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" @@ -75,6 +75,7 @@ #include #include #include +#include #include #include #include @@ -105,8 +106,7 @@ #endif #define LLAMA_MAX_NODES 8192 -#define LLAMA_MAX_EXPERTS 8 - +#define LLAMA_MAX_EXPERTS 60 // // logging @@ -205,11 +205,14 @@ enum llm_arch { LLM_ARCH_REFACT, LLM_ARCH_BERT, LLM_ARCH_NOMIC_BERT, + LLM_ARCH_JINA_BERT_V2, LLM_ARCH_BLOOM, LLM_ARCH_STABLELM, LLM_ARCH_QWEN, LLM_ARCH_QWEN2, + LLM_ARCH_QWEN2MOE, LLM_ARCH_PHI2, + LLM_ARCH_PHI3, LLM_ARCH_PLAMO, LLM_ARCH_CODESHELL, LLM_ARCH_ORION, @@ -220,39 +223,46 @@ enum llm_arch { LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, LLM_ARCH_COMMAND_R, + LLM_ARCH_DBRX, + LLM_ARCH_OLMO, LLM_ARCH_UNKNOWN, }; static const std::map LLM_ARCH_NAMES = { - { LLM_ARCH_LLAMA, "llama" }, - { LLM_ARCH_FALCON, "falcon" }, - { LLM_ARCH_GROK, "grok" }, - { LLM_ARCH_GPT2, "gpt2" }, - { LLM_ARCH_GPTJ, "gptj" }, - { LLM_ARCH_GPTNEOX, "gptneox" }, - { LLM_ARCH_MPT, "mpt" }, - { LLM_ARCH_BAICHUAN, "baichuan" }, - { LLM_ARCH_STARCODER, "starcoder" }, - { LLM_ARCH_PERSIMMON, "persimmon" }, - { LLM_ARCH_REFACT, "refact" }, - { LLM_ARCH_BERT, "bert" }, - { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, - { LLM_ARCH_BLOOM, "bloom" }, - { LLM_ARCH_STABLELM, "stablelm" }, - { LLM_ARCH_QWEN, "qwen" }, - { LLM_ARCH_QWEN2, "qwen2" }, - { LLM_ARCH_PHI2, "phi2" }, - { LLM_ARCH_PLAMO, "plamo" }, - { LLM_ARCH_CODESHELL, "codeshell" }, - { LLM_ARCH_ORION, "orion" }, - { LLM_ARCH_INTERNLM2, "internlm2" }, - { LLM_ARCH_MINICPM, "minicpm" }, - { LLM_ARCH_GEMMA, "gemma" }, - { LLM_ARCH_STARCODER2, "starcoder2" }, - { LLM_ARCH_MAMBA, "mamba" }, - { LLM_ARCH_XVERSE, "xverse" }, - { LLM_ARCH_COMMAND_R, "command-r" }, - { LLM_ARCH_UNKNOWN, "(unknown)" }, + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GROK, "grok" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_BAICHUAN, "baichuan" }, + { LLM_ARCH_STARCODER, "starcoder" }, + { LLM_ARCH_PERSIMMON, "persimmon" }, + { LLM_ARCH_REFACT, "refact" }, + { LLM_ARCH_BERT, "bert" }, + { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, + { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" }, + { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, + { LLM_ARCH_QWEN, "qwen" }, + { LLM_ARCH_QWEN2, "qwen2" }, + { LLM_ARCH_QWEN2MOE, "qwen2moe" }, + { LLM_ARCH_PHI2, "phi2" }, + { LLM_ARCH_PHI3, "phi3" }, + { LLM_ARCH_PLAMO, "plamo" }, + { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, + { LLM_ARCH_INTERNLM2, "internlm2" }, + { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_GEMMA, "gemma" }, + { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_XVERSE, "xverse" }, + { LLM_ARCH_COMMAND_R, "command-r" }, + { LLM_ARCH_DBRX, "dbrx" }, + { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_UNKNOWN, "(unknown)" }, }; enum llm_kv { @@ -308,6 +318,7 @@ enum llm_kv { LLM_KV_SSM_TIME_STEP_RANK, LLM_KV_TOKENIZER_MODEL, + LLM_KV_TOKENIZER_PRE, LLM_KV_TOKENIZER_LIST, LLM_KV_TOKENIZER_TOKEN_TYPE, LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, @@ -318,11 +329,17 @@ enum llm_kv { LLM_KV_TOKENIZER_UNK_ID, LLM_KV_TOKENIZER_SEP_ID, LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_CLS_ID, + LLM_KV_TOKENIZER_MASK_ID, LLM_KV_TOKENIZER_ADD_BOS, LLM_KV_TOKENIZER_ADD_EOS, LLM_KV_TOKENIZER_ADD_PREFIX, LLM_KV_TOKENIZER_HF_JSON, LLM_KV_TOKENIZER_RWKV, + LLM_KV_TOKENIZER_PREFIX_ID, + LLM_KV_TOKENIZER_SUFFIX_ID, + LLM_KV_TOKENIZER_MIDDLE_ID, + LLM_KV_TOKENIZER_EOT_ID, }; static const std::map LLM_KV_NAMES = { @@ -378,6 +395,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, + { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" }, { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, "tokenizer.ggml.token_type_count" }, @@ -388,11 +406,17 @@ static const 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_CLS_ID, "tokenizer.ggml.cls_token_id" }, + { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_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_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" }, { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, + { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" }, + { LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" }, + { LLM_KV_TOKENIZER_MIDDLE_ID, "tokenizer.ggml.middle_token_id" }, + { LLM_KV_TOKENIZER_EOT_ID, "tokenizer.ggml.eot_token_id" }, }; struct LLM_KV { @@ -423,6 +447,7 @@ enum llm_tensor { LLM_TENSOR_ATTN_OUT_NORM, LLM_TENSOR_ATTN_ROT_EMBD, LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_INP_SHEXP, LLM_TENSOR_FFN_NORM, LLM_TENSOR_FFN_GATE, LLM_TENSOR_FFN_DOWN, @@ -434,6 +459,9 @@ enum llm_tensor { LLM_TENSOR_FFN_DOWN_EXPS, // merged experts LLM_TENSOR_FFN_GATE_EXPS, LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_K_NORM, LLM_TENSOR_LAYER_OUT_NORM, @@ -665,6 +693,25 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_JINA_BERT_V2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_BLOOM, { @@ -696,6 +743,8 @@ static const std::map> LLM_TENSOR_NA { 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_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, }, }, { @@ -731,6 +780,28 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_QWEN2MOE, + { + { 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_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_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_PHI2, { @@ -747,6 +818,23 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_PHI3, + { + { 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_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_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_PLAMO, { @@ -926,6 +1014,38 @@ static const std::map> LLM_TENSOR_NA { 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_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + }, + }, + { + LLM_ARCH_DBRX, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, + { + LLM_ARCH_OLMO, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, { @@ -1522,12 +1642,12 @@ struct llama_mlock { }; using llama_mlocks = std::vector>; -static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) { +static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) { std::vector result(8, 0); - const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size()); + const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); if (n_tokens < 0) { result.resize(-n_tokens); - int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size()); + int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); GGML_ASSERT(check == -n_tokens); } else { @@ -1632,17 +1752,17 @@ static size_t llama_get_device_memory(int device) { #if defined(GGML_USE_CUDA) size_t total; size_t free; - ggml_backend_cuda_get_device_memory(device, &total, &free); + ggml_backend_cuda_get_device_memory(device, &free, &total); return free; #elif defined(GGML_USE_SYCL) size_t total; size_t free; - ggml_backend_sycl_get_device_memory(device, &total, &free); + ggml_backend_sycl_get_device_memory(device, &free, &total); return free; #elif defined(GGML_USE_VULKAN) size_t total; size_t free; - ggml_backend_vk_get_device_memory(device, &total, &free); + ggml_backend_vk_get_device_memory(device, &free, &total); return free; #else return 1; @@ -1684,6 +1804,7 @@ enum e_model { MODEL_4B, MODEL_7B, MODEL_8B, + MODEL_12B, MODEL_13B, MODEL_14B, MODEL_15B, @@ -1699,6 +1820,10 @@ enum e_model { MODEL_MEDIUM, MODEL_LARGE, MODEL_XL, + MODEL_A2_7B, + MODEL_8x7B, + MODEL_8x22B, + MODEL_16x12B, }; static const size_t kiB = 1024; @@ -1741,7 +1866,7 @@ struct llama_hparams { float f_logit_scale = 0.0f; bool causal_attn = true; - bool need_kq_pos = false; + bool use_alibi = false; enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; @@ -1831,6 +1956,7 @@ struct llama_cparams { bool embeddings; bool causal_attn; bool offload_kqv; + bool flash_attn; enum llama_pooling_type pooling_type; @@ -1882,6 +2008,12 @@ struct llama_layer { struct ggml_tensor * ffn_down_exps; struct ggml_tensor * ffn_up_exps ; + // ff shared expert (shexp) + struct ggml_tensor * ffn_gate_inp_shexp; + struct ggml_tensor * ffn_gate_shexp; + struct ggml_tensor * ffn_down_shexp; + struct ggml_tensor * ffn_up_shexp; + // ff bias struct ggml_tensor * ffn_down_b; // b2 struct ggml_tensor * ffn_up_b; // b3 @@ -1928,8 +2060,8 @@ struct llama_kv_cache { bool has_shift = false; bool do_defrag = false; bool do_copy = false; - // with recurrent state models, a cell can hold the state for more than one past token - bool recurrent = false; + bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token + bool v_trans = true; // the value tensor is transposed // Note: The value of head isn't only used to optimize searching // for a free KV slot. llama_decode_internal also uses it, so it @@ -2006,7 +2138,8 @@ struct llama_vocab { ttype type; }; - enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; + enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; + enum llama_vocab_pre_type type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; std::unordered_map token_to_id; std::vector id_to_token; @@ -2016,20 +2149,22 @@ struct llama_vocab { std::map, int> bpe_ranks; // default LLaMA special tokens - id special_bos_id = 1; - id special_eos_id = 2; - id special_unk_id = 0; - id special_sep_id = -1; - id special_pad_id = -1; + id special_bos_id = 1; + id special_eos_id = 2; + id special_unk_id = 0; + id special_sep_id = -1; + id special_pad_id = -1; + id special_cls_id = -1; + id special_mask_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; + id special_prefix_id = -1; + id special_suffix_id = -1; + id special_middle_id = -1; + id special_eot_id = -1; // TODO: move above after "eos_id", and here add "file separator" token bool add_space_prefix = true; @@ -2177,7 +2312,7 @@ struct llama_context { std::vector output_ids; // map batch token positions to ids of the logits and embd buffers size_t output_size = 0; // capacity (of tokens positions) for the output buffers - int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch + int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch bool logits_all = false; @@ -2203,7 +2338,6 @@ struct llama_context { struct ggml_tensor * inp_pos; // I32 [n_batch] struct ggml_tensor * inp_out_ids; // I32 [n_outputs] struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_KQ_pos; // F32 [n_kv] struct ggml_tensor * inp_K_shift; // I32 [kv_size] struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] struct ggml_tensor * inp_cls; // I32 [n_batch] @@ -2225,11 +2359,14 @@ struct llama_context { static bool llama_kv_cache_init( struct llama_kv_cache & cache, - const llama_model & model, + const llama_context * ctx, ggml_type type_k, ggml_type type_v, uint32_t kv_size, bool offload) { + const llama_model & model = ctx->model; + const llama_cparams & cparams = ctx->cparams; + const struct llama_hparams & hparams = model.hparams; const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); @@ -2240,8 +2377,9 @@ static bool llama_kv_cache_init( // TODO: find a nicer way to add other recurrent model architectures cache.recurrent = model.arch == LLM_ARCH_MAMBA; + cache.v_trans = !cparams.flash_attn; - // TODO: support mixed reccurent Transformer architectues + // TODO: support mixed recurrent Transformer architectures // NOTE: (!a || b) is a logical implication (a -> b) GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s()); GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s()); @@ -2452,6 +2590,10 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { } cache.head = 0; cache.used = 0; + + for (auto & buf : cache.bufs) { + ggml_backend_buffer_clear(buf, 0); + } } static bool llama_kv_cache_seq_rm( @@ -2772,6 +2914,7 @@ namespace GGUFMeta { case LLAMA_KV_OVERRIDE_TYPE_BOOL: return "bool"; case LLAMA_KV_OVERRIDE_TYPE_INT: return "int"; case LLAMA_KV_OVERRIDE_TYPE_FLOAT: return "float"; + case LLAMA_KV_OVERRIDE_TYPE_STR: return "str"; } return "unknown"; } @@ -2783,13 +2926,16 @@ namespace GGUFMeta { __func__, override_type_to_str(ovrd->tag), ovrd->key); switch (ovrd->tag) { case LLAMA_KV_OVERRIDE_TYPE_BOOL: { - LLAMA_LOG_INFO("%s\n", ovrd->bool_value ? "true" : "false"); + LLAMA_LOG_INFO("%s\n", ovrd->val_bool ? "true" : "false"); } break; case LLAMA_KV_OVERRIDE_TYPE_INT: { - LLAMA_LOG_INFO("%" PRId64 "\n", ovrd->int_value); + LLAMA_LOG_INFO("%" PRId64 "\n", ovrd->val_i64); } break; case LLAMA_KV_OVERRIDE_TYPE_FLOAT: { - LLAMA_LOG_INFO("%.6f\n", ovrd->float_value); + LLAMA_LOG_INFO("%.6f\n", ovrd->val_f64); + } break; + case LLAMA_KV_OVERRIDE_TYPE_STR: { + LLAMA_LOG_INFO("%s\n", ovrd->val_str); } break; default: // Shouldn't be possible to end up here, but just in case... @@ -2808,7 +2954,7 @@ namespace GGUFMeta { static typename std::enable_if::value, bool>::type try_override(OT & target, const struct llama_model_kv_override * ovrd) { if (validate_override(LLAMA_KV_OVERRIDE_TYPE_BOOL, ovrd)) { - target = ovrd->bool_value; + target = ovrd->val_bool; return true; } return false; @@ -2818,7 +2964,7 @@ namespace GGUFMeta { static typename std::enable_if::value && std::is_integral::value, bool>::type try_override(OT & target, const struct llama_model_kv_override * ovrd) { if (validate_override(LLAMA_KV_OVERRIDE_TYPE_INT, ovrd)) { - target = ovrd->int_value; + target = ovrd->val_i64; return true; } return false; @@ -2828,7 +2974,7 @@ namespace GGUFMeta { static typename std::enable_if::value, bool>::type try_override(T & target, const struct llama_model_kv_override * ovrd) { if (validate_override(LLAMA_KV_OVERRIDE_TYPE_FLOAT, ovrd)) { - target = ovrd->float_value; + target = ovrd->val_f64; return true; } return false; @@ -2837,12 +2983,11 @@ namespace GGUFMeta { template static typename std::enable_if::value, bool>::type try_override(T & target, const struct llama_model_kv_override * ovrd) { - (void)target; - (void)ovrd; - if (!ovrd) { 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", - ovrd ? ovrd->key : "NULL")); + if (validate_override(LLAMA_KV_OVERRIDE_TYPE_STR, ovrd)) { + target = ovrd->val_str; + return true; + } + return false; } static bool set(const gguf_context * ctx, const int k, T & target, const struct llama_model_kv_override * ovrd = nullptr) { @@ -2875,6 +3020,7 @@ struct llama_model_loader { size_t n_bytes = 0; bool use_mmap = false; + bool check_tensors; llama_files files; llama_ftype ftype; @@ -2889,9 +3035,13 @@ struct llama_model_loader { ggml_tensor * tensor; - llama_tensor_weight(uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { + llama_tensor_weight(const llama_file * file, uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { const int tensor_idx = gguf_find_tensor(gguf_ctx, name); offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx); + + if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size) { + throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", name)); + } } }; std::vector weights; @@ -2904,7 +3054,7 @@ struct llama_model_loader { 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) { + llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p) { int trace = 0; if (getenv("LLAMA_TRACE")) { trace = atoi(getenv("LLAMA_TRACE")); @@ -2930,15 +3080,15 @@ struct llama_model_loader { get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false); llm_kv = LLM_KV(llm_arch_from_string(arch_name)); + files.emplace_back(new llama_file(fname.c_str(), "rb")); + contexts.emplace_back(ctx); + // Save tensors data offset of the main file. // For subsidiary files, `meta` tensor data offset must not be used, // so we build a unified tensors index for weights. for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { - weights.emplace_back(0, cur->name, meta, cur); + weights.emplace_back(files.back().get(), 0, cur->name, meta, cur); } - files.emplace_back(new llama_file(fname.c_str(), "rb")); - contexts.emplace_back(ctx); - uint16_t n_split = 0; get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false); @@ -2972,12 +3122,13 @@ struct llama_model_loader { throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path)); } + files.emplace_back(new llama_file(split_path, "rb")); + contexts.emplace_back(ctx); + // Save tensors data offset info of the shard. for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { - weights.emplace_back(idx, cur->name, ctx_gguf, cur); + weights.emplace_back(files.back().get(), idx, cur->name, ctx_gguf, cur); } - files.emplace_back(new llama_file(split_path, "rb")); - contexts.emplace_back(ctx); gguf_free(ctx_gguf); } @@ -3000,9 +3151,17 @@ struct llama_model_loader { fver = (enum llama_fver) gguf_get_version(meta); + std::set tensor_names; for (auto & w : weights) { n_elements += ggml_nelements(w.tensor); n_bytes += ggml_nbytes(w.tensor); + // make sure there is no duplicated tensor names + const std::string name(w.tensor->name); + auto found = tensor_names.find(name); + if (found != tensor_names.end()) { + throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", w.tensor->name)); + } + tensor_names.insert(name); } LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", @@ -3036,6 +3195,7 @@ struct llama_model_loader { switch (type_max) { case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; + case GGML_TYPE_BF16: ftype = LLAMA_FTYPE_MOSTLY_BF16; break; case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; @@ -3108,6 +3268,7 @@ struct llama_model_loader { } this->use_mmap = use_mmap; + this->check_tensors = check_tensors; } ~llama_model_loader() { @@ -3187,6 +3348,10 @@ struct llama_model_loader { return nullptr; } + const llama_tensor_weight * get_weight(int i) const { + return get_weight(get_tensor_name(i)); + } + const llama_tensor_weight & require_weight(const char * name) const { const llama_tensor_weight * weight = get_weight(name); if (!weight) { @@ -3362,6 +3527,10 @@ struct llama_model_loader { file->seek(w.offs, SEEK_SET); file->read_raw(cur->data, ggml_nbytes(cur)); } + + if (check_tensors && !ggml_validate_row_data(cur->type, cur->data, ggml_nbytes(cur))) { + throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur))); + } } size_t size_done = 0; @@ -3378,6 +3547,8 @@ struct llama_model_loader { GGML_ASSERT(size_data != 0 && "call init_mappings() first"); std::vector> read_buf; + std::vector>> validation_result; + for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { const auto * weight = get_weight(ggml_get_name(cur)); if (weight == nullptr) { @@ -3399,37 +3570,66 @@ struct llama_model_loader { if (bufs_mmap.count(weight->idx)) { buf_mmap = bufs_mmap.at(weight->idx); } + uint8_t * data = (uint8_t *) mapping->addr + weight->offs; + + if (check_tensors) { + validation_result.emplace_back(std::async(std::launch::async, [cur, data, n_size] { + return std::make_pair(cur, ggml_validate_row_data(cur->type, data, n_size)); + })); + } + GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated if (buf_mmap && cur->data == nullptr) { - ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + weight->offs); + ggml_backend_tensor_alloc(buf_mmap, cur, data); if (lmlocks) { const auto & lmlock = lmlocks->at(weight->idx); - lmlock->grow_to(weight->offs + ggml_nbytes(cur)); + lmlock->grow_to(weight->offs + n_size); } auto & mmap_used = mmaps_used[weight->idx]; mmap_used.first = std::min(mmap_used.first, weight->offs); mmap_used.second = std::max(mmap_used.second, weight->offs + n_size); } else { - ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + weight->offs, 0, n_size); + ggml_backend_tensor_set(cur, data, 0, n_size); } } else { GGML_ASSERT(weight->idx < files.size()); const auto & file = files.at(weight->idx); if (ggml_backend_buffer_is_host(cur->buffer)) { file->seek(weight->offs, SEEK_SET); - file->read_raw(cur->data, ggml_nbytes(cur)); + file->read_raw(cur->data, n_size); + if (check_tensors) { + validation_result.emplace_back(std::async(std::launch::async, [cur, n_size] { + return std::make_pair(cur, ggml_validate_row_data(cur->type, cur->data, n_size)); + })); + } } else { - read_buf.resize(ggml_nbytes(cur)); + read_buf.resize(n_size); file->seek(weight->offs, SEEK_SET); - file->read_raw(read_buf.data(), ggml_nbytes(cur)); + file->read_raw(read_buf.data(), n_size); ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size); + if (check_tensors && !ggml_validate_row_data(cur->type, read_buf.data(), n_size)) { + throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur))); + } } } size_done += n_size; } + // check validation results + bool validation_failed = false; + for (auto & future : validation_result) { + auto result = future.get(); + if (!result.second) { + LLAMA_LOG_ERROR("%s: tensor '%s' has invalid data\n", __func__, ggml_get_name(result.first)); + validation_failed = true; + } + } + if (validation_failed) { + throw std::runtime_error("found tensors with invalid data"); + } + // check if this is the last call and do final cleanup if (size_done >= size_data) { // unmap offloaded tensors and metadata @@ -3487,6 +3687,7 @@ 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 "F16"; + case LLAMA_FTYPE_MOSTLY_BF16: return "BF16"; 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: @@ -3535,6 +3736,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_3B: return "3B"; case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; + case MODEL_12B: return "12B"; case MODEL_13B: return "13B"; case MODEL_14B: return "14B"; case MODEL_15B: return "15B"; @@ -3550,6 +3752,10 @@ static const char * llama_model_type_name(e_model type) { case MODEL_MEDIUM: return "0.4B"; case MODEL_LARGE: return "0.8B"; case MODEL_XL: return "1.5B"; + case MODEL_A2_7B: return "A2.7B"; + case MODEL_8x7B: return "8x7B"; + case MODEL_8x22B: return "8x22B"; + case MODEL_16x12B: return "16x12B"; default: return "?B"; } } @@ -3593,6 +3799,12 @@ static void llm_load_hparams( // get hparams kv ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + + // everything past this point is not vocab-related + if (hparams.vocab_only) { + return; + } + 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); @@ -3664,15 +3876,23 @@ static void llm_load_hparams( { 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; - case 48: model.type = e_model::MODEL_34B; break; - case 60: model.type = e_model::MODEL_30B; break; - case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; - default: model.type = e_model::MODEL_UNKNOWN; + if (hparams.n_expert == 8) { + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_8x7B; break; + case 56: model.type = e_model::MODEL_8x22B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } else { + 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 = hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_34B; break; + case 60: model.type = e_model::MODEL_30B; break; + case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } } } break; case LLM_ARCH_MINICPM: @@ -3768,6 +3988,19 @@ static void llm_load_hparams( model.type = e_model::MODEL_335M; break; // bge-large } } break; + case LLM_ARCH_JINA_BERT_V2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type); + hparams.f_max_alibi_bias = 8.0f; + + switch (hparams.n_layer) { + case 4: model.type = e_model::MODEL_33M; break; // jina-embeddings-small + case 12: model.type = e_model::MODEL_137M; break; // jina-embeddings-base + } + } break; case LLM_ARCH_NOMIC_BERT: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -3814,6 +4047,7 @@ static void llm_load_hparams( switch (hparams.n_layer) { case 24: model.type = e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_3B; break; + case 40: model.type = e_model::MODEL_12B; break; default: model.type = e_model::MODEL_UNKNOWN; } } break; @@ -3838,10 +4072,28 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_QWEN2MOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_A2_7B; 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 24: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_PHI3: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { case 24: model.type = e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_3B; break; @@ -3963,20 +4215,44 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_DBRX: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_16x12B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_OLMO: + { + 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); + + switch (hparams.n_layer) { + case 22: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 80: model.type = e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } model.ftype = ml.ftype; if (hparams.f_max_alibi_bias > 0.0f) { - hparams.need_kq_pos = true; + hparams.use_alibi = true; } hparams.rope_type = llama_rope_type(&model); } // TODO: This should probably be in llama.h -static std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special = false); +static std::vector llama_tokenize_internal( + const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special = false +); static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch); static void llm_load_vocab( @@ -3990,39 +4266,92 @@ static void llm_load_vocab( // determine vocab type { - std::string tokenizer_name; + std::string tokenizer_model; + std::string tokenizer_pre; - ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_name); + ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model); + ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false); - if (tokenizer_name == "no_vocab") { + if (tokenizer_model == "no_vocab") { vocab.type = LLAMA_VOCAB_TYPE_NONE; // default special tokens - vocab.special_bos_id = -1; - vocab.special_eos_id = -1; - vocab.special_unk_id = -1; - vocab.special_sep_id = -1; - vocab.special_pad_id = -1; - vocab.linefeed_id = -1; + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + vocab.linefeed_id = -1; return; - } else if (tokenizer_name == "llama") { + } else if (tokenizer_model == "llama") { vocab.type = LLAMA_VOCAB_TYPE_SPM; // default special tokens - vocab.special_bos_id = 1; - vocab.special_eos_id = 2; - vocab.special_unk_id = 0; - vocab.special_sep_id = -1; - vocab.special_pad_id = -1; + vocab.special_bos_id = 1; + vocab.special_eos_id = 2; + vocab.special_unk_id = 0; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + + // For Fill-In-the-Middle (FIM)/infill models which where converted + // prior to support of FIM special tokens in GGUF, the following + // will allow those models to continue to work. The general names + // of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and + // CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once + // new versions of these models have been published. + std::string gen_name; + ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false); + + std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(), + [](unsigned char c){ return std::tolower(c); }); + + if (gen_name.find("code") != std::string::npos) { + if (model.arch == LLM_ARCH_LLAMA) { + vocab.special_prefix_id = 32007; + vocab.special_suffix_id = 32008; + vocab.special_middle_id = 32009; + vocab.special_eot_id = 32010; + } else if (model.arch == LLM_ARCH_GEMMA) { + vocab.special_prefix_id = 67; + vocab.special_suffix_id = 69; + vocab.special_middle_id = 68; + // TODO: this is not EOT, it is "file separator" token, needs fix + // https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572 + //vocab.special_eot_id = 70; + vocab.special_eot_id = 107; + } + } const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); if (add_space_prefix_keyidx != -1) { vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); } // The default value of add_space_prefix is true. - } else if (tokenizer_name == "gpt2") { - vocab.type = LLAMA_VOCAB_TYPE_BPE; + } else if (tokenizer_model == "bert") { + vocab.type = LLAMA_VOCAB_TYPE_WPM; + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = 100; + vocab.special_sep_id = 102; + vocab.special_pad_id = 0; + vocab.special_cls_id = 101; + vocab.special_mask_id = 103; + vocab.add_space_prefix = false; + } else { + if (tokenizer_model == "gpt2") { + vocab.type = LLAMA_VOCAB_TYPE_BPE; + } else { + LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_model.c_str()); + LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); + vocab.type = LLAMA_VOCAB_TYPE_SPM; + return; + } // read bpe merges and populate bpe ranks const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str()); if (merges_keyidx == -1) { @@ -4049,26 +4378,74 @@ static void llm_load_vocab( } // default special tokens - vocab.special_bos_id = 11; - vocab.special_eos_id = 11; - vocab.special_unk_id = -1; - vocab.special_sep_id = -1; - vocab.special_pad_id = -1; - } else if (tokenizer_name == "bert") { - vocab.type = LLAMA_VOCAB_TYPE_WPM; - - // default special tokens - vocab.special_bos_id = 101; - vocab.special_eos_id = 102; - vocab.special_unk_id = 100; - vocab.special_sep_id = -1; - vocab.special_pad_id = -1; - vocab.add_space_prefix = false; + vocab.special_bos_id = 11; + vocab.special_eos_id = 11; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + } + + // for now, only BPE models have pre-tokenizers + if (vocab.type == LLAMA_VOCAB_TYPE_BPE) { + if (tokenizer_pre.empty()) { + LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__); + LLAMA_LOG_WARN("%s: \n", __func__); + LLAMA_LOG_WARN("%s: ************************************ \n", __func__); + LLAMA_LOG_WARN("%s: GENERATION QUALITY WILL BE DEGRADED! \n", __func__); + LLAMA_LOG_WARN("%s: CONSIDER REGENERATING THE MODEL \n", __func__); + LLAMA_LOG_WARN("%s: ************************************ \n", __func__); + LLAMA_LOG_WARN("%s: \n", __func__); + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + } else if ( + tokenizer_pre == "default") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + } else if ( + tokenizer_pre == "llama3" || + tokenizer_pre == "llama-v3" || + tokenizer_pre == "llama-bpe") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3; + } else if ( + tokenizer_pre == "deepseek-llm") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM; + } else if ( + tokenizer_pre == "deepseek-coder") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER; + } else if ( + tokenizer_pre == "falcon") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_FALCON; + } else if ( + tokenizer_pre == "mpt") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MPT; + } else if ( + tokenizer_pre == "starcoder") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STARCODER; + } else if ( + tokenizer_pre == "gpt-2" || + tokenizer_pre == "jina-es" || + tokenizer_pre == "jina-de") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2; + } else if ( + tokenizer_pre == "refact") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_REFACT; + } else if ( + tokenizer_pre == "command-r") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_COMMAND_R; + } else if ( + tokenizer_pre == "qwen2") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2; + } else if ( + tokenizer_pre == "olmo") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_OLMO; + } else if ( + tokenizer_pre == "dbrx") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX; + } else { + throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); + } } else { - LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str()); - LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); - - vocab.type = LLAMA_VOCAB_TYPE_SPM; + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; } } @@ -4125,12 +4502,19 @@ static void llm_load_vocab( // special tokens { const std::vector> special_token_types = { - { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id }, - { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id }, - { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id }, - { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id }, - { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id }, + { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id }, + { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id }, + { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id }, + { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id }, + { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id }, + { LLM_KV_TOKENIZER_CLS_ID, vocab.special_cls_id }, + { LLM_KV_TOKENIZER_MASK_ID, vocab.special_mask_id }, + { LLM_KV_TOKENIZER_PREFIX_ID, vocab.special_prefix_id }, + { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id }, + { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id }, + { LLM_KV_TOKENIZER_EOT_ID, vocab.special_eot_id }, }; + for (const auto & it : special_token_types) { const std::string & key = kv(std::get<0>(it)); int32_t & id = std::get<1>(it); @@ -4145,7 +4529,6 @@ static void llm_load_vocab( } else { id = new_id; } - } // Handle add_bos_token and add_eos_token @@ -4159,6 +4542,28 @@ static void llm_load_vocab( vocab.special_add_eos = int(temp); } } + + // find EOT token: "<|eot_id|>", "<|im_end|>", "", etc. + // + // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOT_ID + // for now, we apply this workaround to find the EOT token based on its text + if (vocab.special_eot_id == -1) { + for (const auto & t : vocab.token_to_id) { + if ( + // TODO: gemma "" is exported as a normal token, so the following check does not work + // need to fix convert script + //vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL && + (t.first == "<|eot_id|>" || + t.first == "<|im_end|>" || + t.first == "<|end|>" || + t.first == "" + ) + ) { + vocab.special_eot_id = t.second; + break; + } + } + } } // build special tokens cache @@ -4321,12 +4726,19 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { 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.special_cls_id != -1) { LLAMA_LOG_INFO( "%s: CLS token = %d '%s'\n", __func__, vocab.special_cls_id, vocab.id_to_token[vocab.special_cls_id].text.c_str() ); } + if (vocab.special_mask_id != -1) { LLAMA_LOG_INFO( "%s: MASK token = %d '%s'\n", __func__, vocab.special_mask_id, vocab.id_to_token[vocab.special_mask_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_prefix_id != -1) { LLAMA_LOG_INFO( "%s: PRE token = %d '%s'\n", __func__, vocab.special_prefix_id, vocab.id_to_token[vocab.special_prefix_id].text.c_str() ); } + if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_id].text.c_str() ); } + if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); } + if (vocab.special_eot_id != -1) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, vocab.special_eot_id, vocab.id_to_token[vocab.special_eot_id].text.c_str() ); } } // Returns false if cancelled by progress_callback @@ -4344,6 +4756,13 @@ static bool llm_load_tensors( auto & hparams = model.hparams; +#ifdef GGML_USE_SYCL + // disable MoE with SYCL until mul_mat_id is updated + if (hparams.n_expert > 0) { + n_gpu_layers = 0; + } +#endif + model.split_mode = split_mode; model.main_gpu = main_gpu; model.n_gpu_layers = n_gpu_layers; @@ -4441,7 +4860,7 @@ static bool llm_load_tensors( size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output // for moe merged tensors - ctx_size += ggml_tensor_overhead()*hparams.n_expert*n_layer; + ctx_size += ggml_tensor_overhead()*n_layer*3; std::map ctx_map; for (auto & it : buft_layer_count) { @@ -4637,6 +5056,39 @@ static bool llm_load_tensors( layer.layer_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); } } break; + case LLM_ARCH_DBRX: + { + if (n_expert == 0) { + throw std::runtime_error("DBRX model cannot have zero experts"); + } + + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); + + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); + layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}); + layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}); + layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}); + } + } break; case LLM_ARCH_BAICHUAN: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -4831,6 +5283,50 @@ static bool llm_load_tensors( layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd}); } } break; + case LLM_ARCH_JINA_BERT_V2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // word_embeddings + model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); //token_type_embeddings + model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); // LayerNorm + model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}); //LayerNorm bias + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; // JinaBertLayer + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false); + layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, false); + + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false); + layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, false); + + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); //output_dens + layer.bo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); //output_dens + + layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm + layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + + layer.layer_out_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); + layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd}); + } + } break; case LLM_ARCH_BLOOM: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -4951,8 +5447,13 @@ static bool llm_load_tensors( layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, false); layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, false); - layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + // optional q and k layernorms, present in StableLM 2 12B + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, false); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, false); + + // optional FFN norm, not present in StableLM 2 12B which uses parallel residual + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, false); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, false); layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); @@ -4995,7 +5496,13 @@ static bool llm_load_tensors( // output { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } } for (int i = 0; i < n_layer; ++i) { @@ -5023,16 +5530,14 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); } } break; - case LLM_ARCH_PHI2: + case LLM_ARCH_QWEN2MOE: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // output { - model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); - model.output_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT, "bias"), {n_vocab}); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); } for (int i = 0; i < n_layer; ++i) { @@ -5041,18 +5546,68 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, false); - layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, false); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - if (layer.wqkv == nullptr) { - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); - layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + // optional bias tensors + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); - layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); - layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); + + GGML_ASSERT(hparams.n_expert > 0); + GGML_ASSERT(hparams.n_expert_used > 0); + + // MoE branch + auto n_ff_exp = n_ff / hparams.n_expert_used; + layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); + layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}); + layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); + + // Shared expert branch + layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd}); + layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff}); + layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff, n_embd}); + layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff}); + } + } break; + case LLM_ARCH_PHI2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + model.output_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT, "bias"), {n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, false); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, false); + + if (layer.wqkv == nullptr) { + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); @@ -5068,6 +5623,33 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_PHI3: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context* ctx_layer = ctx_for_layer(i); + ggml_context* ctx_split = ctx_for_layer_split(i); + + auto& layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }); + } + } break; case LLM_ARCH_PLAMO: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -5406,11 +5988,47 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + if (n_layer >= 64){ + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}); + } + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; + case LLM_ARCH_OLMO: // adapted from LLM_ARCH_LLAMA with norm params removed + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); @@ -5566,7 +6184,7 @@ static bool llm_load_tensors( // 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, llama_model_params & params) { try { - llama_model_loader ml(fname, params.use_mmap, params.kv_overrides); + llama_model_loader ml(fname, params.use_mmap, params.check_tensors, params.kv_overrides); model.hparams.vocab_only = params.vocab_only; @@ -5604,6 +6222,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam || !( model.ftype == LLAMA_FTYPE_ALL_F32 || model.ftype == LLAMA_FTYPE_MOSTLY_F16 || + model.ftype == LLAMA_FTYPE_MOSTLY_BF16 || model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 ) @@ -5695,37 +6314,47 @@ static struct ggml_tensor * llm_build_inp_embd( static void llm_build_kv_store( struct ggml_context * ctx, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * k_cur, struct ggml_tensor * v_cur, - int64_t n_ctx, int32_t n_tokens, int32_t kv_head, const llm_build_cb & cb, int64_t il) { + const int64_t n_ctx = cparams.n_ctx; + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(kv.size == n_ctx); - // compute the transposed [n_tokens, n_embd] V matrix - assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); - struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); - cb(v_cur_t, "v_cur_t", il); - struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, (ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa))*kv_head); cb(k_cache_view, "k_cache_view", il); - struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, - ( n_ctx)*ggml_element_size(kv.v_l[il]), - (kv_head)*ggml_element_size(kv.v_l[il])); + // note: storing RoPE-ed version of K in the KV cache + ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); + + assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); + + struct ggml_tensor * v_cache_view = nullptr; + + if (cparams.flash_attn) { + v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, + (kv_head)*ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)); + } else { + // note: the V cache is transposed when not using flash attention + v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, + ( n_ctx)*ggml_element_size(kv.v_l[il]), + (kv_head)*ggml_element_size(kv.v_l[il])); + + v_cur = ggml_transpose(ctx, v_cur); + } cb(v_cache_view, "v_cache_view", il); - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); - ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur_t, v_cache_view)); + ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur, v_cache_view)); } static struct ggml_tensor * llm_build_norm( @@ -5774,7 +6403,7 @@ static struct ggml_tensor * llm_build_ffn( llm_ffn_gate_type type_gate, const llm_build_cb & cb, int il) { - struct ggml_tensor * tmp = ggml_mul_mat(ctx, up, cur); + struct ggml_tensor * tmp = up ? ggml_mul_mat(ctx, up, cur) : cur; cb(tmp, "ffn_up", il); if (up_b) { @@ -5851,24 +6480,117 @@ static struct ggml_tensor * llm_build_ffn( return cur; } -// if max_alibi_bias > 0 then apply ALiBi +static struct ggml_tensor * llm_build_moe_ffn( + struct ggml_context * ctx, + struct ggml_tensor * cur, + struct ggml_tensor * gate_inp, + struct ggml_tensor * up_exps, + struct ggml_tensor * gate_exps, + struct ggml_tensor * down_exps, + int64_t n_expert, + int64_t n_expert_used, + llm_ffn_op_type type_op, + bool norm_w, + const llm_build_cb & cb, + int il) { + int64_t n_embd = cur->ne[0]; + int64_t n_tokens = cur->ne[1]; + + ggml_tensor * logits = ggml_mul_mat(ctx, gate_inp, cur); // [n_expert, n_tokens] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens] + cb(probs, "ffn_moe_probs", il); + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_expert_used); // [n_expert_used, n_tokens] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + cb(selected_experts, "ffn_moe_topk", il); + + ggml_tensor * weights = ggml_get_rows(ctx, + ggml_reshape_3d(ctx, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights", il); + + if (norm_w) { + weights = ggml_reshape_2d(ctx, weights, n_expert_used, n_tokens); + + ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens] + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights_norm", il); + + weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens); + } + + cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens); + ggml_tensor * up = ggml_mul_mat_id(ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(up, "ffn_moe_up", il); + + ggml_tensor * gate = ggml_mul_mat_id(ctx, gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(gate, "ffn_moe_gate", il); + + switch (type_op) { + case LLM_FFN_SILU: + { + gate = ggml_silu(ctx, gate); + cb(gate, "ffn_moe_silu", il); + } break; + case LLM_FFN_GELU: + { + gate = ggml_gelu(ctx, gate); + cb(gate, "ffn_moe_gelu", il); + } break; + default: + GGML_ASSERT(false); + } + + ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens] + cb(par, "ffn_moe_gate_par", il); + + ggml_tensor * experts = ggml_mul_mat_id(ctx, down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens] + cb(experts, "ffn_moe_down", il); + + experts = ggml_mul(ctx, experts, weights); + + // aggregate experts + ggml_tensor * moe_out = nullptr; + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens, + experts->nb[2], i*experts->nb[1]); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx, moe_out, cur_expert); + } + } + + if (n_expert_used == 1) { + // avoid returning a non-contiguous tensor + moe_out = ggml_cont(ctx, moe_out); + } + + return moe_out; +} + static struct ggml_tensor * llm_build_kqv( struct ggml_context * ctx, const llama_model & model, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * wo, struct ggml_tensor * wo_b, struct ggml_tensor * q_cur, struct ggml_tensor * kq_mask, - struct ggml_tensor * kq_pos, - int64_t n_ctx, int32_t n_tokens, int32_t n_kv, float kq_scale, const llm_build_cb & cb, int il) { + const int64_t n_ctx = cparams.n_ctx; const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head_k = hparams.n_embd_head_k; @@ -5886,71 +6608,75 @@ static struct ggml_tensor * llm_build_kqv( 0); cb(k, "k", il); - struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); - cb(kq, "kq", il); + struct ggml_tensor * cur; - 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 (cparams.flash_attn) { + GGML_UNUSED(model); + GGML_UNUSED(n_ctx); - if (model.arch == LLM_ARCH_GROK) { - // need to do the following: - // multiply by attn_output_multiplyer of 0.08838834764831845 - // and then : - // kq = 30 * tanh(kq / 30) - // before the softmax below + // split cached v into n_head heads (not transposed) + struct ggml_tensor * v = + ggml_view_3d(ctx, kv.v_l[il], + n_embd_head_v, n_kv, n_head_kv, + ggml_row_size(kv.v_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv.v_l[il]->type, n_embd_head_k), + 0); + cb(v, "v", il); - //try from phi2 - //ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias); - kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); - kq = ggml_scale(ctx, kq, 30); - } + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) { + ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); + } -#if defined(GGML_USE_KOMPUTE) -#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute") -#pragma message(" Falling back to ggml_alibi(). Will become an error in Mar 2024") -#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5488") - if (hparams.f_max_alibi_bias > 0.0f) { - kq = ggml_scale(ctx, kq, kq_scale); - cb(kq, "kq_scaled", il); + cur = ggml_reshape_2d(ctx, cur, n_embd_head_k*n_head, n_tokens); + } else { + struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); + cb(kq, "kq", il); + + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) { + // 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); + } - kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias); - cb(kq, "kq_scaled_alibi", il); + if (model.arch == LLM_ARCH_GROK) { + // need to do the following: + // multiply by attn_output_multiplyer of 0.08838834764831845 + // and then : + // kq = 30 * tanh(kq / 30) + // before the softmax below - kq = ggml_add(ctx, kq, kq_mask); - cb(kq, "kq_masked", il); + //try from phi2 + //ggml_mul_mat_set_prec(kq, GGML_PREC_F32); - kq = ggml_soft_max(ctx, kq); - cb(kq, "kq_soft_max", il); - } else -#endif - { - kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias); + kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); + kq = ggml_scale(ctx, kq, 30); + } + + kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias); cb(kq, "kq_soft_max_ext", il); - } - GGML_ASSERT(kv.size == n_ctx); + GGML_ASSERT(kv.size == n_ctx); - // split cached v into n_head heads - struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v_l[il], - n_kv, n_embd_head_v, n_head_kv, - ggml_element_size(kv.v_l[il])*n_ctx, - ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v, - 0); - cb(v, "v", il); + // split cached v into n_head heads + struct ggml_tensor * v = + ggml_view_3d(ctx, kv.v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv.v_l[il])*n_ctx, + ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v, + 0); + cb(v, "v", il); - struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); - cb(kqv, "kqv", il); + struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); + cb(kqv, "kqv", il); - struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); + struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); - struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); - cb(cur, "kqv_merged_cont", il); + cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); + cb(cur, "kqv_merged_cont", il); + } ggml_build_forward_expand(graph, cur); @@ -5970,6 +6696,7 @@ static struct ggml_tensor * llm_build_kv( struct ggml_context * ctx, const llama_model & model, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * wo, @@ -5978,8 +6705,6 @@ static struct ggml_tensor * llm_build_kv( struct ggml_tensor * v_cur, struct ggml_tensor * q_cur, struct ggml_tensor * kq_mask, - struct ggml_tensor * kq_pos, - int64_t n_ctx, int32_t n_tokens, int32_t kv_head, int32_t n_kv, @@ -5993,12 +6718,12 @@ static struct ggml_tensor * llm_build_kv( ggml_build_forward_expand(graph, k_cur); ggml_build_forward_expand(graph, v_cur); - llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); + llm_build_kv_store(ctx, hparams, cparams, kv, graph, k_cur, v_cur, n_tokens, kv_head, cb, il); struct ggml_tensor * cur; - cur = llm_build_kqv(ctx, model, hparams, kv, graph, wo, wo_b, - q_cur, kq_mask, kq_pos, n_ctx, n_tokens, n_kv, kq_scale, cb, il); + cur = llm_build_kqv(ctx, model, hparams, cparams, kv, graph, wo, wo_b, + q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il); cb(cur, "kqv_out", il); return cur; @@ -6040,6 +6765,8 @@ struct llm_build_context { const int32_t kv_head; // index of where we store new KV data in the cache const int32_t n_orig_ctx; + const bool flash_attn; + const enum llama_pooling_type pooling_type; const enum llama_rope_type rope_type; @@ -6086,6 +6813,7 @@ struct llm_build_context { n_outputs (worst_case ? n_tokens : lctx.n_outputs), kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), n_orig_ctx (cparams.n_yarn_orig_ctx), + flash_attn (cparams.flash_attn), pooling_type (cparams.pooling_type), rope_type (hparams.rope_type), cb (cb), @@ -6102,18 +6830,17 @@ struct llm_build_context { ctx0 = ggml_init(params); - lctx.inp_tokens = nullptr; - lctx.inp_embd = nullptr; - lctx.inp_pos = nullptr; + lctx.inp_tokens = nullptr; + lctx.inp_embd = nullptr; + lctx.inp_pos = nullptr; lctx.inp_out_ids = nullptr; lctx.inp_KQ_mask = nullptr; - lctx.inp_KQ_pos = nullptr; lctx.inp_K_shift = nullptr; - lctx.inp_mean = nullptr; - lctx.inp_cls = nullptr; - lctx.inp_s_copy = nullptr; - lctx.inp_s_mask = nullptr; - lctx.inp_s_seq = nullptr; + lctx.inp_mean = nullptr; + lctx.inp_cls = nullptr; + lctx.inp_s_copy = nullptr; + lctx.inp_s_mask = nullptr; + lctx.inp_s_seq = nullptr; } void free() { @@ -6200,15 +6927,31 @@ struct llm_build_context { ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*id)); - ggml_tensor * view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, i)); + ggml_tensor * view_v_src; + ggml_tensor * view_v_dst; - ggml_tensor * view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, id)); + if (flash_attn) { + // NOTE: the V cache is not transposed when using flash attention + view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*id)); + } else { + view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self.v_l[il]->type, kv_self.size), + ggml_row_size(kv_self.v_l[il]->type, i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self.v_l[il]->type, kv_self.size), + ggml_row_size(kv_self.v_l[il]->type, id)); + } ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); @@ -6238,20 +6981,13 @@ struct llm_build_context { struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { if (causal) { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, n_tokens); + lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); } else { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); } cb(lctx.inp_KQ_mask, "KQ_mask", -1); ggml_set_input(lctx.inp_KQ_mask); - return lctx.inp_KQ_mask; - } - - struct ggml_tensor * build_inp_KQ_pos() { - lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_kv); - cb(lctx.inp_KQ_pos, "KQ_pos", -1); - ggml_set_input(lctx.inp_KQ_pos); - return lctx.inp_KQ_pos; + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask; } struct ggml_tensor * build_inp_mean() { @@ -6357,9 +7093,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -6394,62 +7130,15 @@ struct llm_build_context { 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_exps, 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_exps, 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); - cb(cur_expert, "ffn_moe_gate_par", il); - - cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, 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 = llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + cb, il); + cb(cur, "ffn_moe_out", il); } cur = ggml_add(ctx0, cur, ffn_inp); @@ -6499,9 +7188,6 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - // positions of the tokens in the KV cache - struct ggml_tensor * KQ_pos = build_inp_KQ_pos(); - for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * inpSA = inpL; @@ -6544,9 +7230,9 @@ struct llm_build_context { cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -6616,9 +7302,6 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - // positions of the tokens in the KV cache - struct ggml_tensor * KQ_pos = build_inp_KQ_pos(); - for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * inpSA = inpL; @@ -6651,9 +7334,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -6771,9 +7454,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -6896,9 +7579,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -6928,63 +7611,15 @@ struct llm_build_context { 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_exps, 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_exps, selected_experts, i, cur); - cb(cur_gate, "ffn_moe_gate", il); - - //GeLU - cur_gate = ggml_gelu(ctx0, cur_gate); - cb(cur_gate, "ffn_moe_gelu", il); - - cur_expert = ggml_mul(ctx0, cur_up, cur_gate); - cb(cur_expert, "ffn_moe_gate_par", il); - - cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, 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 = llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_GELU, true, + cb, il); + cb(cur, "ffn_moe_out", il); // Grok // if layer_out_norm is present then apply it before adding the input @@ -6996,7 +7631,6 @@ struct llm_build_context { cb(cur, "layer_out_norm", il); } - cur = ggml_add(ctx0, cur, ffn_inp); cb(cur, "ffn_out", il); @@ -7032,12 +7666,16 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_starcoder() { + struct ggml_cgraph * build_dbrx() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + const int64_t n_embd_head = hparams.n_embd_head_v; const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); struct ggml_tensor * cur; struct ggml_tensor * inpL; @@ -7050,59 +7688,183 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - struct ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - 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); + model.layers[il].attn_norm, NULL, + LLM_NORM, cb, il); cb(cur, "attn_norm", il); // self-attention { + struct ggml_tensor * Qcur = nullptr; + struct ggml_tensor * Kcur = nullptr; + struct ggml_tensor * Vcur = nullptr; + 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); + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", 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))); + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + 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); + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { // skip computing output for unused tokens struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - // add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); + // feed-forward network + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].attn_out_norm, NULL, + LLM_NORM, cb, il); + cb(cur, "attn_out_norm", il); + + cur = llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + cb, il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + 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); + + // 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_starcoder() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + struct ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + 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 + { + 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_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); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + 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, @@ -7285,9 +8047,9 @@ struct llm_build_context { ); cb(Vcur, "Vcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Q, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Q, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7353,9 +8115,6 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - // positions of the tokens in the KV cache - struct ggml_tensor * KQ_pos = build_inp_KQ_pos(); - for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * inpSA = inpL; @@ -7381,9 +8140,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); cb(Qcur, "Qcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7445,8 +8204,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; + struct ggml_tensor * inp_pos = nullptr; - struct ggml_tensor * inp_pos = build_inp_pos(); + if (model.arch != LLM_ARCH_JINA_BERT_V2) { + inp_pos = build_inp_pos(); + } struct ggml_tensor * inp_mean = build_inp_mean(); struct ggml_tensor * inp_cls = build_inp_cls(); @@ -7477,13 +8239,26 @@ struct llm_build_context { struct ggml_tensor * Vcur; // self-attention - if (model.arch == LLM_ARCH_BERT) { + if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) { Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq); cb(Qcur, "Qcur", il); + if (model.layers[il].attn_q_norm) { + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, cb, il); + } + Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk); cb(Kcur, "Kcur", il); + if (model.layers[il].attn_k_norm) { + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, cb, il); + } Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv); cb(Vcur, "Vcur", il); @@ -7523,7 +8298,7 @@ struct llm_build_context { struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); cb(kq, "kq", il); - kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, nullptr, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias); + kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias); cb(kq, "kq_soft_max_ext", il); struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); @@ -7574,6 +8349,13 @@ struct llm_build_context { model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, cb, il); } else { cur = llm_build_ffn(ctx0, cur, model.layers[il].ffn_up, NULL, @@ -7640,9 +8422,6 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - // positions of the tokens in the KV cache - struct ggml_tensor * KQ_pos = build_inp_KQ_pos(); - inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, @@ -7674,9 +8453,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7741,9 +8520,6 @@ struct llm_build_context { // KQ_mask (mask for 1 head, it will be broadcasted to all heads) struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - // positions of the tokens in the KV cache - struct ggml_tensor * KQ_pos = build_inp_KQ_pos(); - if (model.pos_embd) { // inp_pos - contains the positions struct ggml_tensor * inp_pos = build_inp_pos(); @@ -7805,14 +8581,15 @@ struct llm_build_context { 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); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } else { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } } @@ -7884,7 +8661,7 @@ struct llm_build_context { struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; + // norm cur = llm_build_norm(ctx0, inpL, hparams, @@ -7893,6 +8670,8 @@ struct llm_build_context { LLM_NORM, cb, il); cb(cur, "attn_norm", il); + struct ggml_tensor * inpSA = cur; + // self-attention { // compute Q and K and RoPE them @@ -7917,43 +8696,69 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + + if (model.layers[il].attn_q_norm) { + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, cb, il); + cb(Qcur, "Qcur", il); + } + if (model.layers[il].attn_k_norm) { + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, cb, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + ctx0, Qcur, inp_pos, n_rot, rope_type, 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, + ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { // skip computing output for unused tokens struct ggml_tensor * inp_out_ids = build_inp_out_ids(); cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); 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); - + if (model.layers[il].ffn_norm) { + 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); + } else { + // parallel residual + cur = inpSA; + } cur = llm_build_ffn(ctx0, cur, model.layers[il].ffn_up, NULL, model.layers[il].ffn_gate, NULL, @@ -8044,9 +8849,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8143,12 +8948,6 @@ struct llm_build_context { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); cb(Vcur, "Vcur", il); - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(gf, Qcur); - ggml_build_forward_expand(gf, Kcur); - ggml_build_forward_expand(gf, Vcur); - Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, @@ -8163,9 +8962,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8215,16 +9014,17 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_phi2() { + struct ggml_cgraph * build_qwen2moe() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); 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, lctx, hparams, batch, model.tok_embd, cb); @@ -8236,79 +9036,222 @@ struct llm_build_context { struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); 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 - { - struct ggml_tensor * Qcur = nullptr; - struct ggml_tensor * Kcur = nullptr; - struct ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv) { - 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 * inpSA = inpL; - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - 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))); - } else { - Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); - Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); - Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); - } + // norm + 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 + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + 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); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); 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); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 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, n_rot, rope_type, 0, n_orig_ctx, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { // skip computing output for unused tokens struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - // FF + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // 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 * moe_out = + llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + cb, il); + cb(cur, "ffn_moe_out", il); + + // FFN shared expert { - 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, + ggml_tensor * cur_gate_inp = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp_shexp, cur); + cb(cur_gate_inp, "ffn_shexp_gate_inp", il); + + // sigmoid + ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); + cb(cur_gate, "ffn_shexp_gate", il); + + ggml_tensor * cur_ffn = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up_shexp, NULL, + model.layers[il].ffn_gate_shexp, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur_ffn, "ffn_shexp", il); + + ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); + cb(ffn_shexp_out, "ffn_shexp_out", il); + + moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); + cb(moe_out, "ffn_out", il); + + cur = moe_out; + } + + 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); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + 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, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + 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 + { + struct ggml_tensor * Qcur = nullptr; + struct ggml_tensor * Kcur = nullptr; + struct ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + 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); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + 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))); + } else { + Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + 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, n_rot, rope_type, 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, n_rot, rope_type, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); + } + + // 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, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); cb(ffn_output, "ffn_out", il); } @@ -8332,12 +9275,140 @@ struct llm_build_context { cur = ggml_add(ctx0, cur, model.output_b); cb(cur, "result_output", -1); + ggml_build_forward_expand(gf, cur); + return gf; + } + + struct ggml_cgraph * build_phi3() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + auto residual = inpL; + + // self-attention + { + struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + NULL, + LLM_NORM_RMS, cb, il); + cb(attn_norm_output, "attn_norm", il); + + struct ggml_tensor * Qcur = nullptr; + struct ggml_tensor * Kcur = nullptr; + struct ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); + 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))); + } + else { + Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + 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, n_rot, rope_type, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor* inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + } + + cur = ggml_add(ctx0, cur, residual); + residual = cur; + + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // FF + // special-case: the up and gate tensors are merged into a single tensor + // TOOD: support into llm_build_ffn + { + struct ggml_tensor* up = ggml_mul_mat(ctx0, model.layers[il].ffn_up, cur); + cb(up, "ffn_up", il); + + auto g = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), 0)); + auto y = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), up->nb[1] / 2)); + + y = ggml_mul(ctx0, y, ggml_silu(ctx0, g)); + cb(y, "ffn_gate", il); + + auto down = ggml_mul_mat(ctx0, model.layers[il].ffn_down, y); + cb(down, "ffn_down", il); + + cur = down; + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, residual, cur); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + NULL, + LLM_NORM_RMS, 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_plamo() { struct ggml_cgraph * gf = ggml_new_graph(ctx0); @@ -8390,9 +9461,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } struct ggml_tensor * sa_out = cur; @@ -8493,9 +9564,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8600,9 +9671,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8716,9 +9787,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8833,9 +9904,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8963,9 +10034,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9084,9 +10155,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -9203,9 +10274,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9454,6 +10525,31 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } + if (model.layers[il].attn_q_norm) { + Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur) * n_embd_head, + ggml_element_size(Qcur) * n_embd_head * n_head, + 0); + cb(Qcur, "Qcur", il); + Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, + ggml_element_size(Kcur) * n_embd_head, + ggml_element_size(Kcur) * n_embd_head * n_head_kv, + 0); + cb(Kcur, "Kcur", il); + + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, cb, il); + cb(Qcur, "Qcur", il); + + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, cb, il); + cb(Kcur, "Kcur", il); + } + Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, @@ -9468,9 +10564,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9524,28 +10620,161 @@ struct llm_build_context { return gf; } -}; -static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { - llama_batch dummy; - dummy.n_tokens = 0; + // ref: https://allenai.org/olmo + // based on the original build_llama() function, changes: + // * non-parametric layer norm + // * clamp qkv + // * removed bias + // * removed MoE + struct ggml_cgraph * build_olmo() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); - llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; - struct llm_build_context llm(lctx, dummy, cb, false); + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); - llm.init(); + struct ggml_tensor * cur; + struct ggml_tensor * inpL; - struct ggml_cgraph * result = llm.build_defrag(ids); + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); - llm.free(); + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); - return result; -} + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); -static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { - llama_batch dummy; - dummy.n_tokens = 0; + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + NULL, NULL, + 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); + if (hparams.f_clamp_kqv > 0.0f) { + Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 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, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, nullptr, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + 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, + NULL, NULL, + 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, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + NULL, NULL, + 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; + } +}; + +static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { + llama_batch dummy; + dummy.n_tokens = 0; + + llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; + + struct llm_build_context llm(lctx, dummy, cb, false); + + llm.init(); + + struct ggml_cgraph * result = llm.build_defrag(ids); + + llm.free(); + + return result; +} + +static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { + llama_batch dummy; + dummy.n_tokens = 0; llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; @@ -9649,6 +10878,7 @@ static struct ggml_cgraph * llama_build_graph( result = llm.build_refact(); } break; case LLM_ARCH_BERT: + case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_NOMIC_BERT: { result = llm.build_bert(); @@ -9673,10 +10903,18 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_qwen2(); } break; + case LLM_ARCH_QWEN2MOE: + { + result = llm.build_qwen2moe(); + } break; case LLM_ARCH_PHI2: { result = llm.build_phi2(); } break; + case LLM_ARCH_PHI3: + { + result = llm.build_phi3(); + } break; case LLM_ARCH_PLAMO: { result = llm.build_plamo(); @@ -9721,6 +10959,14 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_command_r(); } break; + case LLM_ARCH_DBRX: + { + result = llm.build_dbrx(); + } break; + case LLM_ARCH_OLMO: + { + result = llm.build_olmo(); + } break; default: GGML_ASSERT(false); } @@ -9840,11 +11086,21 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { f = -INFINITY; } else { - f = 0.0f; + if (hparams.use_alibi) { + f = -fabs(lctx.kv_self.cells[i].pos - pos); + } else { + f = 0.0f; + } } data[h*(n_kv*n_tokens) + j*n_kv + i] = f; } } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + } + } } } else { // when using kv cache, the mask needs to match the kv cache size @@ -9863,7 +11119,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { float f = -INFINITY; for (int s = 0; s < batch.n_seq_id[i]; ++s) { if (batch.seq_id[i][s] == seq_id) { - f = 0.0f; + if (hparams.use_alibi) { + f = -fabs(batch.pos[i] - batch.pos[j]); + } else { + f = 0.0f; + } break; } } @@ -9879,19 +11139,6 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (hparams.need_kq_pos) { - const int64_t n_kv = kv_self.n; - - GGML_ASSERT(lctx.inp_KQ_pos); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_pos->buffer)); - - float * data = (float *) lctx.inp_KQ_pos->data; - - for (int i = 0; i < n_kv; ++i) { - data[i] = float(lctx.kv_self.cells[i].pos); - } - } - if (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { const int64_t n_tokens = batch.n_tokens; @@ -10261,7 +11508,7 @@ 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::min(kv_self.size, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); + kv_self.n = std::min(kv_self.size, std::max(256u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 256))); //kv_self.n = llama_kv_cache_cell_max(kv_self); } } @@ -10411,6 +11658,9 @@ static int llama_decode_internal( n_outputs_prev += lctx.n_outputs; } + // set to total number of outputs in the batch, for use in llama_get_logits_ith + lctx.n_outputs = n_outputs; + // wait for the computation to finish (automatically done when obtaining the model output) //llama_synchronize(&lctx); @@ -10426,6 +11676,10 @@ static int llama_decode_internal( } } + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(lctx.sched); + return 0; } @@ -10451,7 +11705,9 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { // each move requires 6*n_layer tensors (see build_defrag) // - source view, destination view, copy operation // - x2 for keys and values - const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer); + //const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer); + // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516 + const uint32_t max_moves = (LLAMA_MAX_NODES - 2*n_layer)/(6*n_layer); // determine which KV cells to move where // @@ -10767,7 +12023,7 @@ static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); GGML_ASSERT(llama_is_byte_token(vocab, id)); - const auto& token_data = vocab.id_to_token.at(id); + const auto & token_data = vocab.id_to_token.at(id); switch (llama_vocab_get_type(vocab)) { case LLAMA_VOCAB_TYPE_SPM: { auto buf = token_data.text.substr(3, 2); @@ -10775,7 +12031,7 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { } case LLAMA_VOCAB_TYPE_BPE: { GGML_ASSERT(false); - return unicode_utf8_to_byte(token_data.text); + return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT? } case LLAMA_VOCAB_TYPE_WPM: { GGML_ASSERT(false); @@ -10997,7 +12253,101 @@ struct llm_tokenizer_bpe { void tokenize(const std::string & text, std::vector & output) { int final_prev_index = -1; - auto word_collection = bpe_gpt2_preprocess(text); + bool ignore_merges = false; + + std::vector word_collection; + switch (vocab.type) { + case LLAMA_VOCAB_TYPE_BPE: + switch (vocab.type_pre) { + case LLAMA_VOCAB_PRE_TYPE_LLAMA3: + ignore_merges = true; + word_collection = unicode_regex_split(text, { + // original regex from tokenizer.json + //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + + // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_DBRX: + word_collection = unicode_regex_split(text, { + // same as llama3 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: + word_collection = unicode_regex_split(text, { + "[\r\n]", + "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", + "\\s?[!-/:-~！-／：-～‘-‟　-。]+", + "\\s+$", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}+", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: + word_collection = unicode_regex_split(text, { + "[\r\n]", + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_FALCON: + word_collection = unicode_regex_split(text, { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "[0-9][0-9][0-9]", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_MPT: + // TODO: MPT pre-tokenization regexes are unknown + // the following are close, but not exact. run the following: + // ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf + GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed"); + word_collection = unicode_regex_split(text, { + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_STARCODER: + case LLAMA_VOCAB_PRE_TYPE_REFACT: + case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: + word_collection = unicode_regex_split(text, { + "\\p{N}", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_GPT2: + case LLAMA_VOCAB_PRE_TYPE_OLMO: + word_collection = unicode_regex_split(text, { + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }); + break; + case LLAMA_VOCAB_PRE_TYPE_QWEN2: + word_collection = unicode_regex_split(text, { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }); + break; + default: + // default regex for BPE tokenization pre-processing + word_collection = unicode_regex_split(text, { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "\\p{N}+", + "[0-9][0-9][0-9]", + }); + break; + } + break; + default: + GGML_ASSERT(false); + break; + } symbols_final.clear(); @@ -11008,6 +12358,11 @@ struct llm_tokenizer_bpe { int index = 0; size_t offset = 0; + if (ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) { + symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()}); + offset = word.size(); + } + while (offset < word.size()) { llm_symbol sym; size_t char_len = std::min(word.size() - offset, (size_t) ::utf8_len(word[offset])); @@ -11054,7 +12409,7 @@ struct llm_tokenizer_bpe { add_new_bigram(bigram.left, left_symbol.next); // right side of current symbol } - // add the fnished tokens to the final list keeping correct order for next and prev + // add the finished tokens to the final list keeping correct order for next and prev for (auto & sym : symbols) { if (sym.n > 0) { sym.prev = final_prev_index; @@ -11124,145 +12479,6 @@ private: work_queue.push(bigram); } - std::vector bpe_gpt2_preprocess(const std::string & text) { - std::vector bpe_words; - std::vector bpe_encoded_words; - - std::string token = ""; - // GPT2 system regex: 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+ - bool collecting_numeric = false; - bool collecting_letter = false; - bool collecting_special = false; - bool collecting_whitespace_lookahead = false; - bool collecting = false; - - std::vector text_utf; - text_utf.reserve(text.size()); - bpe_words.reserve(text.size()); - bpe_encoded_words.reserve(text.size()); - - const auto cpts = unicode_cpts_from_utf8(text); - for (size_t i = 0; i < cpts.size(); ++i) - text_utf.emplace_back(unicode_cpt_to_utf8(cpts[i])); - - for (int i = 0; i < (int)text_utf.size(); i++) { - const std::string & utf_char = text_utf[i]; - bool split_condition = false; - int bytes_remain = text_utf.size() - i; - // forward backward lookups - const std::string & utf_char_next = (i + 1 < (int)text_utf.size()) ? text_utf[i + 1] : ""; - const std::string & utf_char_next_next = (i + 2 < (int)text_utf.size()) ? text_utf[i + 2] : ""; - - // handling contractions - if (!split_condition && bytes_remain >= 2) { - // 's|'t|'m|'d - if (utf_char == "\'" && (utf_char_next == "s" || utf_char_next == "t" || utf_char_next == "m" || utf_char_next == "d")) { - split_condition = true; - } - if (split_condition) { - if (token.size()) { - bpe_words.emplace_back(token); // push previous content as token - } - token = utf_char + utf_char_next; - bpe_words.emplace_back(token); - token = ""; - i++; - continue; - } - } - if (!split_condition && bytes_remain >= 3) { - // 're|'ve|'ll - if (utf_char == "\'" && ( - (utf_char_next == "r" && utf_char_next_next == "e") || - (utf_char_next == "v" && utf_char_next_next == "e") || - (utf_char_next == "l" && utf_char_next_next == "l")) - ) { - split_condition = true; - } - if (split_condition) { - // current token + next token can be defined - if (token.size()) { - bpe_words.emplace_back(token); // push previous content as token - } - token = utf_char + utf_char_next + utf_char_next_next; - bpe_words.emplace_back(token); // the contraction - token = ""; - i += 2; - continue; - } - } - - if (!split_condition && !collecting) { - if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || (!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) { - collecting_letter = true; - collecting = true; - } - else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || (!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) { - collecting_numeric = true; - collecting = true; - } - else if ( - ((unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (unicode_cpt_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) || - (!token.size() && utf_char == " " && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) - ) { - collecting_special = true; - collecting = true; - } - else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) { - collecting_whitespace_lookahead = true; - collecting = true; - } - else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) { - split_condition = true; - } - } - else if (!split_condition && collecting) { - if (collecting_letter && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER) { - split_condition = true; - } - else if (collecting_numeric && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) { - split_condition = true; - } - else if (collecting_special && (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) { - split_condition = true; - } - else if (collecting_whitespace_lookahead && (unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) { - split_condition = true; - } - } - - if (utf_char_next == "") { - split_condition = true; // final - token += utf_char; - } - - if (split_condition) { - if (token.size()) { - bpe_words.emplace_back(token); - } - token = utf_char; - collecting = false; - collecting_letter = false; - collecting_numeric = false; - collecting_special = false; - collecting_whitespace_lookahead = false; - } - else { - token += utf_char; - } - } - - for (std::string & word : bpe_words) { - std::string encoded_token = ""; - for (char & c : word) { - encoded_token += unicode_byte_to_utf8(c); - } - bpe_encoded_words.emplace_back(encoded_token); - } - - return bpe_encoded_words; - } - const llama_vocab & vocab; std::vector symbols; @@ -11323,9 +12539,6 @@ struct llm_tokenizer_wpm { output.push_back(vocab.special_unk_id); } } - - // append eos token - output.push_back(vocab.special_eos_id); } std::vector preprocess(const std::string & text) { @@ -11340,7 +12553,7 @@ struct llm_tokenizer_wpm { continue; } code = unicode_tolower(code); - if (type == CODEPOINT_TYPE_WHITESPACE) { + if (type == CODEPOINT_TYPE_SEPARATOR) { code = ' '; } std::string s = unicode_cpt_to_utf8(code); @@ -11530,30 +12743,28 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< } } -static std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special) { +static std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) { std::vector output; - - // OG tokenizer behavior: - // - // tokenizer.encode('', add_bos=True) returns [1] - // tokenizer.encode('', add_bos=False) returns [] - - if (bos && vocab.special_bos_id != -1) { - output.push_back(vocab.special_bos_id); - } - - if (raw_text.empty()) { - return output; - } - std::forward_list fragment_buffer; - fragment_buffer.emplace_front(raw_text, 0, raw_text.length()); - if (special) tokenizer_st_partition(vocab, fragment_buffer); + if (!raw_text.empty()) { + fragment_buffer.emplace_front(raw_text, 0, raw_text.length()); + if (parse_special) tokenizer_st_partition(vocab, fragment_buffer); + } switch (vocab.type) { case LLAMA_VOCAB_TYPE_SPM: { + // OG tokenizer behavior: + // + // tokenizer.encode('', add_special_tokens=True) returns [1] + // tokenizer.encode('', add_special_tokens=False) returns [] + + if (add_special && vocab.special_add_bos != 0) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + } + for (const auto & fragment : fragment_buffer) { if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { // without adding this leading whitespace, we do not get the same results as the original tokenizer @@ -11579,9 +12790,19 @@ static std::vector llama_tokenize_internal(const llama_vocab & output.push_back(fragment.token); } } + + if (add_special && vocab.special_add_eos == 1) { + GGML_ASSERT(vocab.special_eos_id != -1); + output.push_back(vocab.special_eos_id); + } } break; case LLAMA_VOCAB_TYPE_BPE: { + if (add_special && vocab.special_add_bos != 0) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + } + for (const auto & fragment : fragment_buffer) { if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); @@ -11595,9 +12816,19 @@ static std::vector llama_tokenize_internal(const llama_vocab & output.push_back(fragment.token); } } + + if (add_special && vocab.special_add_eos == 1) { + GGML_ASSERT(vocab.special_add_eos != -1); + output.push_back(vocab.special_eos_id); + } } break; case LLAMA_VOCAB_TYPE_WPM: { + if (add_special) { + GGML_ASSERT(vocab.special_cls_id != -1); + output.push_back(vocab.special_cls_id); + } + for (const auto & fragment : fragment_buffer) { if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); @@ -11611,6 +12842,11 @@ static std::vector llama_tokenize_internal(const llama_vocab & output.push_back(fragment.token); } } + + if (add_special) { + GGML_ASSERT(vocab.special_sep_id != -1); + output.push_back(vocab.special_sep_id); + } } break; case LLAMA_VOCAB_TYPE_NONE: GGML_ASSERT(false); @@ -11777,7 +13013,9 @@ static void llama_grammar_advance_stack( std::vector> & new_stacks) { if (stack.empty()) { - new_stacks.emplace_back(stack); + if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) { + new_stacks.emplace_back(stack); + } return; } @@ -11814,7 +13052,10 @@ static void llama_grammar_advance_stack( } case LLAMA_GRETYPE_CHAR: case LLAMA_GRETYPE_CHAR_NOT: - new_stacks.emplace_back(stack); + if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) { + // only add the stack if it's not a duplicate of one we already have + new_stacks.emplace_back(stack); + } break; default: // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range @@ -11828,12 +13069,13 @@ static void llama_grammar_advance_stack( // be positioned at a character range (see `llama_grammar_advance_stack`), and // produces the N possible stacks if the given char is accepted at those // positions -std::vector> llama_grammar_accept( +void llama_grammar_accept( const std::vector> & rules, const std::vector> & stacks, - const uint32_t chr) { + const uint32_t chr, + std::vector> & new_stacks) { - std::vector> new_stacks; + new_stacks.clear(); for (const auto & stack : stacks) { if (stack.empty()) { @@ -11852,8 +13094,6 @@ std::vector> llama_grammar_accept( llama_grammar_advance_stack(rules, new_stack, new_stacks); } } - - return new_stacks; } static std::vector llama_grammar_reject_candidates( @@ -11867,6 +13107,7 @@ static std::vector llama_grammar_reject_candidates_for_ const std::vector & candidates) { std::vector rejects; + rejects.reserve(candidates.size()); if (stack.empty()) { for (const auto & tok : candidates) { @@ -11880,6 +13121,8 @@ static std::vector llama_grammar_reject_candidates_for_ const llama_grammar_element * stack_pos = stack.back(); std::vector next_candidates; + next_candidates.reserve(candidates.size()); + for (const auto & tok : candidates) { if (*tok.code_points == 0) { // reached end of full codepoints in token, reject iff it ended in a partial sequence @@ -12465,16 +13708,14 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c GGML_ASSERT(ctx); const int64_t t_start_sample_us = ggml_time_us(); - bool allow_eos = false; + bool allow_eog = false; for (const auto & stack : grammar->stacks) { if (stack.empty()) { - allow_eos = true; + allow_eog = true; break; } } - 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; @@ -12482,9 +13723,10 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c for (size_t i = 0; i < candidates->size; ++i) { const llama_token id = candidates->data[i].id; - const std::string piece = llama_token_to_piece(ctx, id); - if (id == eos) { - if (!allow_eos) { + const std::string piece = llama_token_to_piece(ctx, id, false); + + if (llama_token_is_eog(&ctx->model, id)) { + if (!allow_eog) { candidates->data[i].logit = -INFINITY; } } else if (piece.empty() || piece[0] == 0) { @@ -12647,7 +13889,7 @@ llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_da return result; } -llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) { +llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng) { GGML_ASSERT(ctx); const int64_t t_start_sample_us = ggml_time_us(); @@ -12660,7 +13902,6 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra } std::discrete_distribution<> dist(probs.begin(), probs.end()); - auto & rng = ctx->rng; int idx = dist(rng); llama_token result = candidates->data[idx].id; @@ -12670,10 +13911,14 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra return result; } +llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) { + return llama_sample_token_with_rng(ctx, candidates, ctx->rng); +} + void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) { const int64_t t_start_sample_us = ggml_time_us(); - if (token == llama_token_eos(&ctx->model)) { + if (llama_token_is_eog(&ctx->model, token)) { for (const auto & stack : grammar->stacks) { if (stack.empty()) { return; @@ -12682,13 +13927,15 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar GGML_ASSERT(false); } - const std::string piece = llama_token_to_piece(ctx, token); + const std::string piece = llama_token_to_piece(ctx, token, false); // Note terminating 0 in decoded string const auto decoded = decode_utf8(piece, grammar->partial_utf8); const auto & code_points = decoded.first; + std::vector> tmp_new_stacks; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { - grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it); + llama_grammar_accept(grammar->rules, grammar->stacks, *it, tmp_new_stacks); + grammar->stacks = tmp_new_stacks; } grammar->partial_utf8 = decoded.second; GGML_ASSERT(!grammar->stacks.empty()); @@ -12822,6 +14069,11 @@ struct llama_beam_search_data { } llama_logit_info logit_info(ctx); std::vector next_tokens = logit_info.top_k(n_beams); + + // Clear the kv slot so that other beams may try different tokens at this position. The llama_decode() + // call in loop() will conclusively fill in the kv slot once the beams converge at this position. + llama_kv_cache_seq_rm(ctx, 0, n_past, -1); + size_t i=0; if (next_beams.size() < n_beams) { for (; next_beams.size() < n_beams ; ++i) { @@ -12991,13 +14243,16 @@ static void llama_tensor_dequantize_internal( if (qtype.to_float == NULL) { throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type))); } - } else if (tensor->type != GGML_TYPE_F16) { + } else if (tensor->type != GGML_TYPE_F16 && + tensor->type != GGML_TYPE_BF16) { throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type))); } if (nthread < 2) { if (tensor->type == GGML_TYPE_F16) { ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements); + } else if (tensor->type == GGML_TYPE_BF16) { + ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements); } else if (ggml_is_quantized(tensor->type)) { qtype.to_float(tensor->data, f32_output, nelements); } else { @@ -13006,7 +14261,14 @@ static void llama_tensor_dequantize_internal( return; } - size_t block_size = tensor->type == GGML_TYPE_F16 ? 1 : (size_t)ggml_blck_size(tensor->type); + size_t block_size; + if (tensor->type == GGML_TYPE_F16 || + tensor->type == GGML_TYPE_BF16) { + block_size = 1; + } else { + block_size = (size_t)ggml_blck_size(tensor->type); + } + size_t block_size_bytes = ggml_type_size(tensor->type); GGML_ASSERT(nelements % block_size == 0); @@ -13025,6 +14287,8 @@ static void llama_tensor_dequantize_internal( auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) { if (typ == GGML_TYPE_F16) { ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels); + } else if (typ == GGML_TYPE_BF16) { + ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels); } else { qtype.to_float(inbuf, outbuf, nels); } @@ -13320,21 +14584,27 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n return new_type; } -static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector & workers, const int nthread) { - std::mutex mutex; - int counter = 0; - size_t new_size = 0; +static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector & workers, const int nthread) { if (nthread < 2) { // single-thread - return ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix); + size_t new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix); + if (!ggml_validate_row_data(new_type, new_data, new_size)) { + throw std::runtime_error("quantized data validation failed"); + } + return new_size; } - auto compute = [&mutex, &counter, &new_size, new_type, f32_data, new_data, chunk_size, + + std::mutex mutex; + int64_t counter = 0; + size_t new_size = 0; + bool valid = true; + auto compute = [&mutex, &counter, &new_size, &valid, new_type, f32_data, new_data, chunk_size, nrows, n_per_row, imatrix]() { - const int nrows_per_chunk = chunk_size / n_per_row; + const int64_t nrows_per_chunk = chunk_size / n_per_row; size_t local_size = 0; while (true) { std::unique_lock lock(mutex); - int first_row = counter; counter += nrows_per_chunk; + int64_t first_row = counter; counter += nrows_per_chunk; if (first_row >= nrows) { if (local_size > 0) { new_size += local_size; @@ -13342,8 +14612,18 @@ static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const floa break; } lock.unlock(); - const int this_nrow = std::min(nrows - first_row, nrows_per_chunk); - local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix); + const int64_t this_nrow = std::min(nrows - first_row, nrows_per_chunk); + size_t this_size = ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix); + local_size += this_size; + + // validate the quantized data + const size_t row_size = ggml_row_size(new_type, n_per_row); + void * this_data = (char *) new_data + first_row * row_size; + if (!ggml_validate_row_data(new_type, this_data, this_size)) { + std::unique_lock lock(mutex); + valid = false; + break; + } } }; for (int it = 0; it < nthread - 1; ++it) { @@ -13352,6 +14632,9 @@ static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const floa compute(); for (auto & w : workers) { w.join(); } workers.clear(); + if (!valid) { + throw std::runtime_error("quantized data validation failed"); + } return new_size; } @@ -13366,6 +14649,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break; case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break; case LLAMA_FTYPE_MOSTLY_F16: default_type = GGML_TYPE_F16; break; + case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break; case LLAMA_FTYPE_ALL_F32: default_type = GGML_TYPE_F32; break; // K-quants @@ -13414,7 +14698,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s auto v = (std::vector*)params->kv_overrides; kv_overrides = v->data(); } - llama_model_loader ml(fname_inp, use_mmap, kv_overrides); + llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides); ml.init_mappings(false); // no prefetching llama_model model; @@ -13442,17 +14726,23 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s gguf_set_kv (ctx_out, ml.meta); gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); gguf_set_val_u32(ctx_out, "general.file_type", ftype); + // Remove split metadata + gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_NO).c_str()); + gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str()); + gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str()); if (params->kv_overrides) { const std::vector & overrides = *(const std::vector *)params->kv_overrides; for (auto & o : overrides) { if (o.key[0] == 0) break; if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) { - gguf_set_val_f32(ctx_out, o.key, o.float_value); + gguf_set_val_f32(ctx_out, o.key, o.val_f64); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) { - gguf_set_val_i32(ctx_out, o.key, o.int_value); + gguf_set_val_i32(ctx_out, o.key, o.val_i64); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) { - gguf_set_val_bool(ctx_out, o.key, o.bool_value); + gguf_set_val_bool(ctx_out, o.key, o.val_bool); + } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) { + gguf_set_val_str(ctx_out, o.key, o.val_str); } else { LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key); } @@ -13465,7 +14755,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const std::string name = ggml_get_name(meta); // TODO: avoid hardcoded tensor names - use the TN_* constants - if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) { + if (name.find("attn_v.weight") != std::string::npos || + name.find("attn_qkv.weight") != std::string::npos) { ++qs.n_attention_wv; } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) { qs.has_output = true; @@ -13475,7 +14766,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer; // sanity checks - GGML_ASSERT(qs.n_attention_wv == (int)model.hparams.n_layer && "n_attention_wv != n_layer is unexpected"); + // + // - qs.n_attention_wv == 0 for Mamba models + // - qs.n_attention_wv == model.hparams.n_layer for Transformer models + // + GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); size_t total_size_org = 0; size_t total_size_new = 0; @@ -13489,26 +14784,74 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::vector> work; std::vector> f32_conv_buf; + uint16_t n_split = 1; + // Assume split index is continuous + if (params->keep_split) { + for (int i = 0; i < ml.n_tensors; ++i) { + n_split = std::max(uint16_t(ml.get_weight(i)->idx+1), n_split); + } + } + std::vector ctx_outs(n_split, NULL); + ctx_outs[0] = ctx_out; + // populate the original tensors so we get an initial meta data for (int i = 0; i < ml.n_tensors; ++i) { - const struct ggml_tensor * meta = ml.get_tensor_meta(i); - gguf_add_tensor(ctx_out, meta); + auto weight = ml.get_weight(i); + uint16_t i_split = params->keep_split ? weight->idx : 0; + struct ggml_tensor * tensor = weight->tensor; + if (ctx_outs[i_split] == NULL) { + ctx_outs[i_split] = gguf_init_empty(); + } + gguf_add_tensor(ctx_outs[i_split], tensor); } - std::ofstream fout(fname_out, std::ios::binary); - fout.exceptions(std::ofstream::failbit); // fail fast on write errors - - const size_t meta_size = gguf_get_meta_size(ctx_out); - - LLAMA_LOG_INFO("%s: meta size = %zu bytes\n", __func__, meta_size); + // Set split info if needed + if (n_split > 1) { + for (size_t i = 0; i < ctx_outs.size(); ++i) { + gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i); + gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split); + gguf_set_val_i32(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors); + } + } - // placeholder for the meta data - ::zeros(fout, meta_size); + int cur_split = -1; + std::ofstream fout; + auto close_ofstream = [&]() { + // Write metadata and close file handler + if (fout.is_open()) { + fout.seekp(0); + std::vector data(gguf_get_meta_size(ctx_outs[cur_split])); + gguf_get_meta_data(ctx_outs[cur_split], data.data()); + fout.write((const char *) data.data(), data.size()); + fout.close(); + } + }; + auto new_ofstream = [&](int index) { + cur_split = index; + GGML_ASSERT(ctx_outs[cur_split] && "Find uninitialized gguf_context"); + std::string fname = fname_out; + if (params->keep_split) { + char split_path[PATH_MAX] = {0}; + llama_split_path(split_path, sizeof(split_path), fname_out.c_str(), cur_split, n_split); + fname = std::string(split_path); + } - const auto tn = LLM_TN(model.arch); + fout = std::ofstream(fname, std::ios::binary); + fout.exceptions(std::ofstream::failbit); // fail fast on write errors + const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split]); + // placeholder for the meta data + ::zeros(fout, meta_size); + }; + const auto tn = LLM_TN(model.arch); + new_ofstream(0); for (int i = 0; i < ml.n_tensors; ++i) { - struct ggml_tensor * tensor = ml.get_tensor_meta(i); + auto weight = ml.get_weight(i); + struct ggml_tensor * tensor = weight->tensor; + if (weight->idx != cur_split && params->keep_split) { + close_ofstream(); + new_ofstream(weight->idx); + } const std::string name = ggml_get_name(tensor); @@ -13531,6 +14874,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // quantize only 2D and 3D tensors (experts) quantize &= (ggml_n_dims(tensor) >= 2); + + // do not quantize norm tensors + quantize &= name.find("_norm.weight") == std::string::npos; + quantize &= params->quantize_output_tensor || name != "output.weight"; quantize &= !params->only_copy; @@ -13559,10 +14906,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (!params->pure && ggml_is_quantized(default_type)) { new_type = llama_tensor_get_type(qs, new_type, tensor, ftype); } - else if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) { + if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) { new_type = params->token_embedding_type; } - else if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) { + if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) { new_type = params->output_tensor_type; } @@ -13577,7 +14924,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_size = ggml_nbytes(tensor); LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0); } else { - const size_t nelements = ggml_nelements(tensor); + const int64_t nelements = ggml_nelements(tensor); const float * imatrix = nullptr; if (imatrix_data) { @@ -13629,20 +14976,20 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); - if (work.size() < nelements * 4) { + if (work.size() < (size_t)nelements * 4) { work.resize(nelements * 4); // upper bound on size } new_data = work.data(); - const int n_per_row = tensor->ne[0]; - const int nrows = tensor->ne[1]; + const int64_t n_per_row = tensor->ne[0]; + const int64_t nrows = tensor->ne[1]; - static const int min_chunk_size = 32 * 512; - const int chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row); + static const int64_t min_chunk_size = 32 * 512; + const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row); - const int nelements_matrix = tensor->ne[0] * tensor->ne[1]; - const int nchunk = (nelements_matrix + chunk_size - 1)/chunk_size; - const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1; + const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1]; + const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size; + const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1; // quantize each expert separately since they have different importance matrices new_size = 0; @@ -13659,26 +15006,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s total_size_new += new_size; // update the gguf meta data as we go - gguf_set_tensor_type(ctx_out, name.c_str(), new_type); - gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size); + gguf_set_tensor_type(ctx_outs[cur_split], name.c_str(), new_type); + gguf_set_tensor_data(ctx_outs[cur_split], name.c_str(), new_data, new_size); // write tensor data + padding fout.write((const char *) new_data, new_size); zeros(fout, GGML_PAD(new_size, align) - new_size); } - - // go back to beginning of file and write the updated meta data - { - fout.seekp(0); - std::vector data(gguf_get_meta_size(ctx_out)); - gguf_get_meta_data(ctx_out, data.data()); - fout.write((const char *) data.data(), data.size()); + close_ofstream(); + for (auto & c:ctx_outs) { + gguf_free(c); } - fout.close(); - - gguf_free(ctx_out); - LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); @@ -13722,7 +15061,7 @@ static int llama_apply_lora_from_file_internal( std::unique_ptr ml; 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.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr)); ml->init_mappings(/*prefetch*/ false); // no prefetching } @@ -13981,6 +15320,7 @@ struct llama_model_params llama_model_default_params() { /*.vocab_only =*/ false, /*.use_mmap =*/ true, /*.use_mlock =*/ false, + /*.check_tensors =*/ false, }; #ifdef GGML_USE_METAL @@ -14017,6 +15357,7 @@ struct llama_context_params llama_context_default_params() { /*.logits_all =*/ false, /*.embeddings =*/ false, /*.offload_kqv =*/ true, + /*.flash_attn =*/ false, /*.abort_callback =*/ nullptr, /*.abort_callback_data =*/ nullptr, }; @@ -14034,6 +15375,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { /*.quantize_output_tensor =*/ true, /*.only_copy =*/ false, /*.pure =*/ false, + /*.keep_split =*/ false, /*.imatrix =*/ nullptr, /*.kv_overrides =*/ nullptr, }; @@ -14182,6 +15524,7 @@ struct llama_context * llama_new_context_with_model( cparams.defrag_thold = params.defrag_thold; cparams.embeddings = params.embeddings; cparams.offload_kqv = params.offload_kqv; + cparams.flash_attn = params.flash_attn; cparams.pooling_type = params.pooling_type; cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; @@ -14189,12 +15532,20 @@ struct llama_context * llama_new_context_with_model( cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; // this is necessary due to kv_self.n being padded later during inference - cparams.n_ctx = GGML_PAD(cparams.n_ctx, 32); + cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256); // with causal attention, the batch size is limited by the context size cparams.n_batch = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; - cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); + // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask + // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext) + // ref: https://github.com/ggerganov/llama.cpp/pull/5021 + if (cparams.n_batch < GGML_KQ_MASK_PAD) { + LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD); + cparams.n_batch = GGML_KQ_MASK_PAD; + } + + cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); cparams.n_yarn_orig_ctx = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx : @@ -14226,6 +15577,11 @@ struct llama_context * llama_new_context_with_model( } } + if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) { + LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__); + cparams.flash_attn = false; + } + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } @@ -14233,6 +15589,7 @@ struct llama_context * llama_new_context_with_model( LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); + LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); @@ -14361,7 +15718,7 @@ struct llama_context * llama_new_context_with_model( } ctx->backends.push_back(ctx->backend_cpu); - if (!llama_kv_cache_init(ctx->kv_self, ctx->model, type_k, type_v, kv_size, cparams.offload_kqv)) { + if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); return nullptr; @@ -14514,6 +15871,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_REFACT: case LLM_ARCH_BLOOM: case LLM_ARCH_MAMBA: + case LLM_ARCH_JINA_BERT_V2: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values @@ -14527,18 +15885,22 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MINICPM: case LLM_ARCH_XVERSE: case LLM_ARCH_COMMAND_R: + case LLM_ARCH_OLMO: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 case LLM_ARCH_FALCON: case LLM_ARCH_GROK: + case LLM_ARCH_DBRX: case LLM_ARCH_PERSIMMON: case LLM_ARCH_BERT: case LLM_ARCH_NOMIC_BERT: case LLM_ARCH_STABLELM: case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: + case LLM_ARCH_QWEN2MOE: case LLM_ARCH_PHI2: + case LLM_ARCH_PHI3: case LLM_ARCH_GEMMA: case LLM_ARCH_STARCODER2: return LLAMA_ROPE_TYPE_NEOX; @@ -14552,6 +15914,10 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { return LLAMA_ROPE_TYPE_NONE; } +enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx) { + return ctx->cparams.pooling_type; +} + int32_t llama_n_vocab(const struct llama_model * model) { return model->hparams.n_vocab; } @@ -14907,9 +16273,33 @@ void llama_kv_cache_update(struct llama_context * ctx) { llama_kv_cache_update_internal(*ctx); } +// deprecated +size_t llama_get_state_size(const struct llama_context * ctx) { + return llama_state_get_size(ctx); +} + +// deprecated +size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { + return llama_state_get_data(ctx, dst); +} + +// deprecated +size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { + return llama_state_set_data(ctx, src); +} + +// deprecated +bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); +} + +// deprecated +bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + return llama_state_save_file(ctx, path_session, tokens, n_token_count); +} // Returns the *maximum* size of the state -size_t llama_get_state_size(const struct llama_context * ctx) { +size_t llama_state_get_size(const struct llama_context * ctx) { const auto & cparams = ctx->cparams; const auto & hparams = ctx->model.hparams; @@ -14928,6 +16318,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) { const size_t s_kv_head = sizeof(uint32_t); const size_t s_kv_size = sizeof(uint32_t); const size_t s_kv_used = sizeof(uint32_t); + const size_t s_v_trans = sizeof(uint32_t); const size_t s_kv = ctx->kv_self.total_size(); const size_t s_kv_cell = sizeof(llama_pos) + sizeof(size_t) + cparams.n_seq_max*sizeof(llama_seq_id); const size_t s_kv_cells = ctx->kv_self.size * s_kv_cell; @@ -14945,10 +16336,14 @@ size_t llama_get_state_size(const struct llama_context * ctx) { + s_kv_head + s_kv_size + s_kv_used + + s_v_trans + s_kv + s_kv_cells ); + // on session change it is very likely that the state size has changed - so we need to update this function + static_assert(LLAMA_SESSION_VERSION == 6, "So you just bumped the session version - good. But did you remember to update llama_state_get_size?"); + return s_total; } @@ -14997,15 +16392,17 @@ struct llama_data_file_context : llama_data_context { * file context: * llama_file file("/path", "wb"); * llama_data_file_context data_ctx(&file); - * llama_copy_state_data(ctx, &data_ctx); + * llama_state_get_data(ctx, &data_ctx); * * buffer context: * std::vector buf(max_size, 0); * llama_data_buffer_context data_ctx(&buf.data()); - * llama_copy_state_data(ctx, &data_ctx); + * llama_state_get_data(ctx, &data_ctx); * */ -static void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { +static void llama_state_get_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { + llama_synchronize(ctx); + // copy rng { std::ostringstream rng_ss; @@ -15092,11 +16489,13 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat const uint32_t kv_size = kv_self.size; const size_t kv_buf_size = kv_self.total_size() / (kv_size ? kv_size : 1) * kv_head; const uint32_t kv_used = kv_self.used; + const uint32_t v_trans = kv_self.v_trans ? 1 : 0; 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)); + data_ctx->write(&v_trans, sizeof(v_trans)); if (kv_buf_size) { const size_t pre_kv_buf_size = data_ctx->get_size_written(); @@ -15109,7 +16508,7 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size()); data_ctx->write(tmp_buf.data(), tmp_buf.size()); - if (kv_self.recurrent) { + if (kv_self.recurrent || !kv_self.v_trans) { // v is contiguous for recurrent models // TODO: use other tensors for state models than k and v const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); @@ -15149,15 +16548,17 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat } } -size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { +size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst) { llama_data_buffer_context data_ctx(dst); - llama_copy_state_data_internal(ctx, &data_ctx); + llama_state_get_data_internal(ctx, &data_ctx); return data_ctx.get_size_written(); } // Sets the state reading from the specified source address -size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { +size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { + llama_synchronize(ctx); + const uint8_t * inp = src; // set rng @@ -15194,6 +16595,8 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { GGML_ASSERT((uint32_t) id < ctx->cparams.n_batch); ctx->output_ids[id] = i; } + + ctx->n_outputs = n_outputs; } } @@ -15238,11 +16641,15 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { uint32_t kv_head; uint32_t kv_size; uint32_t kv_used; + uint32_t v_trans; 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); + memcpy(&v_trans, inp, sizeof(v_trans)); inp += sizeof(v_trans); + + GGML_ASSERT(kv_self.v_trans == (bool) v_trans); // incompatible V transposition if (kv_self.size != kv_size) { // the KV cache needs to be big enough to load all the KV cells from the saved state @@ -15252,6 +16659,8 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { __func__, kv_head, kv_size, kv_self.size); } + llama_kv_cache_clear(ctx); + if (kv_buf_size) { const size_t pre_kv_buf_size = inp - src; @@ -15263,7 +16672,7 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); inp += k_size; - if (kv_self.recurrent) { + if (kv_self.recurrent || !kv_self.v_trans) { // v is contiguous for recurrent models // TODO: use other tensors for state models than k and v const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); @@ -15285,8 +16694,6 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { GGML_ASSERT(kv_buf_size == inp - src - pre_kv_buf_size); } - llama_kv_cache_clear(ctx); - ctx->kv_self.head = kv_head; ctx->kv_self.used = kv_used; @@ -15309,14 +16716,14 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { } const size_t nread = inp - src; - const size_t max_size = llama_get_state_size(ctx); + const size_t max_size = llama_state_get_size(ctx); GGML_ASSERT(nread <= max_size); return nread; } -static bool llama_load_session_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { llama_file file(path_session, "rb"); // sanity checks @@ -15354,7 +16761,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c // restore the context state { const size_t n_state_size_cur = file.size - file.tell(); - const size_t n_state_size_max = llama_get_state_size(ctx); + const size_t n_state_size_max = llama_state_get_size(ctx); if (n_state_size_cur > n_state_size_max) { LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur); @@ -15364,22 +16771,22 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c std::vector state_data(n_state_size_max); file.read_raw(state_data.data(), n_state_size_cur); - llama_set_state_data(ctx, state_data.data()); + llama_state_set_data(ctx, state_data.data()); } return true; } -bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { try { - return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); + return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading session file: %s\n", err.what()); return false; } } -bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { +static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { llama_file file(path_session, "wb"); file.write_u32(LLAMA_SESSION_MAGIC); @@ -15393,11 +16800,479 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi // save the context state using stream saving llama_data_file_context data_ctx(&file); - llama_copy_state_data_internal(ctx, &data_ctx); + llama_state_get_data_internal(ctx, &data_ctx); return true; } +bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + try { + return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("error saving session file: %s\n", err.what()); + return false; + } +} + +size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) { + // save the size of size_t as a uint32_t for safety check + const size_t size_t_size_size = sizeof(uint32_t); + + // other values + const size_t s_cell_count_size = sizeof(uint32_t); + const size_t s_layer_count_size = sizeof(uint32_t); + const size_t n_embd_v_gqa_size = sizeof(uint32_t); + + size_t s_cell_count = 0; + size_t s_cell_data_size = 0; + const auto & kv_self = ctx->kv_self; + const auto & hparams = ctx->model.hparams; + + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); + + for (uint32_t i = 0; i < kv_self.size; ++i) { + const auto & cell = kv_self.cells[i]; + if (cell.seq_id.count(seq_id) > 0) { + ++s_cell_count; + s_cell_data_size += sizeof(llama_pos); + } + } + + for (int il = 0; il < (int)n_layer; ++il) { + // types of keys and values + s_cell_data_size += sizeof(int32_t) * 2; + // k_size_row and v_size_el values of layer + s_cell_data_size += sizeof(size_t) * 2; + + // keys + const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); + s_cell_data_size += k_size_row * s_cell_count; + + // values (transposed) + const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); + s_cell_data_size += v_size_el * s_cell_count * n_embd_v_gqa; + } + + const size_t s_total = ( + size_t_size_size + + s_cell_count_size + + s_layer_count_size + + n_embd_v_gqa_size + + s_cell_data_size + ); + + return s_total; +} + +static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_context & data_ctx, llama_seq_id seq_id) { + llama_synchronize(ctx); + + const auto & kv_self = ctx->kv_self; + GGML_ASSERT(!kv_self.recurrent); // not implemented + + // Save the size of size_t as a uint32_t for safety check + const uint32_t size_t_size = sizeof(size_t); + data_ctx.write(&size_t_size, sizeof(size_t_size)); + + std::vector> cell_ranges; // ranges, from inclusive, to exclusive + uint32_t cell_count = 0; + + // Count the number of cells with the specified seq_id + // Find all the ranges of cells with this seq id + { + uint32_t cell_range_begin = kv_self.size; + for (uint32_t i = 0; i < kv_self.size; ++i) { + const auto & cell = kv_self.cells[i]; + if (cell.has_seq_id(seq_id)) { + ++cell_count; + if (cell_range_begin == kv_self.size) { + cell_range_begin = i; + } + } + else { + if (cell_range_begin != kv_self.size) { + cell_ranges.push_back({ cell_range_begin, i }); + cell_range_begin = kv_self.size; + } + } + } + if (cell_range_begin != kv_self.size) { + cell_ranges.push_back({ cell_range_begin, kv_self.size }); + } + + // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count + uint32_t cell_count_check = 0; + for (const auto & range : cell_ranges) { + cell_count_check += range.second - range.first; + } + GGML_ASSERT(cell_count == cell_count_check); + } + + // Write the cell count + data_ctx.write(&cell_count, sizeof(cell_count)); + + const auto & hparams = ctx->model.hparams; + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); + + // Write the layer count + data_ctx.write(&n_layer, sizeof(n_layer)); + + // Write n_embd_v_gqa + data_ctx.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + + // Iterate the ranges and write all the pos (this is the token position in the prompt) + for (const auto & range : cell_ranges) { + for (uint32_t i = range.first; i < range.second; ++i) { + const auto & cell = kv_self.cells[i]; + data_ctx.write(&cell.pos, sizeof(cell.pos)); + } + } + + // Iterate and write all the keys first, each row is a cell + // Get whole range at a time + std::vector tmp_buf; + for (int il = 0; il < (int)n_layer; ++il) { + // Write key type + const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; + data_ctx.write(&k_type_i, sizeof(k_type_i)); + + // Write row size of key + const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); + data_ctx.write(&k_size_row, sizeof(k_size_row)); + + // Read each range of cells of k_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + tmp_buf.resize(range_size * k_size_row); + ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), range.first * k_size_row, range_size * k_size_row); + data_ctx.write(tmp_buf.data(), tmp_buf.size()); + } + } + + // TODO: simplify, reduce copy-paste + if (!kv_self.v_trans) { + for (int il = 0; il < (int)n_layer; ++il) { + // Write value type + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + data_ctx.write(&v_type_i, sizeof(v_type_i)); + + // Write row size of value + const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); + data_ctx.write(&v_size_row, sizeof(v_size_row)); + + // Read each range of cells of v_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + tmp_buf.resize(range_size * v_size_row); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row); + data_ctx.write(tmp_buf.data(), tmp_buf.size()); + } + } + } else { + // For the values, they are transposed, so we also need the element size and get the element ranges from each row + const uint32_t kv_size = kv_self.size; + for (int il = 0; il < (int)n_layer; ++il) { + // Write value type + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + data_ctx.write(&v_type_i, sizeof(v_type_i)); + + // Write element size + const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); + data_ctx.write(&v_size_el, sizeof(v_size_el)); + + // For each row, we get the element values of each cell + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + // Read each range of cells of v_size_el length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t src_offset = (range.first + j * kv_size) * v_size_el; + tmp_buf.resize(range_size * v_size_el); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size()); + data_ctx.write(tmp_buf.data(), tmp_buf.size()); + } + } + } + } + + return data_ctx.get_size_written(); +} + +size_t llama_state_seq_get_data(struct llama_context* ctx, uint8_t* dst, llama_seq_id seq_id) { + llama_data_buffer_context data_ctx(dst); + return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); +} + +size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, llama_seq_id dest_seq_id) { + llama_synchronize(ctx); + + auto & kv_self = ctx->kv_self; + GGML_ASSERT(!kv_self.recurrent); // not implemented + + // Wipe the slot + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + + const uint8_t * inp = src; + + // Read size of size_t + uint32_t size_t_size; + memcpy(&size_t_size, inp, sizeof(size_t_size)); + inp += sizeof(size_t_size); + if (size_t_size != sizeof(size_t)) { + LLAMA_LOG_ERROR("%s: size_t size mismatch\n", __func__); + return 0; + } + + // Read the cell count + uint32_t cell_count; + memcpy(&cell_count, inp, sizeof(cell_count)); + inp += sizeof(cell_count); + + // Read the layer count + uint32_t n_layer_ref; + memcpy(&n_layer_ref, inp, sizeof(n_layer_ref)); + inp += sizeof(n_layer_ref); + + // Read n_embd_v_gqa + uint32_t n_embd_v_gqa_ref; + memcpy(&n_embd_v_gqa_ref, inp, sizeof(n_embd_v_gqa_ref)); + inp += sizeof(n_embd_v_gqa_ref); + + // Sanity check model compatibility + const auto & hparams = ctx->model.hparams; + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); + if (n_layer != n_layer_ref) { + LLAMA_LOG_ERROR("%s: mismatched n_layer (%d != %d)\n", __func__, n_layer, n_layer_ref); + return 0; + } + if (n_embd_v_gqa != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref); + return 0; + } + + // Allocate the new cells for the slot + if (cell_count) { + llama_batch batch = llama_batch_init(cell_count, 0, 1); + batch.n_tokens = cell_count; + for (uint32_t i = 0; i < cell_count; ++i) { + llama_pos pos; + memcpy(&pos, inp, sizeof(pos)); + inp += sizeof(pos); + + batch.pos[i] = pos; + batch.n_seq_id[i] = 1; + batch.seq_id[i][0] = dest_seq_id; + } + if (!llama_kv_cache_find_slot(kv_self, batch)) { + llama_batch_free(batch); + LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); + return 0; + } + + // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values) + // Assume that this is one contiguous block of cells + GGML_ASSERT(kv_self.head + cell_count <= kv_self.size); + GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]); + GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id)); + GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id)); + + // Cleanup + llama_batch_free(batch); + } + + const uint32_t kv_size = kv_self.size; + const uint32_t kv_head = kv_self.head; + + // For each layer, read the keys for each cell, one row is one cell, read as one contiguous blo + for (int il = 0; il < (int)n_layer; ++il) { + // Read type of key + int32_t k_type_i_ref; + memcpy(&k_type_i_ref, inp, sizeof(k_type_i_ref)); + inp += sizeof(k_type_i_ref); + const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; + if (k_type_i != k_type_i_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); + return 0; + } + + // Read row size of key + size_t k_size_row_ref; + memcpy(&k_size_row_ref, inp, sizeof(k_size_row_ref)); + inp += sizeof(k_size_row_ref); + const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); + if (k_size_row != k_size_row_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, k_size_row_ref, il); + return 0; + } + + if (cell_count) { + // Read and set the keys for the whole cell range + ggml_backend_tensor_set(kv_self.k_l[il], inp, kv_head * k_size_row, cell_count * k_size_row); + inp += cell_count * k_size_row; + } + } + + // TODO: simplify, reduce copy-paste + if (!kv_self.v_trans) { + for (int il = 0; il < (int)n_layer; ++il) { + // Read type of value + int32_t v_type_i_ref; + memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); + inp += sizeof(v_type_i_ref); + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + if (v_type_i != v_type_i_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return 0; + } + + // Read row size of value + size_t v_size_row_ref; + memcpy(&v_size_row_ref, inp, sizeof(v_size_row_ref)); + inp += sizeof(v_size_row_ref); + const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); + if (v_size_row != v_size_row_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, v_size_row_ref, il); + return 0; + } + + if (cell_count) { + // Read and set the values for the whole cell range + ggml_backend_tensor_set(kv_self.v_l[il], inp, kv_head * v_size_row, cell_count * v_size_row); + inp += cell_count * v_size_row; + } + } + } else { + // For each layer, read the values for each cell (transposed) + for (int il = 0; il < (int)n_layer; ++il) { + // Read type of value + int32_t v_type_i_ref; + memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); + inp += sizeof(v_type_i_ref); + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + if (v_type_i != v_type_i_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return 0; + } + + // Read element size of value + size_t v_size_el_ref; + memcpy(&v_size_el_ref, inp, sizeof(v_size_el_ref)); + inp += sizeof(v_size_el_ref); + const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); + if (v_size_el != v_size_el_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, v_size_el_ref, il); + return 0; + } + + if (cell_count) { + // For each row in the transposed matrix, read the values for the whole cell range + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + const size_t dst_offset = (kv_head + j * kv_size) * v_size_el; + ggml_backend_tensor_set(kv_self.v_l[il], inp, dst_offset, cell_count * v_size_el); + inp += cell_count * v_size_el; + } + } + } + } + + const size_t nread = inp - src; + + return nread; +} + +static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { + llama_file file(filepath, "wb"); + + file.write_u32(LLAMA_STATE_SEQ_MAGIC); + file.write_u32(LLAMA_STATE_SEQ_VERSION); + + // save the prompt + file.write_u32((uint32_t)n_token_count); + file.write_raw(tokens, sizeof(llama_token) * n_token_count); + + // save the context state using stream saving + llama_data_file_context data_ctx(&file); + llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); + + const size_t res = file.tell(); + GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + data_ctx.get_size_written()); + return res; +} + +static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + llama_file file(filepath, "rb"); + + // version checks + { + const uint32_t magic = file.read_u32(); + const uint32_t version = file.read_u32(); + + if (magic != LLAMA_STATE_SEQ_MAGIC || version != LLAMA_STATE_SEQ_VERSION) { + LLAMA_LOG_ERROR("%s: unknown (magic, version) for sequence state file: %08x, %08x\n", __func__, magic, version); + return 0; + } + } + + // load the prompt + { + const uint32_t n_token_count = file.read_u32(); + + if (n_token_count > n_token_capacity) { + LLAMA_LOG_ERROR("%s: token count in sequence state file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity); + return 0; + } + + file.read_raw(tokens_out, sizeof(llama_token) * n_token_count); + *n_token_count_out = n_token_count; + } + + // restore the context state + { + const size_t state_size = file.size - file.tell(); + std::vector state_data(state_size); + file.read_raw(state_data.data(), state_size); + const size_t nread = llama_state_seq_set_data(ctx, state_data.data(), dest_seq_id); + if (!nread) { + LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__); + return 0; + } + GGML_ASSERT(nread <= state_size); + GGML_ASSERT(nread + sizeof(uint32_t) * 3 + sizeof(llama_token) * *n_token_count_out == file.tell()); + } + + return file.tell(); +} + +size_t llama_state_seq_save_file(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { + try { + return llama_state_seq_save_file_internal(ctx, filepath, seq_id, tokens, n_token_count); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("error saving sequence state file: %s\n", err.what()); + return 0; + } +} + +size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + try { + return llama_state_seq_load_file_internal(ctx, filepath, dest_seq_id, tokens_out, n_token_capacity, n_token_count_out); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("error loading sequence state file: %s\n", err.what()); + return 0; + } +} + void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) { ctx->cparams.n_threads = n_threads; ctx->cparams.n_threads_batch = n_threads_batch; @@ -15511,23 +17386,31 @@ float * llama_get_logits(struct llama_context * ctx) { } float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { + int32_t j = -1; llama_synchronize(ctx); try { if (ctx->logits == nullptr) { throw std::runtime_error("no logits"); } - if ((size_t) i >= ctx->output_ids.size()) { + + if (i < 0) { + j = ctx->n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); + } + } else if ((size_t) i >= ctx->output_ids.size()) { throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size())); + } else { + j = ctx->output_ids[i]; } - const int32_t j = ctx->output_ids[i]; if (j < 0) { throw std::runtime_error(format("batch.logits[%d] != true", i)); } - if ((size_t) j >= ctx->output_size) { + if (j >= ctx->n_outputs) { // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, output_size=%lu)", j, ctx->output_size)); + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); } return ctx->logits + j*ctx->model.hparams.n_vocab; @@ -15547,23 +17430,32 @@ float * llama_get_embeddings(struct llama_context * ctx) { } float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) { + int32_t j = -1; + llama_synchronize(ctx); try { if (ctx->embd == nullptr) { throw std::runtime_error("no embeddings"); } - if ((size_t) i >= ctx->output_ids.size()) { + + if (i < 0) { + j = ctx->n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); + } + } else if ((size_t) i >= ctx->output_ids.size()) { throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size())); + } else { + j = ctx->output_ids[i]; } - const int32_t j = ctx->output_ids[i]; if (j < 0) { throw std::runtime_error(format("batch.logits[%d] != true", i)); } - if ((size_t) j >= ctx->output_size) { + if (j >= ctx->n_outputs) { // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, output_size=%lu)", j, ctx->output_size)); + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); } return ctx->embd + j*ctx->model.hparams.n_embd; @@ -15602,6 +17494,13 @@ llama_token_type llama_token_get_type(const struct llama_model * model, llama_to return model->vocab.id_to_token[token].type; } +bool llama_token_is_eog(const struct llama_model * model, llama_token token) { + return token != -1 && ( + token == llama_token_eos(model) || + token == llama_token_eot(model) + ); +} + llama_token llama_token_bos(const struct llama_model * model) { return model->vocab.special_bos_id; } @@ -15610,6 +17509,14 @@ llama_token llama_token_eos(const struct llama_model * model) { return model->vocab.special_eos_id; } +llama_token llama_token_cls(const struct llama_model * model) { + return model->vocab.special_cls_id; +} + +llama_token llama_token_sep(const struct llama_model * model) { + return model->vocab.special_sep_id; +} + llama_token llama_token_nl(const struct llama_model * model) { return model->vocab.linefeed_id; } @@ -15644,9 +17551,9 @@ int32_t llama_tokenize( int32_t text_len, llama_token * tokens, int32_t n_tokens_max, - bool add_bos, - bool special) { - auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos, special); + bool add_special, + bool parse_special) { + auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_special, parse_special); if (n_tokens_max < (int) res.size()) { // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); @@ -15662,16 +17569,17 @@ int32_t llama_tokenize( static std::string llama_decode_text(const std::string & text) { std::string decoded_text; - auto unicode_sequences = unicode_cpts_from_utf8(text); - for (auto & unicode_sequence : unicode_sequences) { - decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(unicode_sequence)); + + const auto cpts = unicode_cpts_from_utf8(text); + for (const auto cpt : cpts) { + decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(cpt)); } return decoded_text; } // does not write null-terminator to buf -int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length) { +int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) { if (0 <= token && token < llama_n_vocab(model)) { switch (llama_vocab_get_type(model->vocab)) { case LLAMA_VOCAB_TYPE_WPM: @@ -15686,7 +17594,9 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token } memcpy(buf, result.c_str(), result.length()); return result.length(); - } else if (llama_is_user_defined_token(model->vocab, token)) { + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { std::string result = model->vocab.id_to_token[token].text; if (length < (int) result.length()) { return -(int) result.length(); @@ -15699,8 +17609,6 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token } memcpy(buf, "\xe2\x96\x85", 3); return 3; - } else if (llama_is_control_token(model->vocab, token)) { - ; } else if (llama_is_byte_token(model->vocab, token)) { if (length < 1) { return -1; @@ -15721,15 +17629,15 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token } memcpy(buf, result.c_str(), result.length()); return result.length(); - } else if (llama_is_user_defined_token(model->vocab, token)) { + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { std::string result = model->vocab.id_to_token[token].text; if (length < (int) result.length()) { return -(int) result.length(); } memcpy(buf, result.c_str(), result.length()); return result.length(); - } else if (llama_is_control_token(model->vocab, token)) { - ; } break; } @@ -15912,6 +17820,39 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "### Response:\n"; } + } else if (tmpl == "command-r" || (tmpl.find("<|START_OF_TURN_TOKEN|>") != std::string::npos && tmpl.find("<|USER_TOKEN|>") != std::string::npos)) { + // CohereForAI/c4ai-command-r-plus + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>"; + } else if (role == "user") { + ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>"; + } else if (role == "assistant") { + ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>"; + } + } + if (add_ass) { + ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"; + } + } else if (tmpl == "llama3" || (tmpl.find("<|start_header_id|>") != std::string::npos && tmpl.find("<|end_header_id|>") != std::string::npos)) { + // Llama 3 + for (auto message : chat) { + std::string role(message->role); + ss << "<|start_header_id|>" << role << "<|end_header_id|>\n\n" << trim(message->content) << "<|eot_id|>"; + } + if (add_ass) { + ss << "<|start_header_id|>assistant<|end_header_id|>\n\n"; + } + } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos )) { + // Phi 3 + for (auto message : chat) { + std::string role(message->role); + ss << "<|" << role << "|>\n" << trim(message->content) << "<|end|>\n"; + } + if (add_ass) { + ss << "<|assistant|>\n"; + } } else { // template not supported return -1; @@ -15995,7 +17936,7 @@ struct llama_timings llama_get_timings(struct llama_context * ctx) { /*.t_eval_ms =*/ 1e-3 * ctx->t_eval_us, /*.n_sample =*/ std::max(1, ctx->n_sample), - /*.n_p_eval =*/ std::max(1, ctx->n_p_eval), + /*.n_p_eval =*/ std::max(0, ctx->n_p_eval), /*.n_eval =*/ std::max(1, ctx->n_eval), }; @@ -16044,6 +17985,11 @@ const char * llama_print_system_info(void) { s += "SSSE3 = " + std::to_string(ggml_cpu_has_ssse3()) + " | "; s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | "; +#ifdef GGML_USE_LLAMAFILE + s += "LLAMAFILE = 1 | "; +#else + s += "LLAMAFILE = 0 | "; +#endif return s.c_str(); }