MODEL_7B; break; case 40: model.type = e_model::MODEL_13B; break; @@ -2135,7 +2663,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_STARCODER: { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { case 24: model.type = e_model::MODEL_1B; break; case 36: model.type = e_model::MODEL_3B; break; @@ -2146,7 +2674,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_PERSIMMON: { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { case 36: model.type = e_model::MODEL_8B; break; default: model.type = e_model::MODEL_UNKNOWN; @@ -2154,7 +2682,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_REFACT: { - GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { case 32: model.type = e_model::MODEL_1B; break; default: model.type = e_model::MODEL_UNKNOWN; @@ -2162,7 +2690,7 @@ static void llm_load_hparams( } break; case LLM_ARCH_BLOOM: { - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { case 24: model.type = e_model::MODEL_1B; break; @@ -2177,9 +2705,9 @@ static void llm_load_hparams( { hparams.f_clamp_kqv = 0.0f; - GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); - GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_CLAMP_KQV)); - GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS)); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv, false); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); switch (hparams.n_layer) { case 32: model.type = e_model::MODEL_7B; break; @@ -2187,6 +2715,35 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_STABLELM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_PHI2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + default: (void)0; } @@ -2227,7 +2784,7 @@ static void llm_load_vocab( { std::string tokenizer_name; - GGUF_GET_KEY(ctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL)); + ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_name); if (tokenizer_name == "llama") { vocab.type = LLAMA_VOCAB_TYPE_SPM; @@ -2317,16 +2874,30 @@ static void llm_load_vocab( }; for (const auto & it : special_token_types) { const std::string & key = kv(std::get<0>(it)); - int32_t & id = std::get<1>(it), old_id = id; + int32_t & id = std::get<1>(it); - GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key); - // Must be >= -1 and < vocab size. Since the key is unsigned, -1 - // can only come from the default value, so there's no point in - // validating that. - if (size_t(id + 1) > vocab.id_to_token.size()) { - LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n", - __func__, key.c_str(), id, old_id); - id = old_id; + uint32_t new_id; + if (!ml.get_key(std::get<0>(it), new_id, false)) { + continue; + } + if (new_id >= vocab.id_to_token.size()) { + LLAMA_LOG_WARN("%s: bad special token: '%s' = %ud, using default id %d\n", + __func__, key.c_str(), new_id, id); + } else { + id = new_id; + } + + } + + // Handle add_bos_token and add_eos_token + { + bool temp = true; + + if (ml.get_key(LLM_KV_TOKENIZER_ADD_BOS, temp, false)) { + vocab.special_add_bos = int(temp); + } + if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) { + vocab.special_add_eos = int(temp); } } } @@ -2338,7 +2909,7 @@ static void llm_load_vocab( // The assumption is, since special tokens aren't meant to be exposed to end user, they are designed // to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer // are special tokens. - // From testing, this appears to corelate 1:1 with special tokens. + // From testing, this appears to correlate 1:1 with special tokens. // // Counting special tokens and verifying in only one direction @@ -2451,6 +3022,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); + LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); + LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str()); LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); @@ -2459,25 +3032,26 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9); - if (ml.n_bytes < GB) { - LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); + if (ml.n_bytes < GiB) { + LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } else { LLAMA_LOG_INFO("%s: model size = %.2f GiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } // general kv - LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, model.name.c_str()); + LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, model.name.c_str()); // special tokens - if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); } - if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); } - if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); } - if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); } - if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); } - if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); } + if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); } + if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); } + if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); } + if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); } + if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); } + if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); } } -static void llm_load_tensors( +// Returns false if cancelled by progress_callback +static bool llm_load_tensors( llama_model_loader & ml, llama_model & model, int n_gpu_layers, @@ -2493,25 +3067,16 @@ static void llm_load_tensors( model.n_gpu_layers = n_gpu_layers; - size_t ctx_size; - size_t mmapped_size; + size_t ctx_size = ggml_tensor_overhead() * ml.n_tensors; - ml.calc_sizes(ctx_size, mmapped_size); - - LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, ctx_size/1024.0/1024.0); // create the ggml context { - model.buf.resize(ctx_size); - if (use_mlock) { - model.mlock_buf.init (model.buf.data); - model.mlock_buf.grow_to(model.buf.size); - } - struct ggml_init_params params = { - /*.mem_size =*/ model.buf.size, - /*.mem_buffer =*/ model.buf.data, - /*.no_alloc =*/ ml.use_mmap, + /*.mem_size =*/ ctx_size, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, }; model.ctx = ggml_init(params); @@ -2521,22 +3086,25 @@ static void llm_load_tensors( } (void) main_gpu; -#ifdef GGML_USE_CUBLAS - LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); - ggml_cuda_set_main_device(main_gpu); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT + + enum ggml_backend_type llama_backend_offload = GGML_BACKEND_CPU; + enum ggml_backend_type llama_backend_offload_split = GGML_BACKEND_CPU; + +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (ggml_cublas_loaded()) { + LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); + ggml_cuda_set_main_device(main_gpu); + + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU_SPLIT; + } #elif defined(GGML_USE_CLBLAST) - LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU -#else -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_CPU + LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU; #endif - // prepare memory for the weights - size_t vram_weights = 0; + // create tensors for the weights { const int64_t n_embd = hparams.n_embd; const int64_t n_embd_gqa = hparams.n_embd_gqa(); @@ -2556,15 +3124,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2572,13 +3133,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -2588,8 +3142,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2600,17 +3154,33 @@ static void llm_load_tensors( layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + // optional bias tensors + layer.bq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, backend, false); + layer.bk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, backend, false); + layer.bv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, backend, false); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend, false); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_gate_inp = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd}, backend, false); - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); + if (layer.ffn_gate_inp == nullptr) { + GGML_ASSERT(hparams.n_expert == 0); + GGML_ASSERT(hparams.n_expert_used == 0); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + } else { + GGML_ASSERT(hparams.n_expert > 0); + GGML_ASSERT(hparams.n_expert_used > 0); + + // MoE branch + for (uint32_t x = 0; x < hparams.n_expert; ++x) { + layer.ffn_gate_exp[x] = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), {n_embd, n_ff}, backend_split); + layer.ffn_down_exp[x] = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd}, backend_split); + layer.ffn_up_exp[x] = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, x), {n_embd, n_ff}, backend_split); + } } } } break; @@ -2622,15 +3192,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2638,13 +3201,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -2654,8 +3210,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2671,19 +3227,10 @@ static void llm_load_tensors( layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } } break; case LLM_ARCH_FALCON: { - // TODO: CPU-only for now - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output @@ -2692,15 +3239,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2709,14 +3249,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -2726,8 +3258,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2737,11 +3269,6 @@ static void llm_load_tensors( if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(layer.attn_norm_2); - vram_weights += ggml_nbytes(layer.attn_norm_2_b); - } } layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); @@ -2749,13 +3276,6 @@ static void llm_load_tensors( layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + - ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + - ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); - } } } break; case LLM_ARCH_STARCODER: @@ -2769,15 +3289,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2786,14 +3299,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -2803,8 +3308,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2825,16 +3330,6 @@ static void llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + - ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + - ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + - ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b) + - ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b); - } } } break; case LLM_ARCH_PERSIMMON: @@ -2846,15 +3341,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2863,22 +3351,14 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; const int i_gpu_start = n_layer - n_gpu_layers; model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; auto & layer = model.layers[i]; layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); @@ -2900,8 +3380,6 @@ static void llm_load_tensors( } break; case LLM_ARCH_BLOOM: { - // TODO: CPU-only for now - model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); @@ -2912,15 +3390,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2929,14 +3400,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - vram_weights += ggml_nbytes(model.output_norm_b); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -2946,8 +3409,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2968,16 +3431,6 @@ static void llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + - ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + - ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + - ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b) + - ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b); - } } } break; case LLM_ARCH_MPT: @@ -2990,15 +3443,8 @@ static void llm_load_tensors( ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU -#ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; -#else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; -#endif // _WIN32 - - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -3006,13 +3452,6 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.output_norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } } const uint32_t n_ff = hparams.n_ff; @@ -3022,8 +3461,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -3035,94 +3474,285 @@ static void llm_load_tensors( layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attn_norm) + - ggml_nbytes(layer.wqkv) + - ggml_nbytes(layer.wo) + - ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.ffn_down) + - ggml_nbytes(layer.ffn_up); - } } } break; - default: - throw std::runtime_error("unknown architecture"); - } - } + case LLM_ARCH_STABLELM: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - ml.done_getting_tensors(); + // output + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - // print memory requirements - { - // this is the total memory required to run the inference - size_t mem_required = - ctx_size + - mmapped_size - vram_weights; // weights in VRAM not in memory + if (n_gpu_layers > int(n_layer)) { + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } - LLAMA_LOG_INFO("%s: mem required = %7.2f MB\n", __func__, mem_required / 1024.0 / 1024.0); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + } -#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) - const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); + const uint32_t n_ff = hparams.n_ff; - LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu); - if (n_gpu_layers > (int) hparams.n_layer) { - LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__); - } + const int i_gpu_start = n_layer - n_gpu_layers; -#ifdef GGML_USE_CUBLAS - const int max_backend_supported_layers = hparams.n_layer + 3; - const int max_offloadable_layers = hparams.n_layer + 3; -#elif GGML_USE_CLBLAST - const int max_backend_supported_layers = hparams.n_layer + 1; - const int max_offloadable_layers = hparams.n_layer + 1; -#endif // GGML_USE_CUBLAS + model.layers.resize(n_layer); - LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); - LLAMA_LOG_INFO("%s: VRAM used: %.2f MB\n", __func__, vram_weights / 1024.0 / 1024.0); -#else - (void) n_gpu_layers; -#endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) - } + for (uint32_t i = 0; i < n_layer; ++i) { + /* + llama_model_loader: - tensor 4: blk.0.attn_output.weight f16 [ 2560, 2560, 1, 1 ] + */ + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT - // populate `tensors_by_name` - for (int i = 0; i < ml.n_tensors; ++i) { - struct ggml_tensor * cur = ggml_get_tensor(ctx, ml.get_tensor_name(i)); - model.tensors_by_name.emplace_back(ggml_get_name(cur), cur); - } + auto & layer = model.layers[i]; - (void) tensor_split; -#ifdef GGML_USE_CUBLAS - { - ggml_cuda_set_tensor_split(tensor_split); - } -#endif + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - ml.load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL); + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - if (progress_callback) { - progress_callback(1.0f, progress_callback_user_data); - } + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - model.mapping = std::move(ml.mapping); + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + } + } break; + case LLM_ARCH_QWEN: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; - // loading time will be recalculate after the first eval, so - // we take page faults deferred by mmap() into consideration - model.t_load_us = ggml_time_us() - model.t_start_us; -} + if (n_gpu_layers > int(n_layer)) { + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } -static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { - try { - llama_model_loader ml(fname, params.use_mmap); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + } - model.hparams.vocab_only = params.vocab_only; + const uint32_t n_ff = hparams.n_ff / 2; - llm_load_arch (ml, model); - llm_load_hparams(ml, model); - llm_load_vocab (ml, model); + const int i_gpu_start = n_layer - n_gpu_layers; - llm_load_print_meta(ml, model); + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd * 3}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd * 3}, backend); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + } + } break; + case LLM_ARCH_PHI2: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; + + if (n_gpu_layers > int(n_layer)) { + backend_norm = llama_backend_offload; + backend_output = llama_backend_offload; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + model.output_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "bias"), {n_vocab}, backend_output); + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); + + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); + + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + } + } break; + default: + throw std::runtime_error("unknown architecture"); + } + } + + ml.done_getting_tensors(); + + ml.init_mapping(); + + // allocate tensors + size_t vram_weights = 0; + size_t buf_size = 0; + + ggml_backend_buffer_type_t buft = llama_default_buffer_type(n_gpu_layers); + + for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { + // GGML_BACKEND_GPU tensors are for CUDA and OpenCL only, which are handled separately without ggml-backend + if (t->backend == GGML_BACKEND_CPU) { + buf_size += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), ggml_backend_buft_get_alignment(buft)); + } else { + vram_weights += ggml_nbytes(t); + } + } + + // create backend buffer + ggml_backend_buffer_t buf_mmap = nullptr; + +#ifdef GGML_USE_METAL + if (n_gpu_layers > 0) { + if (ml.use_mmap) { + const size_t max_size = ggml_get_max_tensor_size(ctx); + model.buf = ggml_backend_metal_buffer_from_ptr(ml.mapping->addr, ml.mapping->size, max_size); + buf_mmap = model.buf; + } else { + model.buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_metal_buffer_type()); + } + } +#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST) + // for testing only + if (n_gpu_layers > 0) { + model.buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cuda_buffer_type(0)); + } +#endif + + if (model.buf == nullptr) { + // CPU backend, and indirectly CUDA and OpenCL + if (ml.use_mmap) { + model.buf = ggml_backend_cpu_buffer_from_ptr(ml.mapping->addr, ml.mapping->size); + buf_mmap = model.buf; + } else { + // allocate only CPU tensors + model.buf = ggml_backend_buft_alloc_buffer(buft, buf_size); + ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(model.buf); + for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { + if (t->backend == GGML_BACKEND_CPU) { + ggml_tallocr_alloc(alloc, t); + } + } + ggml_tallocr_free(alloc); + } + } + + if (use_mlock && ggml_backend_buffer_is_host(model.buf)) { + model.mlock_buf.init (ggml_backend_buffer_get_base(model.buf)); + model.mlock_buf.grow_to(ggml_backend_buffer_get_size(model.buf)); + } + + // print memory requirements + { + size_t sys_mem_required = ctx_size + buf_size; + + if (sys_mem_required > 0) { + LLAMA_LOG_INFO("%s: system memory used = %7.2f MiB\n", __func__, sys_mem_required / 1024.0 / 1024.0); + } + if (vram_weights > 0) { + LLAMA_LOG_INFO("%s: VRAM used = %7.2f MiB\n", __func__, vram_weights / 1024.0 / 1024.0); + } + +#if (defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)) || defined(GGML_USE_CLBLAST) + const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); + + LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu); + if (n_gpu_layers > (int) hparams.n_layer) { + LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__); + } + + const int max_backend_supported_layers = hparams.n_layer + 1; + const int max_offloadable_layers = hparams.n_layer + 1; + + LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); +#endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) + } + +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + ggml_cuda_set_tensor_split(tensor_split); +#else + GGML_UNUSED(tensor_split); +#endif // GGML_USE_CUBLAS + + // populate tensors_by_name + for (int i = 0; i < ml.n_tensors; ++i) { + struct ggml_tensor * cur = ggml_get_tensor(ctx, ml.get_tensor_name(i)); + model.tensors_by_name.emplace_back(ggml_get_name(cur), cur); + } + + if (!ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL)) { + return false; + } + + model.mapping = std::move(ml.mapping); + + // loading time will be recalculate after the first eval, so + // we take page faults deferred by mmap() into consideration + model.t_load_us = ggml_time_us() - model.t_start_us; + return true; +} + +// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback +static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { + try { + llama_model_loader ml(fname, params.use_mmap, params.kv_overrides); + + model.hparams.vocab_only = params.vocab_only; + + llm_load_arch (ml, model); + llm_load_hparams(ml, model); + llm_load_vocab (ml, model); + + llm_load_print_meta(ml, model); if (model.hparams.n_vocab != model.vocab.id_to_token.size()) { throw std::runtime_error("vocab size mismatch"); @@ -3130,19 +3760,21 @@ static bool llama_model_load(const std::string & fname, llama_model & model, con if (params.vocab_only) { LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); - return true; + return 0; } - llm_load_tensors( + if (!llm_load_tensors( ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock, params.progress_callback, params.progress_callback_user_data - ); + )) { + return -2; + } } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); - return false; + return -1; } - return true; + return 0; } // @@ -3210,7 +3842,7 @@ static void llm_build_k_shift( struct ggml_cgraph * graph, llm_rope_type type, int64_t n_ctx, - int64_t n_rot, + int n_rot, float freq_base, float freq_scale, const llm_build_cb & cb) { @@ -3241,11 +3873,11 @@ static void llm_build_k_shift( struct ggml_tensor * tmp = // we rotate only the first n_rot dimensions ggml_rope_custom_inplace(ctx, - ggml_view_3d(ctx, kv.k, - n_rot, n_head_kv, n_ctx, - ggml_element_size(kv.k)*n_embd_head, - ggml_element_size(kv.k)*n_embd_gqa, - ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il), + ggml_view_3d(ctx, kv.k_l[il], + n_embd_head, n_head_kv, n_ctx, + ggml_row_size(kv.k_l[il]->type, n_embd_head), + ggml_row_size(kv.k_l[il]->type, n_embd_gqa), + 0), K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(tmp, "K_shifted", il); @@ -3272,13 +3904,13 @@ static void llm_build_kv_store( //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed cb(v_cur_t, "v_cur_t", il); - struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k, n_tokens*n_embd_gqa, - (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head)); + struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_gqa, + (ggml_row_size(kv.k_l[il]->type, n_embd_gqa))*kv_head); cb(k_cache_view, "k_cache_view", il); - struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv.v), - (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v)); + struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv.v_l[il]), + (kv_head)*ggml_element_size(kv.v_l[il])); cb(v_cache_view, "v_cache_view", il); // important: storing RoPE-ed version of K in the KV cache! @@ -3407,17 +4039,18 @@ static struct ggml_tensor * llm_build_ffn( // if max_alibi_bias > 0 then apply ALiBi static struct ggml_tensor * llm_build_kqv( struct ggml_context * ctx, + const llama_model & model, const llama_hparams & hparams, const llama_kv_cache & kv, struct ggml_tensor * wo, struct ggml_tensor * wo_b, struct ggml_tensor * q_cur, - struct ggml_tensor * kq_scale, struct ggml_tensor * kq_mask, int64_t n_ctx, int32_t n_tokens, int32_t n_kv, float max_alibi_bias, + float kq_scale, const llm_build_cb & cb, int il) { const int64_t n_embd = hparams.n_embd; @@ -3430,40 +4063,52 @@ static struct ggml_tensor * llm_build_kqv( cb(q, "q", il); struct ggml_tensor * k = - ggml_view_3d(ctx, kv.k, + ggml_view_3d(ctx, kv.k_l[il], n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv.k)*n_embd_gqa, - ggml_element_size(kv.k)*n_embd_head, - ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il); + ggml_row_size(kv.k_l[il]->type, n_embd_gqa), + ggml_row_size(kv.k_l[il]->type, n_embd_head), + 0); cb(k, "k", il); struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); cb(kq, "kq", il); - kq = ggml_scale(ctx, kq, kq_scale); - cb(kq, "kq_scaled", il); + if (model.arch == LLM_ARCH_PHI2) { + // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs + // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 + ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + } if (max_alibi_bias > 0.0f) { - // TODO: n_head or n_head_kv - // TODO: K-shift is likely not working - // TODO: change to ggml_add - kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias); - cb(kq, "kq_scaled_alibi", il); - } + // temporary branch until we figure out how to handle ggml_alibi through ggml_add + kq = ggml_scale(ctx, kq, kq_scale); + cb(kq, "kq_scaled", il); + + if (max_alibi_bias > 0.0f) { + // TODO: n_head or n_head_kv + // TODO: K-shift is likely not working + // TODO: change to ggml_add + kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias); + cb(kq, "kq_scaled_alibi", il); + } - kq = ggml_add(ctx, kq, kq_mask); - cb(kq, "kq_masked", il); + kq = ggml_add(ctx, kq, kq_mask); + cb(kq, "kq_masked", il); - kq = ggml_soft_max(ctx, kq); - cb(kq, "kq_soft_max", il); + kq = ggml_soft_max(ctx, kq); + cb(kq, "kq_soft_max", il); + } else { + kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale); + cb(kq, "kq_soft_max_ext", il); + } // split cached v into n_head heads struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v, + ggml_view_3d(ctx, kv.v_l[il], n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv.v)*n_ctx, - ggml_element_size(kv.v)*n_ctx*n_embd_head, - ggml_element_size(kv.v)*n_ctx*n_embd_gqa*il); + ggml_element_size(kv.v_l[il])*n_ctx, + ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head, + 0); cb(v, "v", il); struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); @@ -3501,6 +4146,8 @@ struct llm_build_context { const int64_t n_head_kv; const int64_t n_embd_head; const int64_t n_embd_gqa; + const int64_t n_expert; + const int64_t n_expert_used; const float freq_base; const float freq_scale; @@ -3520,7 +4167,7 @@ struct llm_build_context { const llm_build_cb & cb; - llama_buffer & buf_compute; + std::vector & buf_compute_meta; struct ggml_context * ctx0 = nullptr; @@ -3530,33 +4177,35 @@ struct llm_build_context { const llama_batch & batch, const llm_build_cb & cb, bool worst_case) : - model (lctx.model), - hparams (model.hparams), - cparams (lctx.cparams), - batch (batch), - kv_self (lctx.kv_self), - n_embd (hparams.n_embd), - n_layer (hparams.n_layer), - n_ctx (cparams.n_ctx), - n_head (hparams.n_head), - n_head_kv (hparams.n_head_kv), - n_embd_head (hparams.n_embd_head()), - n_embd_gqa (hparams.n_embd_gqa()), - freq_base (cparams.rope_freq_base), - freq_scale (cparams.rope_freq_scale), - ext_factor (cparams.yarn_ext_factor), - attn_factor (cparams.yarn_attn_factor), - beta_fast (cparams.yarn_beta_fast), - beta_slow (cparams.yarn_beta_slow), - norm_eps (hparams.f_norm_eps), - norm_rms_eps (hparams.f_norm_rms_eps), - n_tokens (batch.n_tokens), - n_kv (worst_case ? n_ctx : kv_self.n), - kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), - n_orig_ctx (cparams.n_yarn_orig_ctx), - do_rope_shift (worst_case || kv_self.has_shift), - cb (cb), - buf_compute (lctx.buf_compute) { + model (lctx.model), + hparams (model.hparams), + cparams (lctx.cparams), + batch (batch), + kv_self (lctx.kv_self), + n_embd (hparams.n_embd), + n_layer (hparams.n_layer), + n_ctx (cparams.n_ctx), + n_head (hparams.n_head), + n_head_kv (hparams.n_head_kv), + n_embd_head (hparams.n_embd_head()), + n_embd_gqa (hparams.n_embd_gqa()), + n_expert (hparams.n_expert), + n_expert_used (hparams.n_expert_used), + freq_base (cparams.rope_freq_base), + freq_scale (cparams.rope_freq_scale), + ext_factor (cparams.yarn_ext_factor), + attn_factor (cparams.yarn_attn_factor), + beta_fast (cparams.yarn_beta_fast), + beta_slow (cparams.yarn_beta_slow), + norm_eps (hparams.f_norm_eps), + norm_rms_eps (hparams.f_norm_rms_eps), + n_tokens (batch.n_tokens), + n_kv (worst_case ? n_ctx : kv_self.n), + kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), + n_orig_ctx (cparams.n_yarn_orig_ctx), + do_rope_shift (worst_case || kv_self.has_shift), + cb (cb), + buf_compute_meta (lctx.buf_compute_meta) { GGML_ASSERT(!!kv_self.ctx); // all initializations should be done in init() @@ -3564,8 +4213,8 @@ struct llm_build_context { void init() { struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, + /*.mem_size =*/ buf_compute_meta.size(), + /*.mem_buffer =*/ buf_compute_meta.data(), /*.no_alloc =*/ true, }; @@ -3580,7 +4229,7 @@ struct llm_build_context { } struct ggml_cgraph * build_llama() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); GGML_ASSERT(n_embd_head == hparams.n_rot); @@ -3594,10 +4243,6 @@ struct llm_build_context { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); cb(inp_pos, "inp_pos", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -3621,12 +4266,24 @@ struct llm_build_context { // compute Q and K and RoPE them struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, @@ -3644,9 +4301,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, - model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -3654,7 +4311,7 @@ struct llm_build_context { cb(ffn_inp, "ffn_inp", il); // feed-forward network - { + if (model.layers[il].ffn_gate_inp == nullptr) { cur = llm_build_norm(ctx0, ffn_inp, hparams, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, cb, il); @@ -3666,6 +4323,69 @@ struct llm_build_context { model.layers[il].ffn_down, NULL, LLM_FFN_SILU, LLM_FFN_PAR, cb, il); cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] + cb(probs, "ffn_moe_probs", il); + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + + ggml_tensor * weights = ggml_get_rows(ctx0, + ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); + cb(weights, "ffn_moe_weights", il); + + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] + + ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] + cb(weights, "ffn_moe_weights_norm", il); + + // compute expert outputs + ggml_tensor * moe_out = nullptr; + + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert; + + ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur); + cb(cur_up, "ffn_moe_up", il); + + ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exp, n_expert, selected_experts, i, cur); + cb(cur_gate, "ffn_moe_gate", il); + + cur_gate = ggml_silu(ctx0, cur_gate); + cb(cur_gate, "ffn_moe_silu", il); + + cur_expert = ggml_mul(ctx0, cur_up, cur_gate); // [n_tokens, n_embd] + cb(cur_expert, "ffn_moe_gate_par", il); + + cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_expert); // [n_tokens, n_embd] + cb(cur_expert, "ffn_moe_down", il); + + cur_expert = ggml_mul(ctx0, cur_expert, + ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0])); + cb(cur_expert, "ffn_moe_weighted", il); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx0, moe_out, cur_expert); + cb(moe_out, "ffn_moe_out", il); + } + } + + cur = moe_out; } cur = ggml_add(ctx0, cur, ffn_inp); @@ -3692,7 +4412,7 @@ struct llm_build_context { } struct ggml_cgraph * build_baichuan() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -3704,10 +4424,6 @@ struct llm_build_context { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); cb(inp_pos, "inp_pos", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -3764,9 +4480,9 @@ struct llm_build_context { // apply ALiBi for 13B model const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -3812,7 +4528,7 @@ struct llm_build_context { } struct ggml_cgraph * build_falcon() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -3824,10 +4540,6 @@ struct llm_build_context { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); cb(inp_pos, "inp_pos", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -3888,9 +4600,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -3934,7 +4646,7 @@ struct llm_build_context { } struct ggml_cgraph * build_starcoder() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * pos; @@ -3947,10 +4659,6 @@ struct llm_build_context { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); cb(inp_pos, "inp_pos", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -3988,9 +4696,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4033,7 +4741,7 @@ struct llm_build_context { } struct ggml_cgraph * build_persimmon() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); const int64_t n_rot = n_embd_head / 2; @@ -4043,13 +4751,11 @@ struct llm_build_context { inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); cb(inpL, "imp_embd", -1); + // inp_pos - contains the positions struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); cb(inp_pos, "inp_pos", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -4178,7 +4884,7 @@ struct llm_build_context { struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass); cb(Kcur, "Kcur", il); - struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 1, 2, 0, 3)); + struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3)); cb(Q, "Q", il); Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3)); @@ -4195,9 +4901,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); // TODO: not tested, could be broken - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, model.layers[il].bo, - Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + Q, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4243,7 +4949,7 @@ struct llm_build_context { } struct ggml_cgraph * build_refact() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4251,10 +4957,6 @@ struct llm_build_context { inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); cb(inpL, "inp_embd", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -4286,9 +4988,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4334,7 +5036,7 @@ struct llm_build_context { } struct ggml_cgraph * build_bloom() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4342,10 +5044,6 @@ struct llm_build_context { inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); cb(inpL, "inp_embd", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -4383,9 +5081,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4428,7 +5126,7 @@ struct llm_build_context { } struct ggml_cgraph * build_mpt() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -4436,10 +5134,6 @@ struct llm_build_context { inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); cb(inpL, "inp_embd", -1); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - cb(KQ_scale, "KQ_scale", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); cb(KQ_mask, "KQ_mask", -1); @@ -4477,9 +5171,9 @@ struct llm_build_context { llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - cur = llm_build_kqv(ctx0, hparams, kv_self, + cur = llm_build_kqv(ctx0, model, hparams, kv_self, model.layers[il].wo, NULL, - Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il); + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4498,27 +5192,357 @@ struct llm_build_context { cur = llm_build_ffn(ctx0, cur, model.layers[il].ffn_up, NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + NULL, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_stablelm() { + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_qwen() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_custom( + ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_phi2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * attn_norm_output; + struct ggml_tensor * ffn_output; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + attn_norm_output = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(attn_norm_output, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_custom( + ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + // with phi2, we scale the Q to avoid precision issues + // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 + Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il); + cb(cur, "kqv_out", il); + } + + // FF + { + ffn_output = llm_build_ffn(ctx0, attn_norm_output, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); + cb(ffn_output, "ffn_out", il); } - cur = ggml_add(ctx0, cur, ffn_inp); + cur = ggml_add(ctx0, cur, ffn_output); + cb(cur, "l_out", il); + + cur = ggml_add(ctx0, cur, inpL); cb(cur, "l_out", il); - // input for next layer inpL = cur; } - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, + cur = llm_build_norm(ctx0, inpL, hparams, model.output_norm, - NULL, + model.output_norm_b, LLM_NORM, cb, -1); cb(cur, "result_norm", -1); cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output_no_bias", -1); + + cur = ggml_add(ctx0, cur, model.output_b); cb(cur, "result_output", -1); ggml_build_forward_expand(gf, cur); @@ -4535,10 +5559,10 @@ struct llm_build_context { enum llm_offload_func_e { OFFLOAD_FUNC_NOP, OFFLOAD_FUNC, - OFFLOAD_FUNC_KQ, - OFFLOAD_FUNC_V, + OFFLOAD_FUNC_FRC, // force offload + OFFLOAD_FUNC_KQV, OFFLOAD_FUNC_NR, - OFFLOAD_FUNC_EMB, + OFFLOAD_FUNC_EMB, // embeddings OFFLOAD_FUNC_OUT, }; @@ -4622,11 +5646,11 @@ static const std::unordered_map k_offload_map //{ "inp_embd", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel { "pos_embd", OFFLOAD_FUNC_NR }, - { "inp_pos", OFFLOAD_FUNC_KQ }, // this is often used for KQ ops (e.g. rope) - { "KQ_scale", OFFLOAD_FUNC_KQ }, - { "KQ_mask", OFFLOAD_FUNC_KQ }, - { "K_shift", OFFLOAD_FUNC_KQ }, - { "K_shifted", OFFLOAD_FUNC_KQ }, + { "inp_pos", OFFLOAD_FUNC_FRC }, // this is often used for KQ ops (e.g. rope) + { "KQ_mask", OFFLOAD_FUNC_FRC }, + { "K_shift", OFFLOAD_FUNC_FRC }, + + { "K_shifted", OFFLOAD_FUNC }, { "inp_norm", OFFLOAD_FUNC_NR }, { "inp_norm_w", OFFLOAD_FUNC_NR }, @@ -4639,37 +5663,38 @@ static const std::unordered_map k_offload_map { "attn_norm", OFFLOAD_FUNC }, { "attn_norm_2", OFFLOAD_FUNC }, - { "wqkv", OFFLOAD_FUNC_KQ }, - { "bqkv", OFFLOAD_FUNC_KQ }, - { "wqkv_clamped", OFFLOAD_FUNC_KQ }, - - { "tmpk", OFFLOAD_FUNC_KQ }, - { "tmpq", OFFLOAD_FUNC_KQ }, - { "tmpv", OFFLOAD_FUNC_V }, - { "Kcur", OFFLOAD_FUNC_KQ }, - { "Qcur", OFFLOAD_FUNC_KQ }, - { "Vcur", OFFLOAD_FUNC_V }, - - { "krot", OFFLOAD_FUNC_KQ }, - { "qrot", OFFLOAD_FUNC_KQ }, - { "kpass", OFFLOAD_FUNC_KQ }, - { "qpass", OFFLOAD_FUNC_KQ }, - { "krotated", OFFLOAD_FUNC_KQ }, - { "qrotated", OFFLOAD_FUNC_KQ }, - - { "q", OFFLOAD_FUNC_KQ }, - { "k", OFFLOAD_FUNC_KQ }, - { "kq", OFFLOAD_FUNC_KQ }, - { "kq_scaled", OFFLOAD_FUNC_KQ }, - { "kq_scaled_alibi", OFFLOAD_FUNC_KQ }, - { "kq_masked", OFFLOAD_FUNC_KQ }, - { "kq_soft_max", OFFLOAD_FUNC_V }, - { "v", OFFLOAD_FUNC_V }, - { "kqv", OFFLOAD_FUNC_V }, - { "kqv_merged", OFFLOAD_FUNC_V }, - { "kqv_merged_cont", OFFLOAD_FUNC_V }, - { "kqv_wo", OFFLOAD_FUNC_V }, - { "kqv_out", OFFLOAD_FUNC_V }, + { "wqkv", OFFLOAD_FUNC_KQV }, + { "bqkv", OFFLOAD_FUNC_KQV }, + { "wqkv_clamped", OFFLOAD_FUNC_KQV }, + + { "tmpk", OFFLOAD_FUNC_KQV }, + { "tmpq", OFFLOAD_FUNC_KQV }, + { "tmpv", OFFLOAD_FUNC_KQV }, + { "Kcur", OFFLOAD_FUNC_KQV }, + { "Qcur", OFFLOAD_FUNC_KQV }, + { "Vcur", OFFLOAD_FUNC_KQV }, + + { "krot", OFFLOAD_FUNC_KQV }, + { "qrot", OFFLOAD_FUNC_KQV }, + { "kpass", OFFLOAD_FUNC_KQV }, + { "qpass", OFFLOAD_FUNC_KQV }, + { "krotated", OFFLOAD_FUNC_KQV }, + { "qrotated", OFFLOAD_FUNC_KQV }, + + { "q", OFFLOAD_FUNC_KQV }, + { "k", OFFLOAD_FUNC_KQV }, + { "kq", OFFLOAD_FUNC_KQV }, + { "kq_scaled", OFFLOAD_FUNC_KQV }, + { "kq_scaled_alibi", OFFLOAD_FUNC_KQV }, + { "kq_masked", OFFLOAD_FUNC_KQV }, + { "kq_soft_max", OFFLOAD_FUNC_KQV }, + { "kq_soft_max_ext", OFFLOAD_FUNC_KQV }, + { "v", OFFLOAD_FUNC_KQV }, + { "kqv", OFFLOAD_FUNC_KQV }, + { "kqv_merged", OFFLOAD_FUNC_KQV }, + { "kqv_merged_cont", OFFLOAD_FUNC_KQV }, + { "kqv_wo", OFFLOAD_FUNC_KQV }, + { "kqv_out", OFFLOAD_FUNC_KQV }, { "ffn_inp", OFFLOAD_FUNC }, { "ffn_norm", OFFLOAD_FUNC }, @@ -4688,9 +5713,24 @@ static const std::unordered_map k_offload_map { "ffn_relu", OFFLOAD_FUNC }, { "ffn_sqr(relu)", OFFLOAD_FUNC }, + { "ffn_moe_logits", OFFLOAD_FUNC }, + { "ffn_moe_probs", OFFLOAD_FUNC }, + { "ffn_moe_argsort", OFFLOAD_FUNC }, + { "ffn_moe_weights", OFFLOAD_FUNC }, + { "ffn_moe_weights_sum", OFFLOAD_FUNC }, + { "ffn_moe_weights_norm", OFFLOAD_FUNC }, + { "ffn_moe_weighted", OFFLOAD_FUNC }, + { "ffn_moe_up", OFFLOAD_FUNC }, + { "ffn_moe_gate", OFFLOAD_FUNC }, + { "ffn_moe_silu", OFFLOAD_FUNC }, + { "ffn_moe_gate_par", OFFLOAD_FUNC }, + { "ffn_moe_down", OFFLOAD_FUNC }, + { "ffn_moe_out", OFFLOAD_FUNC }, + { "l_out", OFFLOAD_FUNC }, { "result_norm", OFFLOAD_FUNC_EMB }, + { "result_output_no_bias", OFFLOAD_FUNC_EMB }, { "result_output", OFFLOAD_FUNC_OUT }, }; @@ -4708,11 +5748,10 @@ static struct ggml_cgraph * llama_build_graph( bool alloc_inp_tokens = false; bool alloc_inp_embd = false; bool alloc_inp_pos = false; - bool alloc_inp_KQ_scale = false; bool alloc_inp_KQ_mask = false; bool alloc_inp_K_shift = false; -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) const bool do_offload = true; #else const bool do_offload = true; // TODO: set to false after finishing refactoring @@ -4740,7 +5779,7 @@ static struct ggml_cgraph * llama_build_graph( if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) { const int64_t n_tokens = cur->ne[0]; - memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur)); + ggml_backend_tensor_set(cur, batch.token, 0, n_tokens*ggml_element_size(cur)); } alloc_inp_tokens = true; @@ -4753,7 +5792,7 @@ static struct ggml_cgraph * llama_build_graph( const int64_t n_embd = cur->ne[0]; const int64_t n_tokens = cur->ne[1]; - memcpy(cur->data, batch.embd, n_tokens*n_embd*ggml_element_size(cur)); + ggml_backend_tensor_set(cur, batch.embd, 0, n_tokens*n_embd*ggml_element_size(cur)); } alloc_inp_embd = true; @@ -4765,27 +5804,13 @@ static struct ggml_cgraph * llama_build_graph( if (!ggml_allocr_is_measure(lctx.alloc) && batch.pos) { const int64_t n_tokens = cur->ne[0]; - int32_t * data = (int32_t *) cur->data; - - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } + static_assert(std::is_same::value, "llama_pos must be int32_t"); + ggml_backend_tensor_set(cur, batch.pos, 0, n_tokens*ggml_element_size(cur)); } alloc_inp_pos = true; } - if (!alloc_inp_KQ_scale && strcmp(name, "KQ_scale") == 0) { - ggml_allocr_alloc(lctx.alloc, cur); - - if (!ggml_allocr_is_measure(lctx.alloc)) { - const int64_t n_embd_head = model.hparams.n_embd_head(); - ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head))); - } - - alloc_inp_KQ_scale = true; - } - if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) { ggml_allocr_alloc(lctx.alloc, cur); @@ -4793,8 +5818,13 @@ static struct ggml_cgraph * llama_build_graph( const int64_t n_kv = cur->ne[0]; const int64_t n_tokens = cur->ne[1]; - float * data = (float *) cur->data; - memset(data, 0, ggml_nbytes(cur)); + float * data; + if (ggml_backend_buffer_is_host(cur->buffer)) { + data = (float *) cur->data; + } else { + lctx.buf_copy.resize(ggml_nbytes(cur)); + data = (float *) lctx.buf_copy.data(); + } for (int h = 0; h < 1; ++h) { for (int j = 0; j < n_tokens; ++j) { @@ -4802,12 +5832,20 @@ static struct ggml_cgraph * llama_build_graph( const llama_seq_id seq_id = batch.seq_id[j][0]; for (int i = 0; i < n_kv; ++i) { + float f; if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; + f = -INFINITY; + } else { + f = 0; } + data[h*(n_kv*n_tokens) + j*n_kv + i] = f; } } } + + if (data != cur->data) { + ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); + } } alloc_inp_KQ_mask = true; @@ -4819,11 +5857,21 @@ static struct ggml_cgraph * llama_build_graph( if (!ggml_allocr_is_measure(lctx.alloc)) { const int64_t n_ctx = cur->ne[0]; - int32_t * data = (int32_t *) cur->data; + int32_t * data; + if (ggml_backend_buffer_is_host(cur->buffer)) { + data = (int32_t *) cur->data; + } else { + lctx.buf_copy.resize(ggml_nbytes(cur)); + data = (int32_t *) lctx.buf_copy.data(); + } for (int i = 0; i < n_ctx; ++i) { data[i] = lctx.kv_self.cells[i].delta; } + + if (data != cur->data) { + ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); + } } alloc_inp_K_shift = true; @@ -4860,16 +5908,16 @@ static struct ggml_cgraph * llama_build_graph( static const std::unordered_map> k_offload_func_name = { { OFFLOAD_FUNC_NOP, "CPU" }, { OFFLOAD_FUNC_OUT, "CPU" }, -#ifdef GGML_USE_CUBLAS - { OFFLOAD_FUNC, "GPU (CUDA)" }, - { OFFLOAD_FUNC_KQ, "GPU (CUDA) KQ" }, - { OFFLOAD_FUNC_V, "GPU (CUDA) V" }, - { OFFLOAD_FUNC_NR, "GPU (CUDA) NR" }, +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + { OFFLOAD_FUNC, "GPU (CUDA)" }, + { OFFLOAD_FUNC_FRC, "GPU (CUDA) FRC" }, + { OFFLOAD_FUNC_KQV, "GPU (CUDA) KQV" }, + { OFFLOAD_FUNC_NR, "GPU (CUDA) NR" }, { OFFLOAD_FUNC_EMB, "GPU (CUDA) EMB" }, #else { OFFLOAD_FUNC, "CPU" }, - { OFFLOAD_FUNC_KQ, "CPU" }, - { OFFLOAD_FUNC_V, "CPU" }, + { OFFLOAD_FUNC_FRC, "CPU" }, + { OFFLOAD_FUNC_KQV, "CPU" }, { OFFLOAD_FUNC_NR, "CPU" }, { OFFLOAD_FUNC_EMB, "CPU" }, #endif // GGML_USE_CUBLAS @@ -4902,18 +5950,23 @@ static struct ggml_cgraph * llama_build_graph( } } break; - case OFFLOAD_FUNC_NR: - if (n_gpu_layers <= n_layer + 0) { + case OFFLOAD_FUNC_FRC: + if (!lctx.cparams.offload_kqv) { func_e = OFFLOAD_FUNC_NOP; - } - break; - case OFFLOAD_FUNC_V: - if (n_gpu_layers <= n_layer + 1) { + } break; + case OFFLOAD_FUNC_KQV: + if (!lctx.cparams.offload_kqv) { func_e = OFFLOAD_FUNC_NOP; + } else { + if (n_gpu_layers < n_layer) { + if (il < i_gpu_start) { + func_e = OFFLOAD_FUNC_NOP; + } + } } break; - case OFFLOAD_FUNC_KQ: - if (n_gpu_layers <= n_layer + 2) { + case OFFLOAD_FUNC_NR: + if (n_gpu_layers <= n_layer + 0) { func_e = OFFLOAD_FUNC_NOP; } break; @@ -4928,7 +5981,7 @@ static struct ggml_cgraph * llama_build_graph( offload_func_t func = ggml_offload_nop; // this is needed for compatibility with Metal for example -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) static offload_func_t ggml_offload_gpu = ggml_cuda_assign_buffers_no_alloc; #else static offload_func_t ggml_offload_gpu = ggml_offload_nop; @@ -4938,8 +5991,8 @@ static struct ggml_cgraph * llama_build_graph( case OFFLOAD_FUNC_NOP: case OFFLOAD_FUNC_OUT: func = ggml_offload_nop; break; case OFFLOAD_FUNC: - case OFFLOAD_FUNC_KQ: - case OFFLOAD_FUNC_V: + case OFFLOAD_FUNC_KQV: + case OFFLOAD_FUNC_FRC: case OFFLOAD_FUNC_NR: case OFFLOAD_FUNC_EMB: func = ggml_offload_gpu; break; default: GGML_ASSERT(false); @@ -4994,6 +6047,18 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_mpt(); } break; + case LLM_ARCH_STABLELM: + { + result = llm.build_stablelm(); + } break; + case LLM_ARCH_QWEN: + { + result = llm.build_qwen(); + } break; + case LLM_ARCH_PHI2: + { + result = llm.build_phi2(); + } break; default: GGML_ASSERT(false); } @@ -5071,7 +6136,7 @@ static int llama_decode_internal( const int64_t n_embd = hparams.n_embd; const int64_t n_vocab = hparams.n_vocab; - // helpers for smoother batch API transistion + // helpers for smoother batch API transition // after deprecating the llama_eval calls, these will be removed std::vector pos; @@ -5103,6 +6168,12 @@ static int llama_decode_internal( batch.seq_id = seq_id_arr.data(); } + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (kv_self.head > kv_self.used + 2*n_tokens) { + kv_self.head = 0; + } + if (!llama_kv_cache_find_slot(kv_self, batch)) { return 1; } @@ -5110,10 +6181,10 @@ static int llama_decode_internal( // a heuristic, to avoid attending the full cache if it is not yet utilized // after enough generations, the benefit from this heuristic disappears // if we start defragmenting the cache, the benefit from this will be more important - //kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)); // TODO: this might be better for CUDA? - kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self))); + kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); + //kv_self.n = llama_kv_cache_cell_max(kv_self); - //printf("kv_self.n = %d\n", kv_self.n); + //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); ggml_allocr_reset(lctx.alloc); @@ -5121,18 +6192,23 @@ static int llama_decode_internal( ggml_allocr_alloc_graph(lctx.alloc, gf); - struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; - struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2]; - - GGML_ASSERT(strcmp(res->name, "result_output") == 0); - GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0); + // the output is always the last tensor in the graph + struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; + GGML_ASSERT(strcmp(res->name, "result_output") == 0); + // the embeddings could be the second to last tensor, or the third to last tensor + struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2]; + if (strcmp(embeddings->name, "result_norm") != 0) { + embeddings = gf->nodes[gf->n_nodes - 3]; + GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0); + } -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + char * buf_alloc_base = (char *)ggml_backend_buffer_get_base(lctx.buf_alloc); for (int i = 0; i < gf->n_leafs; i++) { ggml_tensor * node = gf->leafs[i]; if (node->backend == GGML_BACKEND_GPU && node->extra == NULL) { - ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data); + ggml_cuda_assign_scratch_offset(node, (char *)node->data - buf_alloc_base); ggml_cuda_copy_to_device(node); } } @@ -5140,7 +6216,7 @@ static int llama_decode_internal( for (int i = 0; i < gf->n_nodes; i++) { ggml_tensor * node = gf->nodes[i]; if (node->backend == GGML_BACKEND_GPU && node->extra == NULL) { - ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data); + ggml_cuda_assign_scratch_offset(node, (char *)node->data - buf_alloc_base); } } @@ -5162,36 +6238,28 @@ static int llama_decode_internal( n_threads = std::min(4, n_threads); } - // If all tensors can be run on the GPU then using more than 1 thread is detrimental. - const bool full_offload_supported = - model.arch == LLM_ARCH_LLAMA || - model.arch == LLM_ARCH_BAICHUAN || - model.arch == LLM_ARCH_FALCON || - model.arch == LLM_ARCH_REFACT || - model.arch == LLM_ARCH_MPT; - - const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3; - if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) { + const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 1; + if (ggml_cpu_has_cublas() && fully_offloaded) { n_threads = 1; } -#if GGML_USE_MPI +#ifdef GGML_USE_MPI const int64_t n_layer = hparams.n_layer; ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer); #endif #ifdef GGML_USE_METAL - if (lctx.ctx_metal) { - ggml_metal_set_n_cb (lctx.ctx_metal, n_threads); - ggml_metal_graph_compute(lctx.ctx_metal, gf); - } else { - ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads); + if (ggml_backend_is_metal(lctx.backend)) { + ggml_backend_metal_set_n_cb(lctx.backend, n_threads); } -#else - ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads); #endif -#if GGML_USE_MPI + if (ggml_backend_is_cpu(lctx.backend)) { + ggml_backend_cpu_set_n_threads(lctx.backend, n_threads); + } + ggml_backend_graph_compute(lctx.backend, gf); + +#ifdef GGML_USE_MPI ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer); #endif @@ -5229,20 +6297,37 @@ static int llama_decode_internal( { auto & logits_out = lctx.logits; +#ifndef NDEBUG + auto & logits_valid = lctx.logits_valid; + logits_valid.clear(); + logits_valid.resize(n_tokens); + + logits_out.clear(); +#endif + if (batch.logits) { logits_out.resize(n_vocab * n_tokens); for (uint32_t i = 0; i < n_tokens; i++) { if (batch.logits[i] == 0) { continue; } - memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab); + ggml_backend_tensor_get(res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float)); +#ifndef NDEBUG + logits_valid[i] = true; +#endif } } else if (lctx.logits_all) { logits_out.resize(n_vocab * n_tokens); - memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*n_vocab*n_tokens); + ggml_backend_tensor_get(res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float)); +#ifndef NDEBUG + std::fill(logits_valid.begin(), logits_valid.end(), true); +#endif } else { logits_out.resize(n_vocab); - memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(n_tokens - 1)), sizeof(float)*n_vocab); + ggml_backend_tensor_get(res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float)); +#ifndef NDEBUG + logits_valid[0] = true; +#endif } } @@ -5251,7 +6336,7 @@ static int llama_decode_internal( auto & embedding_out = lctx.embedding; embedding_out.resize(n_embd); - memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(n_tokens - 1)), sizeof(float)*n_embd); + ggml_backend_tensor_get(embeddings, embedding_out.data(), (n_embd*(n_tokens - 1))*sizeof(float), n_embd*sizeof(float)); } // measure the performance only for the single-token evals @@ -5852,12 +6937,12 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< // loop over the text while (true) { - // find the first occurence of a given special token in this fragment + // find the first occurrence of a given special token in this fragment // passing offset argument only limit the "search area" but match coordinates // are still relative to the source full raw_text auto match = raw_text->find(special_token, raw_text_base_offset); - // no occurences found, stop processing this fragment for a given special token + // no occurrences found, stop processing this fragment for a given special token if (match == std::string::npos) break; // check if match is within bounds of offset <-> length @@ -5960,7 +7045,10 @@ static std::vector llama_tokenize_internal(const llama_vocab & // by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer // and passing 'add space prefix' as bool argument // - auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length); + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + if (&fragment == &fragment_buffer.front()) { + raw_text = " " + raw_text; // prefix with space if the first token is not special + } #ifdef PRETOKENIZERDEBUG fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); @@ -6026,11 +7114,13 @@ struct llama_grammar_candidate { // Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as // pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`. static std::pair, llama_partial_utf8> decode_utf8( - const char * src, + const std::string & src, llama_partial_utf8 partial_start) { static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 }; - const char * pos = src; + const char * pos = src.c_str(); std::vector code_points; + // common english strings have the same number of codepoints and bytes. `+ 1` for the terminating 0. + code_points.reserve(src.size() + 1); uint32_t value = partial_start.value; int n_remain = partial_start.n_remain; @@ -6634,6 +7724,7 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c // Replace the data in candidates with the new_candidates data std::copy(new_candidates.begin(), new_candidates.end(), candidates->data); candidates->size = new_candidates.size(); + candidates->sorted = false; if (ctx) { ctx->t_sample_us += ggml_time_us() - t_start_sample_us; @@ -6718,7 +7809,9 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c const llama_token eos = llama_token_eos(&ctx->model); std::vector, llama_partial_utf8>> candidates_decoded; + candidates_decoded.reserve(candidates->size); std::vector candidates_grammar; + candidates_grammar.reserve(candidates->size); for (size_t i = 0; i < candidates->size; ++i) { const llama_token id = candidates->data[i].id; @@ -6730,7 +7823,7 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c } else if (piece.empty() || piece[0] == 0) { candidates->data[i].logit = -INFINITY; } else { - candidates_decoded.push_back(decode_utf8(piece.c_str(), grammar->partial_utf8)); + candidates_decoded.push_back(decode_utf8(piece, grammar->partial_utf8)); candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second }); } } @@ -6937,7 +8030,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar const std::string piece = llama_token_to_piece(ctx, token); // Note terminating 0 in decoded string - const auto decoded = decode_utf8(piece.c_str(), grammar->partial_utf8); + const auto decoded = decode_utf8(piece, grammar->partial_utf8); const auto & code_points = decoded.first; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it); @@ -7048,7 +8141,7 @@ struct llama_beam_search_data { } // Min-heaps are used to efficiently collect the top-k elements (k=n_beams). - // The repetative patterns below reflect the 2 stages of heaps: + // The repetitive patterns below reflect the 2 stages of heaps: // * Gather elements until the vector is full, then call std::make_heap() on it. // * If the heap is full and a new element is found that should be included, pop the // least element to the back(), replace it with the new, then push it into the heap. @@ -7201,12 +8294,6 @@ void llama_beam_search(llama_context * ctx, // quantization // -template -struct no_init { - T value; - no_init() { /* do nothing */ } -}; - struct quantize_state_internal { const llama_model & model; const llama_model_quantize_params * params; @@ -7255,18 +8342,21 @@ static void llama_convert_tensor_internal( return; } - auto block_size = tensor->type == GGML_TYPE_F16 ? 