diff --git a/examples/chat-persistent.sh b/examples/chat-persistent.sh index e0c251e5b..22f5b83d3 100755 --- a/examples/chat-persistent.sh +++ b/examples/chat-persistent.sh @@ -9,7 +9,7 @@ if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then exit 1 fi -MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}" +MODEL="${MODEL:-./models/llama-13b/ggml-model-q4_0.gguf}" PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}" USER_NAME="${USER_NAME:-User}" AI_NAME="${AI_NAME:-ChatLLaMa}" @@ -61,9 +61,9 @@ fi if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then echo 'Prompt cache does not exist, building...' - # Default batch_size to 8 here for better user feedback during initial prompt processing + # Default batch_size to 64 here for better user feedback during initial prompt processing ./main 2>>"$LOG" \ - --batch_size 8 \ + --batch_size 64 \ "${OPTS[@]}" \ --prompt-cache "$PROMPT_CACHE_FILE" \ --file "$CUR_PROMPT_FILE" \ @@ -132,7 +132,7 @@ while read -e line; do # HACK get num tokens from debug message # TODO get both messages in one go if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" || - ! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then + ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then echo >&2 "Couldn't get number of tokens from ./main output!" exit 1 fi diff --git a/llama.cpp b/llama.cpp index 666acc212..37c517028 100644 --- a/llama.cpp +++ b/llama.cpp @@ -7044,16 +7044,6 @@ struct llama_data_file_context : llama_data_context { * */ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { - // TODO: does not support multi-sequence states - { - const auto & kv_self = ctx->kv_self; - for (uint32_t i = 0; i < kv_self.head; ++i) { - GGML_ASSERT(kv_self.cells[i].pos == (int32_t) i); - GGML_ASSERT(kv_self.cells[i].seq_id.size() == 1); - GGML_ASSERT(kv_self.cells[i].has_seq_id(0)); - } - } - // copy rng { std::stringstream rng_ss; @@ -7106,36 +7096,38 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat const auto & hparams = ctx->model.hparams; const auto & cparams = ctx->cparams; - const int n_layer = hparams.n_layer; - const int n_embd = hparams.n_embd_gqa(); - const int n_ctx = cparams.n_ctx; + const auto n_layer = hparams.n_layer; + const auto n_embd = hparams.n_embd_gqa(); + const auto n_ctx = cparams.n_ctx; - const size_t kv_size = kv_self.buf.size; - const int kv_ntok = kv_self.head; + const size_t kv_buf_size = kv_self.buf.size; + const uint32_t kv_head = kv_self.head; + const uint32_t kv_size = kv_self.size; - data_ctx->write(&kv_size, sizeof(kv_size)); - data_ctx->write(&kv_ntok, sizeof(kv_ntok)); + 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)); - if (kv_size) { + if (kv_buf_size) { const size_t elt_size = ggml_element_size(kv_self.k); ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); ggml_cgraph gf{}; - ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); + ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); std::vector kout3d_data(ggml_nbytes(kout3d), 0); kout3d->data = kout3d_data.data(); - ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); + ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer); std::vector vout3d_data(ggml_nbytes(vout3d), 0); vout3d->data = vout3d_data.data(); ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_ntok, n_layer, + n_embd, kv_head, n_layer, elt_size*n_embd, elt_size*n_embd*n_ctx, 0); ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_ntok, n_embd, n_layer, + kv_head, n_embd, n_layer, elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d)); @@ -7149,6 +7141,20 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat data_ctx->write(kout3d_data.data(), kout3d_data.size()); data_ctx->write(vout3d_data.data(), vout3d_data.size()); } + + for (uint32_t i = 0; i < kv_size; ++i) { + const auto & cell = kv_self.cells[i]; + + const llama_pos pos = cell.pos; + const size_t seq_id_size = cell.seq_id.size(); + + data_ctx->write(&pos, sizeof(pos)); + data_ctx->write(&seq_id_size, sizeof(seq_id_size)); + + for (auto seq_id : cell.seq_id) { + data_ctx->write(&seq_id, sizeof(seq_id)); + } + } } } @@ -7220,34 +7226,36 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { const int n_embd = hparams.n_embd_gqa(); const int n_ctx = cparams.n_ctx; - size_t kv_size; - int kv_ntok; + size_t kv_buf_size; + uint32_t kv_head; + uint32_t kv_size; - memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size); - memcpy(&kv_ntok, inp, sizeof(kv_ntok)); inp += sizeof(kv_ntok); + 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); - if (kv_size) { - GGML_ASSERT(kv_self.buf.size == kv_size); + if (kv_buf_size) { + GGML_ASSERT(kv_self.buf.size == kv_buf_size); const size_t elt_size = ggml_element_size(kv_self.k); ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); ggml_cgraph gf{}; - ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); + ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); kin3d->data = (void *) inp; inp += ggml_nbytes(kin3d); - ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); + ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer); vin3d->data = (void *) inp; inp += ggml_nbytes(vin3d); ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, - n_embd, kv_ntok, n_layer, + n_embd, kv_head, n_layer, elt_size*n_embd, elt_size*n_embd*n_ctx, 0); ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, - kv_ntok, n_embd, n_layer, + kv_head, n_embd, n_layer, elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d)); @@ -7257,8 +7265,27 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { ggml_free(cpy_ctx); } - ctx->kv_self.head = kv_ntok; + ctx->kv_self.head = kv_head; ctx->kv_self.size = kv_size; + + ctx->kv_self.cells.resize(kv_size); + + for (uint32_t i = 0; i < kv_size; ++i) { + llama_pos pos; + size_t seq_id_size; + + memcpy(&pos, inp, sizeof(pos)); inp += sizeof(pos); + memcpy(&seq_id_size, inp, sizeof(seq_id_size)); inp += sizeof(seq_id_size); + + ctx->kv_self.cells[i].pos = pos; + + llama_seq_id seq_id; + + for (size_t j = 0; j < seq_id_size; ++j) { + memcpy(&seq_id, inp, sizeof(seq_id)); inp += sizeof(seq_id); + ctx->kv_self.cells[i].seq_id.insert(seq_id); + } + } } const size_t nread = inp - src; diff --git a/llama.h b/llama.h index 96ff1f09c..a5190a8a5 100644 --- a/llama.h +++ b/llama.h @@ -42,7 +42,7 @@ #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn' #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN -#define LLAMA_SESSION_VERSION 1 +#define LLAMA_SESSION_VERSION 2 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) // Defined when llama.cpp is compiled with support for offloading model layers to GPU.