1 : (size_t)ggml_blck_size(tensor->type); - auto block_size_bytes = ggml_type_size(tensor->type); + size_t block_size = tensor->type == GGML_TYPE_F16 ? 1 : (size_t)ggml_blck_size(tensor->type); + size_t block_size_bytes = ggml_type_size(tensor->type); GGML_ASSERT(nelements % block_size == 0); - auto nblocks = nelements / block_size; - auto blocks_per_thread = nblocks / nthread; - auto spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count + size_t nblocks = nelements / block_size; + size_t blocks_per_thread = nblocks / nthread; + size_t spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count + + size_t in_buff_offs = 0; + size_t out_buff_offs = 0; - for (auto tnum = 0, in_buff_offs = 0, out_buff_offs = 0; tnum < nthread; tnum++) { - auto thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread - auto thr_elems = thr_blocks * block_size; // number of elements for this thread - auto thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread + for (int tnum = 0; tnum < nthread; tnum++) { + size_t thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread + size_t thr_elems = thr_blocks * block_size; // number of elements for this thread + size_t thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) { if (typ == GGML_TYPE_F16) { @@ -7283,11 +8373,9 @@ static void llama_convert_tensor_internal( workers.clear(); } -static ggml_type get_k_quant_type( - quantize_state_internal & qs, - ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype -) { +static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) { const std::string name = ggml_get_name(tensor); + // TODO: avoid hardcoded tensor names - use the TN_* constants const llm_arch arch = qs.model.arch; const auto tn = LLM_TN(arch); @@ -7321,7 +8409,18 @@ static ggml_type get_k_quant_type( // nearly negligible increase in model size by quantizing this tensor with more bits: if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K; } + if (qs.model.hparams.n_expert == 8) { + // for the 8-expert model, bumping this to Q8_0 trades just ~128MB + // TODO: explore better strategies + new_type = GGML_TYPE_Q8_0; + } ++qs.i_attention_wv; + } else if (name.find("attn_k.weight") != std::string::npos) { + if (qs.model.hparams.n_expert == 8) { + // for the 8-expert model, bumping this to Q8_0 trades just ~128MB + // TODO: explore better strategies + new_type = GGML_TYPE_Q8_0; + } } else if (name.find("ffn_down.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { @@ -7436,10 +8535,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s constexpr bool use_mmap = false; #endif - llama_model_loader ml(fname_inp, use_mmap); - if (ml.use_mmap) { - ml.mapping.reset(new llama_mmap(&ml.file, /* prefetch */ 0, ggml_is_numa())); - } + llama_model_loader ml(fname_inp, use_mmap, NULL); + ml.init_mapping(false); // no prefetching? llama_model model; llm_load_arch(ml, model); @@ -7530,10 +8627,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'? // quantize only 2D tensors - quantize &= (tensor->n_dims == 2); + quantize &= (ggml_n_dims(tensor) == 2); quantize &= params->quantize_output_tensor || name != "output.weight"; quantize &= !params->only_copy; + // do not quantize expert gating tensors + quantize &= name.find("ffn_gate_inp.weight") == std::string::npos; + enum ggml_type new_type; void * new_data; size_t new_size; @@ -7612,7 +8712,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s workers.clear(); } - LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); + LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); int64_t tot_count = 0; for (size_t i = 0; i < hist_cur.size(); i++) { hist_all[i] += hist_cur[i]; @@ -7682,74 +8782,63 @@ static int llama_apply_lora_from_file_internal( const int64_t t_start_lora_us = ggml_time_us(); - auto fin = std::ifstream(path_lora, std::ios::binary); - if (!fin) { - LLAMA_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_lora); - return 1; - } + llama_file fin(path_lora, "rb"); // verify magic and version { - uint32_t magic; - fin.read((char *) &magic, sizeof(magic)); - uint32_t format_version; - fin.read((char *) &format_version, sizeof(format_version)); + uint32_t magic = fin.read_u32(); + if (magic != LLAMA_FILE_MAGIC_GGLA) { + LLAMA_LOG_ERROR("%s: bad file magic\n", __func__); + return 1; + } + uint32_t format_version = fin.read_u32(); if (format_version != 1) { LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ ); return 1; } } - int32_t lora_r; - int32_t lora_alpha; - fin.read((char *) &lora_r, sizeof(lora_r)); - fin.read((char *) &lora_alpha, sizeof(lora_alpha)); + int32_t lora_r = fin.read_u32(); + int32_t lora_alpha = fin.read_u32(); float scaling = scale * (float)lora_alpha / (float)lora_r; LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling); + // create a name -> tensor map of the model to accelerate lookups + // find the max tensor size to estimate the required temporary buffer size + size_t max_tensor_size = 0; + std::unordered_map model_tensors; + for (const auto & kv : model.tensors_by_name) { + model_tensors.insert(kv); + size_t f32_size = ggml_nelements(kv.second) * sizeof(float); + max_tensor_size = std::max(max_tensor_size, f32_size); + } + // create a temporary ggml context to store the lora tensors - // todo: calculate size from biggest possible tensor - std::vector lora_buf(1024ull * 1024ull * 1024ull); + // TODO: use ggml-alloc + size_t lora_ctx_size = max_tensor_size * 3; + LLAMA_LOG_INFO("%s: allocating %.f MB for lora temporary buffer\n", __func__, lora_ctx_size / 1024.0 / 1024.0); + std::vector lora_buf(lora_ctx_size); + struct ggml_init_params params; params.mem_size = lora_buf.size(); params.mem_buffer = lora_buf.data(); params.no_alloc = false; - ggml_context * lora_ctx = ggml_init(params); - std::unordered_map lora_tensors; + using unique_context = std::unique_ptr; - // create a name -> tensor map of the model to accelerate lookups - std::unordered_map model_tensors; - for (const auto & kv : model.tensors_by_name) { - model_tensors.insert(kv); - } + unique_context lora_ctx(nullptr, ggml_free); + lora_ctx.reset(ggml_init(params)); + std::unordered_map lora_tensors; // load base model std::unique_ptr ml; - ggml_context * base_ctx = NULL; - std::vector base_buf; - if (path_base_model) { - LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model); - ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true)); - - size_t ctx_size; - size_t mmapped_size; - ml->calc_sizes(ctx_size, mmapped_size); - base_buf.resize(ctx_size); - ggml_init_params base_params; - base_params.mem_size = base_buf.size(); - base_params.mem_buffer = base_buf.data(); - base_params.no_alloc = ml->use_mmap; - - base_ctx = ggml_init(base_params); - - // maybe this should in llama_model_loader - if (ml->use_mmap) { - ml->mapping.reset(new llama_mmap(&ml->file, /* prefetch */ 0, ggml_is_numa())); - } + if (path_base_model) { + LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model); + ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*kv_overrides*/ nullptr)); + ml->init_mapping(false); // no prefetching } // read tensors and apply @@ -7759,27 +8848,35 @@ static int llama_apply_lora_from_file_internal( std::vector work_buffer; while (true) { + if (fin.tell() == fin.size) { + // eof + break; + } + int32_t n_dims; - int32_t length; + int32_t name_len; int32_t ftype; - fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); - fin.read(reinterpret_cast(&length), sizeof(length)); - fin.read(reinterpret_cast(&ftype), sizeof(ftype)); - if (fin.eof()) { - break; + fin.read_raw(&n_dims, sizeof(n_dims)); + fin.read_raw(&name_len, sizeof(name_len)); + fin.read_raw(&ftype, sizeof(ftype)); + + if (n_dims != 1 && n_dims != 2) { + LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims); + return 1; } int32_t ne[2] = { 1, 1 }; for (int i = 0; i < n_dims; ++i) { - fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); + fin.read_raw(&ne[i], sizeof(ne[i])); } std::string name; { + GGML_ASSERT(name_len <= 1024); char buf[1024]; - fin.read(buf, length); - name = std::string(buf, length); + fin.read_raw(buf, name_len); + name = std::string(buf, name_len); } // check for lora suffix and get the type of tensor @@ -7793,7 +8890,7 @@ static int llama_apply_lora_from_file_internal( std::string lora_type = name.substr(pos + lora_suffix.length()); std::string base_name = name; base_name.erase(pos); - // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); + // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(), base_name.c_str(), lora_type.c_str()); if (model_tensors.find(base_name) == model_tensors.end()) { LLAMA_LOG_ERROR("%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); @@ -7812,22 +8909,15 @@ static int llama_apply_lora_from_file_internal( return false; } } - ggml_tensor * lora_tensor; - if (n_dims == 2) { - lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]); - } - else { - LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims); - return 1; - } - ggml_set_name(lora_tensor, "lora_tensor"); + ggml_tensor * lora_tensor = ggml_new_tensor_2d(lora_ctx.get(), wtype, ne[0], ne[1]); + ggml_set_name(lora_tensor, name.c_str()); // load tensor data - size_t offset = fin.tellg(); + size_t offset = fin.tell(); size_t tensor_data_size = ggml_nbytes(lora_tensor); offset = (offset + 31) & -32; - fin.seekg(offset); - fin.read((char*)lora_tensor->data, tensor_data_size); + fin.seek(offset, SEEK_SET); + fin.read_raw(lora_tensor->data, tensor_data_size); lora_tensors[name] = lora_tensor; @@ -7840,7 +8930,7 @@ static int llama_apply_lora_from_file_internal( offload_func_t offload_func = ggml_offload_nop; offload_func_t offload_func_force_inplace = ggml_offload_nop; -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) { if (dest_t->type != GGML_TYPE_F16) { throw std::runtime_error(format( @@ -7857,13 +8947,11 @@ static int llama_apply_lora_from_file_internal( // load from base model if (gguf_find_tensor(ctx_gguf, base_name.c_str()) < 0) { - // TODO: throw LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); return 1; } - // TODO: not tested!! maybe not working! - base_t = ml->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU); + base_t = ml->get_tensor_meta(base_name.c_str()); ml->load_data_for(base_t); } else { base_t = dest_t; @@ -7892,43 +8980,42 @@ static int llama_apply_lora_from_file_internal( } // w = w + BA*s - ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB); + ggml_tensor * BA = ggml_mul_mat(lora_ctx.get(), loraA, loraB); offload_func(BA); ggml_set_name(BA, "BA"); if (scaling != 1.0f) { - ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling); - ggml_set_name(scale_tensor, "scale_tensor"); - - BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor); + BA = ggml_scale_inplace(lora_ctx.get(), BA, scaling); offload_func(BA); ggml_set_name(BA, "BA_scaled"); } ggml_tensor * r; if (base_t == dest_t) { - r = ggml_add_inplace(lora_ctx, dest_t, BA); + r = ggml_add_inplace(lora_ctx.get(), dest_t, BA); offload_func_force_inplace(r); ggml_set_name(r, "r_add_inplace"); } else { - r = ggml_add(lora_ctx, base_t, BA); + r = ggml_add(lora_ctx.get(), base_t, BA); offload_func(r); ggml_set_name(r, "r_add"); - r = ggml_cpy(lora_ctx, r, dest_t); + r = ggml_cpy(lora_ctx.get(), r, dest_t); offload_func(r); ggml_set_name(r, "r_cpy"); } - struct ggml_cgraph * gf = ggml_new_graph(lora_ctx); + struct ggml_cgraph * gf = ggml_new_graph(lora_ctx.get()); ggml_build_forward_expand(gf, r); ggml_graph_compute_helper(work_buffer, gf, n_threads); + // the tensors in the adapter must be sorted such that loraA and loraB of the same tensor are next to each other + GGML_ASSERT(lora_tensors.size() == 2); + // we won't need these tensors again, reset the context to save memory - ggml_free(lora_ctx); - lora_ctx = ggml_init(params); + lora_ctx.reset(ggml_init(params)); lora_tensors.clear(); n_tensors++; @@ -7938,12 +9025,6 @@ static int llama_apply_lora_from_file_internal( } } - // TODO: this should be in a destructor, it will leak on failure - ggml_free(lora_ctx); - if (base_ctx) { - ggml_free(base_ctx); - } - const int64_t t_lora_us = ggml_time_us() - t_start_lora_us; LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0); @@ -7960,6 +9041,7 @@ struct llama_model_params llama_model_default_params() { /*.tensor_split =*/ nullptr, /*.progress_callback =*/ nullptr, /*.progress_callback_user_data =*/ nullptr, + /*.kv_overrides =*/ nullptr, /*.vocab_only =*/ false, /*.use_mmap =*/ true, /*.use_mlock =*/ false, @@ -7987,10 +9069,12 @@ struct llama_context_params llama_context_default_params() { /*.yarn_beta_fast =*/ 32.0f, /*.yarn_beta_slow =*/ 1.0f, /*.yarn_orig_ctx =*/ 0, + /*.type_k =*/ GGML_TYPE_F16, + /*.type_v =*/ GGML_TYPE_F16, /*.mul_mat_q =*/ true, - /*.f16_kv =*/ true, /*.logits_all =*/ false, /*.embedding =*/ false, + /*.offload_kqv =*/ true, }; return result; @@ -8070,11 +9154,18 @@ struct llama_model * llama_load_model_from_file( LLAMA_LOG_INFO("\n"); } } + return true; }; } - if (!llama_model_load(path_model, *model, params)) { - LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); + int status = llama_model_load(path_model, *model, params); + GGML_ASSERT(status <= 0); + if (status < 0) { + if (status == -1) { + LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); + } else if (status == -2) { + LLAMA_LOG_INFO("%s: cancelled model load\n", __func__); + } delete model; return nullptr; } @@ -8107,6 +9198,7 @@ struct llama_context * llama_new_context_with_model( cparams.yarn_beta_fast = params.yarn_beta_fast; cparams.yarn_beta_slow = params.yarn_beta_slow; cparams.mul_mat_q = params.mul_mat_q; + cparams.offload_kqv = params.offload_kqv; cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; @@ -8140,19 +9232,68 @@ struct llama_context * llama_new_context_with_model( ctx->rng = std::mt19937(params.seed); ctx->logits_all = params.logits_all; - ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32; + const ggml_type type_k = params.type_k; + const ggml_type type_v = params.type_v; + + GGML_ASSERT(hparams.n_embd_head() % ggml_blck_size(type_k) == 0); + GGML_ASSERT(hparams.n_embd_head() % ggml_blck_size(type_v) == 0); // reserve memory for context buffers if (!hparams.vocab_only) { - if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, cparams.n_ctx, model->n_gpu_layers)) { + // initialize backend +#ifdef GGML_USE_METAL + if (model->n_gpu_layers > 0) { + ctx->backend = ggml_backend_metal_init(); + if (ctx->backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__); + } + } +#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST) + // for testing only + if (model->n_gpu_layers > 0) { + ctx->backend = ggml_backend_cuda_init(0); + if (ctx->backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize CUDA backend\n", __func__); + } + } +#endif + + if (ctx->backend == nullptr && ggml_backend_buffer_is_host(model->buf)) { + ctx->backend = ggml_backend_cpu_init(); + if (ctx->backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__); + } + } + + if (ctx->backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize a backend\n", __func__); + delete ctx; + return nullptr; + } + + if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, type_k, type_v, + cparams.n_ctx, model->n_gpu_layers, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); return nullptr; } { - const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v); - LLAMA_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); + size_t memory_size_k = 0; + size_t memory_size_v = 0; + + for (auto & k : ctx->kv_self.k_l) { + memory_size_k += ggml_nbytes(k); + } + + for (auto & v : ctx->kv_self.v_l) { + memory_size_v += ggml_nbytes(v); + } + + LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, + (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), + ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), + ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); } // resized during inference @@ -8167,12 +9308,11 @@ struct llama_context * llama_new_context_with_model( } { - static const size_t tensor_alignment = 32; // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data - ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead()); + ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead()); // create measure allocator - ctx->alloc = ggml_allocr_new_measure(tensor_alignment); + ctx->alloc = ggml_allocr_new_measure_from_backend(ctx->backend); // build worst-case graph int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch); @@ -8180,96 +9320,50 @@ struct llama_context * llama_new_context_with_model( llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0)); -#ifdef GGML_USE_METAL - if (model->n_gpu_layers > 0) { - ggml_metal_log_set_callback(llama_log_callback_default, NULL); - - ctx->ctx_metal = ggml_metal_init(1); - if (!ctx->ctx_metal) { - LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__); - llama_free(ctx); - return NULL; - } - //ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false); - //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal)); - } -#endif // measure memory requirements for the graph - size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment; + size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf); - LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MiB\n", __func__, (ctx->buf_compute_meta.size() + alloc_size) / 1024.0 / 1024.0); - // recreate allocator with exact memory requirements + // create allocator again with exact memory requirements ggml_allocr_free(ctx->alloc); - ctx->buf_alloc.resize(alloc_size); - ctx->alloc = ggml_allocr_new(ctx->buf_alloc.data, ctx->buf_alloc.size, tensor_alignment); -#ifdef GGML_USE_METAL - if (ctx->ctx_metal) { - //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal)); - } -#endif -#ifdef GGML_USE_CUBLAS - ggml_cuda_set_scratch_size(alloc_size); - LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MB\n", __func__, alloc_size / 1024.0 / 1024.0); + ctx->buf_alloc = ggml_backend_alloc_buffer(ctx->backend, alloc_size); + ctx->alloc = ggml_allocr_new_from_buffer(ctx->buf_alloc); +#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST) + if (model->n_gpu_layers > 0) { + ggml_cuda_set_scratch_size(alloc_size); + LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0); - // calculate total VRAM usage - auto add_tensor = [](const ggml_tensor * t, size_t & size) { - if (t->backend == GGML_BACKEND_GPU || t->backend == GGML_BACKEND_GPU_SPLIT) { - size += ggml_nbytes(t); + // calculate total VRAM usage + auto add_tensor = [](const ggml_tensor * t, size_t & size) { + if (t->backend == GGML_BACKEND_GPU || t->backend == GGML_BACKEND_GPU_SPLIT) { + size += ggml_nbytes(t); + } + }; + size_t model_vram_size = 0; + for (const auto & kv : model->tensors_by_name) { + add_tensor(kv.second, model_vram_size); } - }; - size_t model_vram_size = 0; - for (const auto & kv : model->tensors_by_name) { - add_tensor(kv.second, model_vram_size); - } - - size_t kv_vram_size = 0; - add_tensor(ctx->kv_self.k, kv_vram_size); - add_tensor(ctx->kv_self.v, kv_vram_size); - - size_t ctx_vram_size = alloc_size + kv_vram_size; - size_t total_vram_size = model_vram_size + ctx_vram_size; - - LLAMA_LOG_INFO("%s: total VRAM used: %.2f MB (model: %.2f MB, context: %.2f MB)\n", __func__, - total_vram_size / 1024.0 / 1024.0, - model_vram_size / 1024.0 / 1024.0, - ctx_vram_size / 1024.0 / 1024.0); -#endif - } - -#ifdef GGML_USE_METAL - if (model->n_gpu_layers > 0) { - // this allocates all Metal resources and memory buffers - - void * data_ptr = NULL; - size_t data_size = 0; - - if (ctx->model.mapping) { - data_ptr = ctx->model.mapping->addr; - data_size = ctx->model.mapping->size; - } else { - data_ptr = ggml_get_mem_buffer(ctx->model.ctx); - data_size = ggml_get_mem_size (ctx->model.ctx); - } - const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx); + size_t kv_vram_size = 0; + for (auto & k : ctx->kv_self.k_l) { + add_tensor(k, kv_vram_size); + } + for (auto & v : ctx->kv_self.v_l) { + add_tensor(v, kv_vram_size); + } - LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); + size_t ctx_vram_size = alloc_size + kv_vram_size; + size_t total_vram_size = model_vram_size + ctx_vram_size; -#define LLAMA_METAL_CHECK_BUF(result) \ - if (!(result)) { \ - LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \ - llama_free(ctx); \ - return NULL; \ + LLAMA_LOG_INFO("%s: total VRAM used: %.2f MiB (model: %.2f MiB, context: %.2f MiB)\n", __func__, + total_vram_size / 1024.0 / 1024.0, + model_vram_size / 1024.0 / 1024.0, + ctx_vram_size / 1024.0 / 1024.0); } - - LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size)); - LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0)); - LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0)); -#undef LLAMA_METAL_CHECK_BUF - } #endif + } } #ifdef GGML_USE_MPI @@ -8297,10 +9391,14 @@ const llama_model * llama_get_model(const struct llama_context * ctx) { return &ctx->model; } -int llama_n_ctx(const struct llama_context * ctx) { +uint32_t llama_n_ctx(const struct llama_context * ctx) { return ctx->cparams.n_ctx; } +uint32_t llama_n_batch(const struct llama_context * ctx) { + return ctx->cparams.n_batch; +} + enum llama_vocab_type llama_vocab_type(const struct llama_model * model) { return model->vocab.type; } @@ -8321,6 +9419,45 @@ float llama_rope_freq_scale_train(const struct llama_model * model) { return model->hparams.rope_freq_scale_train; } +int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) { + const auto & it = model->gguf_kv.find(key); + if (it == model->gguf_kv.end()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + +int llama_model_meta_count(const struct llama_model * model) { + return (int)model->gguf_kv.size(); +} + +int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->first.c_str()); +} + +int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s %s %s", llama_model_arch_name(model->arch).c_str(), @@ -8379,8 +9516,107 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha } } +struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq) { + struct llama_kv_cache_view result = { + /*.n_cells = */ 0, + /*.n_max_seq = */ n_max_seq, + /*.token_count = */ 0, + /*.used_cells = */ llama_get_kv_cache_used_cells(ctx), + /*.max_contiguous = */ 0, + /*.max_contiguous_idx = */ -1, + /*.cells = */ nullptr, + /*.cells_sequences = */ nullptr, + }; + return result; +} + +void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { + if (view->cells != nullptr) { + free(view->cells); + view->cells = nullptr; + } + if (view->cells_sequences != nullptr) { + free(view->cells_sequences); + view->cells_sequences = nullptr; + } +} + +void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) { + if (uint32_t(view->n_cells) < ctx->kv_self.size || view->cells == nullptr) { + view->n_cells = int32_t(ctx->kv_self.size); + void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells); + GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells"); + view->cells = (struct llama_kv_cache_view_cell *)p; + p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_max_seq * view->n_cells); + GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences"); + view->cells_sequences = (llama_seq_id *)p; + } + + const std::vector & kv_cells = ctx->kv_self.cells; + llama_kv_cache_view_cell * c_curr = view->cells; + llama_seq_id * cs_curr = view->cells_sequences; + int32_t used_cells = 0; + int32_t token_count = 0; + int32_t curr_contig_idx = -1; + uint32_t max_contig = 0; + int32_t max_contig_idx = -1; + + for (int32_t i = 0; i < int32_t(ctx->kv_self.size); i++, c_curr++, cs_curr += view->n_max_seq) { + const size_t curr_size = kv_cells[i].seq_id.size(); + token_count += curr_size; + c_curr->pos = kv_cells[i].pos + kv_cells[i].delta; + + if (curr_size > 0) { + if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) { + max_contig = i - curr_contig_idx; + max_contig_idx = curr_contig_idx; + } + curr_contig_idx = -1; + } else if (curr_contig_idx < 0) { + curr_contig_idx = i; + } + + int seq_idx = 0; + for (const llama_seq_id it : kv_cells[i].seq_id) { + if (seq_idx >= view->n_max_seq) { + break; + } + cs_curr[seq_idx] = it; + seq_idx++; + } + if (seq_idx != 0) { + used_cells++; + } + for (; seq_idx < view->n_max_seq; seq_idx++) { + cs_curr[seq_idx] = -1; + } + } + if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) { + max_contig_idx = curr_contig_idx; + max_contig = kv_cells.size() - curr_contig_idx; + } + view->max_contiguous = max_contig; + view->max_contiguous_idx = max_contig_idx; + view->token_count = token_count; + view->used_cells = used_cells; + if (uint32_t(used_cells) != ctx->kv_self.used) { + LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n", + __func__, ctx->kv_self.used, used_cells); + } +} + int llama_get_kv_cache_token_count(const struct llama_context * ctx) { - return ctx->kv_self.head; + int result = 0; + + for (uint32_t i = 0; i < ctx->kv_self.size; i++) { + result += ctx->kv_self.cells[i].seq_id.size(); + } + + return result; +} + +int llama_get_kv_cache_used_cells(const struct llama_context * ctx) { + return ctx->kv_self.used; } void llama_kv_cache_clear(struct llama_context * ctx) { @@ -8419,7 +9655,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) { const size_t s_embedding = ctx->embedding.size() * sizeof(float); const size_t s_kv_size = sizeof(size_t); const size_t s_kv_ntok = sizeof(int); - const size_t s_kv = ctx->kv_self.buf.size; + const size_t s_kv = ggml_backend_buffer_get_size(ctx->kv_self.buf); const size_t s_total = ( + s_rng_size @@ -8547,46 +9783,59 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat const auto n_embd = hparams.n_embd_gqa(); const auto n_ctx = cparams.n_ctx; - const size_t kv_buf_size = kv_self.buf.size; + const size_t kv_buf_size = ggml_backend_buffer_get_size(kv_self.buf); const uint32_t kv_head = kv_self.head; const uint32_t kv_size = kv_self.size; + const uint32_t kv_used = kv_self.used; data_ctx->write(&kv_buf_size, sizeof(kv_buf_size)); data_ctx->write(&kv_head, sizeof(kv_head)); data_ctx->write(&kv_size, sizeof(kv_size)); + data_ctx->write(&kv_used, sizeof(kv_used)); if (kv_buf_size) { - const size_t elt_size = ggml_element_size(kv_self.k); + const size_t elt_size = ggml_element_size(kv_self.k_l[0]); - ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_context * cpy_ctx = ggml_init({ 6*n_layer*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); ggml_cgraph * gf = ggml_new_graph(cpy_ctx); - ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); - std::vector kout3d_data(ggml_nbytes(kout3d), 0); - kout3d->data = kout3d_data.data(); + std::vector kout2d(n_layer); + std::vector vout2d(n_layer); + + for (int il = 0; il < (int) n_layer; ++il) { + kout2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.k_l[il]->type, n_embd, kv_head); + vout2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.v_l[il]->type, kv_head, n_embd); + + ggml_tensor * k2d = ggml_view_2d(cpy_ctx, kv_self.k_l[il], + n_embd, kv_head, + elt_size*n_embd, 0); + + ggml_tensor * v2d = ggml_view_2d(cpy_ctx, kv_self.v_l[il], + kv_head, n_embd, + elt_size*n_ctx, 0); - ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer); - std::vector vout3d_data(ggml_nbytes(vout3d), 0); - vout3d->data = vout3d_data.data(); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k2d, kout2d[il])); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v2d, vout2d[il])); + } + + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(cpy_ctx, ctx->backend); - ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_head, n_layer, - elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + ggml_backend_graph_compute(ctx->backend, gf); - ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_head, n_embd, n_layer, - elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + std::vector tmp_buf; + for (int il = 0; il < (int) n_layer; ++il) { + tmp_buf.resize(ggml_nbytes(kout2d[il])); + ggml_backend_tensor_get(kout2d[il], tmp_buf.data(), 0, tmp_buf.size()); + data_ctx->write(tmp_buf.data(), tmp_buf.size()); - ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d)); - ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d)); - ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); + tmp_buf.resize(ggml_nbytes(vout2d[il])); + ggml_backend_tensor_get(vout2d[il], tmp_buf.data(), 0, tmp_buf.size()); + data_ctx->write(tmp_buf.data(), tmp_buf.size()); + } ggml_free(cpy_ctx); - // our data is now in the kout3d_data and vout3d_data buffers - // write them to file - data_ctx->write(kout3d_data.data(), kout3d_data.size()); - data_ctx->write(vout3d_data.data(), vout3d_data.size()); + ggml_backend_buffer_free(buf); } for (uint32_t i = 0; i < kv_size; ++i) { @@ -8676,44 +9925,61 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { size_t kv_buf_size; uint32_t kv_head; uint32_t kv_size; + uint32_t kv_used; memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size); memcpy(&kv_head, inp, sizeof(kv_head)); inp += sizeof(kv_head); memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size); + memcpy(&kv_used, inp, sizeof(kv_used)); inp += sizeof(kv_used); if (kv_buf_size) { - GGML_ASSERT(kv_self.buf.size == kv_buf_size); + GGML_ASSERT(ggml_backend_buffer_get_size(kv_self.buf) == kv_buf_size); - const size_t elt_size = ggml_element_size(kv_self.k); + const size_t elt_size = ggml_element_size(kv_self.k_l[0]); - ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_context * cpy_ctx = ggml_init({ 6*n_layer*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); ggml_cgraph * gf = ggml_new_graph(cpy_ctx); - ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); - kin3d->data = (void *) inp; - inp += ggml_nbytes(kin3d); + std::vector kin2d(n_layer); + std::vector vin2d(n_layer); + + for (int il = 0; il < n_layer; ++il) { + kin2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.k_l[il]->type, n_embd, kv_head); + vin2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.v_l[il]->type, kv_head, n_embd); + + ggml_tensor * k2d = ggml_view_2d(cpy_ctx, kv_self.k_l[il], + n_embd, kv_head, + elt_size*n_embd, 0); + + ggml_tensor * v2d = ggml_view_2d(cpy_ctx, kv_self.v_l[il], + kv_head, n_embd, + elt_size*n_ctx, 0); + + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin2d[il], k2d)); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin2d[il], v2d)); + } - ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer); - vin3d->data = (void *) inp; - inp += ggml_nbytes(vin3d); + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(cpy_ctx, ctx->backend); - ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_head, n_layer, - elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + // load data into the tensors + for (int il = 0; il < n_layer; ++il) { + ggml_backend_tensor_set(kin2d[il], inp, 0, ggml_nbytes(kin2d[il])); + inp += ggml_nbytes(kin2d[il]); - ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_head, n_embd, n_layer, - elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + ggml_backend_tensor_set(vin2d[il], inp, 0, ggml_nbytes(vin2d[il])); + inp += ggml_nbytes(vin2d[il]); + } - ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d)); - ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d)); - ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); + ggml_backend_graph_compute(ctx->backend, gf); ggml_free(cpy_ctx); + + ggml_backend_buffer_free(buf); } ctx->kv_self.head = kv_head; ctx->kv_self.size = kv_size; + ctx->kv_self.used = kv_used; ctx->kv_self.cells.resize(kv_size); @@ -8931,6 +10197,7 @@ float * llama_get_logits(struct llama_context * ctx) { } float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { + assert(ctx->logits_valid.at(i)); return ctx->logits.data() + i*ctx->model.hparams.n_vocab; } @@ -8962,6 +10229,14 @@ llama_token llama_token_nl(const struct llama_model * model) { return model->vocab.linefeed_id; } +int llama_add_bos_token(const struct llama_model * model) { + return model->vocab.special_add_bos; +} + +int llama_add_eos_token(const struct llama_model * model) { + return model->vocab.special_add_eos; +} + llama_token llama_token_prefix(const struct llama_model * model) { return model->vocab.special_prefix_id; } @@ -9168,6 +10443,9 @@ const std::vector> & llama_internal void llama_log_set(ggml_log_callback log_callback, void * user_data) { g_state.log_callback = log_callback ? log_callback : llama_log_callback_default; g_state.log_callback_user_data = user_data; +#ifdef GGML_USE_METAL + ggml_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data); +#endif } static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {