llama_cpp 0.2.2 → 0.3.1

data/ext/llama_cpp/src/llama.cpp

@@ -21,9 +21,13 @@
 #endif
 #ifdef GGML_USE_K_QUANTS
 #ifndef QK_K
+#ifdef GGML_QKK_64
+#define QK_K 64
+#else
 #define QK_K 256
 #endif
 #endif
+#endif
 
 #include <array>
 #include <ctime>
@@ -182,6 +186,19 @@ struct llama_kv_cache {
     }
 };
 
+struct llama_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    struct token_score {
+        token tok;
+        float score;
+    };
+
+    std::unordered_map<token, id> token_to_id;
+    std::vector<token_score> id_to_token;
+};
+
 struct llama_model {
     e_model type = MODEL_UNKNOWN;
 
@@ -198,10 +215,6 @@ struct llama_model {
     // context
     struct ggml_context * ctx = NULL;
 
-    // key + value cache for the self attention
-    // TODO: move to llama_state
-    struct llama_kv_cache kv_self;
-
     // the model memory buffer
     llama_ctx_buffer buf;
 
@@ -215,6 +228,11 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
+    int64_t t_load_us = 0;
+    int64_t t_start_us = 0;
+
+    llama_vocab vocab;
+
     ~llama_model() {
         if (ctx) {
             ggml_free(ctx);
@@ -233,24 +251,11 @@ struct llama_model {
     }
 };
 
-struct llama_vocab {
-    using id    = int32_t;
-    using token = std::string;
-
-    struct token_score {
-        token tok;
-        float score;
-    };
-
-    std::unordered_map<token, id> token_to_id;
-    std::vector<token_score> id_to_token;
-};
-
 struct llama_context {
+    llama_context(const llama_model & model, const llama_vocab & vocab) : model(model), vocab(vocab), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {}
+
     std::mt19937 rng;
 
-    int64_t t_load_us = 0;
-    int64_t t_start_us = 0;
     bool has_evaluated_once = false;
 
     int64_t t_sample_us = 0;
@@ -261,8 +266,16 @@ struct llama_context {
     int32_t n_eval   = 0; // number of eval calls
     int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
 
-    llama_model model;
-    llama_vocab vocab;
+    const llama_model & model;
+    const llama_vocab & vocab;
+
+    bool model_owner = false;
+
+    int64_t t_load_us;
+    int64_t t_start_us;
+
+    // key + value cache for the self attention
+    struct llama_kv_cache kv_self;
 
     size_t mem_per_token = 0;
 
@@ -351,96 +364,14 @@ static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml
     return size / ggml_blck_size(type);
 }
 
-struct llama_load_tensor_shard {
-    std::vector<uint32_t> ne;
-    size_t size;
-    enum ggml_type type;
-    size_t file_idx;
-    size_t file_off;
-
-    void calc_size() {
-        size = llama_calc_tensor_size(ne, type);
-    }
-};
-
-enum llama_split_type {
-    SPLIT_NONE,
-    SPLIT_BY_COLUMNS,
-    SPLIT_BY_ROWS
-};
-
 struct llama_load_tensor {
-    std::vector<llama_load_tensor_shard> shards;
-
     std::string name;
     enum ggml_type type = GGML_TYPE_F32;
-    llama_split_type split_type = SPLIT_NONE;
     std::vector<uint32_t> ne;
+    size_t file_off;
     size_t size;
     struct ggml_tensor * ggml_tensor = NULL;
     uint8_t * data;
-
-    llama_load_tensor(const std::string & name) : name(name) {}
-
-    void calc_all() {
-        calc_type();
-        calc_split_type();
-        calc_ne();
-        calc_size();
-    }
-
-    void calc_type() {
-        const auto & first_shard = shards.at(0);
-        for (const auto & shard : shards) {
-            if (shard.type != first_shard.type) {
-                throw std::runtime_error(format("inconsistent tensor shard type in '%s'", name.c_str()));
-            }
-        }
-        type = first_shard.type;
-    }
-
-    void calc_split_type() {
-        if (shards.at(0).ne.size() == 1 || // 1D tensors are just duplicated in every file
-            shards.size() == 1) { // only one file?
-            split_type = SPLIT_NONE;
-        } else if (name.find("tok_embeddings.") == 0 ||
-            name.find(".attention.wo.weight") != std::string::npos ||
-            name.find(".feed_forward.w2.weight") != std::string::npos) {
-            split_type = SPLIT_BY_COLUMNS;
-        } else {
-            split_type = SPLIT_BY_ROWS;
-        }
-    }
-
-    void calc_ne() {
-        const auto & first_shard = shards.at(0);
-        for (const auto & shard : shards) {
-            if (shard.ne != first_shard.ne) {
-                throw std::runtime_error(format("inconsistent tensor shard shape in '%s': first was %s, other was %s",
-                             name.c_str(), llama_format_tensor_shape(first_shard.ne).c_str(), llama_format_tensor_shape(shard.ne).c_str()));
-            }
-        }
-        ne = first_shard.ne;
-        LLAMA_ASSERT(shards.size() <= UINT32_MAX);
-        uint32_t n_shards = (uint32_t) shards.size();
-        switch (split_type) {
-            case SPLIT_NONE:
-                ne = first_shard.ne;
-                break;
-            case SPLIT_BY_COLUMNS:
-                ne = {checked_mul<uint32_t>(first_shard.ne[0], n_shards),
-                      first_shard.ne[1]};
-                break;
-            case SPLIT_BY_ROWS:
-                ne = {first_shard.ne[0],
-                      checked_mul<uint32_t>(first_shard.ne[1], n_shards)};
-                break;
-        }
-    }
-
-    void calc_size() {
-        size = llama_calc_tensor_size(ne, type);
-    }
 };
 
 struct llama_load_tensors_map {
@@ -463,13 +394,13 @@ struct llama_file_loader {
     llama_hparams hparams;
     llama_vocab vocab;
 
-    llama_file_loader(const char * fname, size_t file_idx, llama_load_tensors_map & tensors_map)
+    llama_file_loader(const char * fname, llama_load_tensors_map & tensors_map)
         : file(fname, "rb") {
         fprintf(stderr, "llama.cpp: loading model from %s\n", fname);
         read_magic();
         read_hparams();
         read_vocab();
-        read_tensor_metadata(file_idx, tensors_map);
+        read_tensor_metadata(tensors_map);
     }
     void read_magic() {
         uint32_t magic = file.read_u32();
@@ -526,19 +457,19 @@ struct llama_file_loader {
             tok_score.score = score;
         }
     }
-    void read_tensor_metadata(size_t file_idx, llama_load_tensors_map & tensors_map) {
+    void read_tensor_metadata(llama_load_tensors_map & tensors_map) {
         while (file.tell() < file.size) {
-            llama_load_tensor_shard shard;
+            llama_load_tensor tensor;
             uint32_t n_dims = file.read_u32();
             uint32_t name_len = file.read_u32();
-            shard.type = (enum ggml_type) file.read_u32();
-            shard.ne.resize(n_dims);
-            file.read_raw(shard.ne.data(), sizeof(shard.ne[0]) * n_dims);
+            tensor.type = (enum ggml_type) file.read_u32();
+            tensor.ne.resize(n_dims);
+            file.read_raw(tensor.ne.data(), sizeof(tensor.ne[0]) * n_dims);
             std::string name = file.read_string(name_len);
             if (n_dims < 1 || n_dims > 2) {
                 throw std::runtime_error(format("llama.cpp: tensor '%s' should not be %u-dimensional", name.c_str(), n_dims));
             }
-            switch (shard.type) {
+            switch (tensor.type) {
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q4_0:
@@ -553,30 +484,20 @@ struct llama_file_loader {
                 case GGML_TYPE_Q6_K:
                     break;
                 default: {
-                    throw std::runtime_error(format("unrecognized tensor type %u\n", shard.type));
+                    throw std::runtime_error(format("unrecognized tensor type %u\n", tensor.type));
                 }
             }
 
-            if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) {
-                // skip to the next multiple of 32 bytes
-                file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
-            }
-            shard.file_idx = file_idx;
-            shard.file_off = file.tell();
+            // skip to the next multiple of 32 bytes
+            file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
 
-            shard.calc_size();
-            file.seek(shard.size, SEEK_CUR);
+            tensor.file_off = file.tell();
+            tensor.name = name;
+            tensor.size = llama_calc_tensor_size(tensor.ne, tensor.type);
+            file.seek(tensor.size, SEEK_CUR);
 
-            auto it = tensors_map.name_to_idx.find(name);
-            size_t idx;
-            if (it != tensors_map.name_to_idx.end()) {
-                idx = it->second;
-            } else {
-                tensors_map.tensors.emplace_back(name);
-                idx = tensors_map.tensors.size() - 1;
-                tensors_map.name_to_idx.emplace(name, idx);
-            }
-            tensors_map.tensors.at(idx).shards.push_back(shard);
+            tensors_map.tensors.push_back(tensor);
+            tensors_map.name_to_idx[name] = tensors_map.tensors.size() - 1;
         }
     }
 };
@@ -646,56 +567,19 @@ struct llama_file_saver {
 };
 
 struct llama_model_loader {
-    std::vector<std::unique_ptr<llama_file_loader>> file_loaders;
+    std::unique_ptr<llama_file_loader> file_loader;
     llama_load_tensors_map tensors_map;
     bool use_mmap;
     size_t num_ggml_tensors_created = 0;
     struct ggml_context * ggml_ctx = NULL;
     std::unique_ptr<llama_mmap> mapping;
 
-    llama_model_loader(const std::string & fname_base, bool use_mmap, bool vocab_only) {
-        auto * first_file = new llama_file_loader(fname_base.c_str(), 0, tensors_map);
-        file_loaders.emplace_back(first_file);
-        uint32_t n_parts = vocab_only ? 1 : guess_n_parts();
-        for (uint32_t i = 1; i < n_parts; i++) {
-            std::string fname = fname_base + "." + std::to_string(i);
-            auto * ith_file = new llama_file_loader(fname.c_str(), i, tensors_map);
-            file_loaders.emplace_back(ith_file);
-            if (ith_file->hparams != first_file->hparams) {
-                throw std::runtime_error(format("llama.cpp: hparams inconsistent between files"));
-            }
-        }
+    llama_model_loader(const std::string & fname_base, bool use_mmap) {
+        file_loader = std::unique_ptr<llama_file_loader>(new llama_file_loader(fname_base.c_str(), tensors_map));
         if (!llama_mmap::SUPPORTED) {
             use_mmap = false;
         }
-        if (use_mmap && alignment_prevents_mmap()) {
-            fprintf(stderr, "llama.cpp: can't use mmap because tensors are not aligned; convert to new format to avoid this\n");
-            use_mmap = false;
-        }
         this->use_mmap = use_mmap;
-        for (llama_load_tensor & lt : tensors_map.tensors) {
-            lt.calc_all();
-        }
-    }
-
-    bool alignment_prevents_mmap() {
-        for (const llama_load_tensor & lt : tensors_map.tensors) {
-            for (const llama_load_tensor_shard & shard : lt.shards) {
-                if (shard.file_off & 3) {
-                    return true;
-                }
-            }
-        }
-        return false;
-    }
-
-    uint32_t guess_n_parts() const {
-        auto it = tensors_map.name_to_idx.find("tok_embeddings.weight");
-        if (it == tensors_map.name_to_idx.end()) {
-            throw std::runtime_error(std::string("missing tok_embeddings.weight"));
-        }
-        const llama_load_tensor & lt = tensors_map.tensors.at(it->second);
-        return file_loaders.at(0)->hparams.n_embd / lt.shards.at(0).ne.at(0);
     }
 
     void calc_sizes(size_t * ctx_size_p, size_t * mmapped_size_p) const {
@@ -761,7 +645,7 @@ struct llama_model_loader {
         }
 
         if (use_mmap) {
-            mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size));
+            mapping.reset(new llama_mmap(&file_loader->file, prefetch_size, ggml_is_numa()));
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
@@ -817,45 +701,13 @@ struct llama_model_loader {
 
     void load_data_for(llama_load_tensor & lt) {
         if (use_mmap) {
-            LLAMA_ASSERT(lt.shards.size() == 1);
-            lt.data = (uint8_t *) mapping->addr + lt.shards.at(0).file_off;
-        } else if (lt.split_type == SPLIT_NONE) {
-            llama_file & file = file_loaders.at(lt.shards.at(0).file_idx)->file;
-            file.seek(lt.shards.at(0).file_off, SEEK_SET);
+            lt.data = (uint8_t *) mapping->addr + lt.file_off;
+        } else {
+            llama_file & file = file_loader->file;
+            file.seek(lt.file_off, SEEK_SET);
             file.read_raw(lt.data, lt.size);
-        } else if (lt.split_type == SPLIT_BY_ROWS) {
-            size_t offset = 0;
-            for (llama_load_tensor_shard & shard : lt.shards) {
-                llama_file & file = file_loaders.at(shard.file_idx)->file;
-                file.seek(shard.file_off, SEEK_SET);
-                file.read_raw(lt.data + offset, shard.size);
-                offset += shard.size;
-            }
-            LLAMA_ASSERT(offset == lt.size);
-        } else if (lt.split_type == SPLIT_BY_COLUMNS) {
-            // Let's load the data into temporary buffers to ensure the OS performs large loads.
-            std::vector<llama_buffer> tmp_bufs(lt.shards.size());
-            for (size_t i = 0; i < lt.shards.size(); i++) {
-                llama_load_tensor_shard & shard = lt.shards.at(i);
-                llama_file & file = file_loaders.at(shard.file_idx)->file;
-                file.seek(shard.file_off, SEEK_SET);
-                tmp_bufs.at(i).resize(shard.size);
-                file.read_raw(tmp_bufs.at(i).addr, shard.size);
-            }
-            // Then reshape.
-            size_t num_rows = lt.ne.at(1);
-            size_t per_shard_row_size = lt.shards.at(0).size / num_rows;
-            size_t out_offset = 0;
-            for (size_t row = 0; row < num_rows; row++) {
-                for (llama_buffer & tmp_buf : tmp_bufs) {
-                    memcpy(lt.data + out_offset,
-                           tmp_buf.addr + row * per_shard_row_size,
-                           per_shard_row_size);
-                    out_offset += per_shard_row_size;
-                }
-            }
-            LLAMA_ASSERT(out_offset == lt.size);
         }
+
         if (0) {
             print_checksum(lt);
         }
@@ -925,7 +777,7 @@ static bool kv_cache_init(
 
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
-        /*.seed                        =*/ -1,
+        /*.seed                        =*/ LLAMA_DEFAULT_SEED,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
         /*.gpu_layers                  =*/ 0,
@@ -964,7 +816,7 @@ bool llama_mlock_supported() {
     return llama_mlock::SUPPORTED;
 }
 
-void llama_init_backend() {
+void llama_init_backend(bool numa) {
     ggml_time_init();
 
     // needed to initialize f16 tables
@@ -973,6 +825,10 @@ void llama_init_backend() {
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
+
+    if (numa) {
+        ggml_numa_init();
+    }
 }
 
 int64_t llama_time_us() {
@@ -1033,7 +889,8 @@ static const char *llama_model_type_name(e_model type) {
 
 static void llama_model_load_internal(
         const std::string & fname,
-        llama_context & lctx,
+        llama_model & model,
+        llama_vocab & vocab,
         int n_ctx,
         int n_batch,
         int n_gpu_layers,
@@ -1047,15 +904,14 @@ static void llama_model_load_internal(
         llama_progress_callback progress_callback,
         void * progress_callback_user_data) {
 
-    lctx.t_start_us = ggml_time_us();
+    model.t_start_us = ggml_time_us();
 
-    std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname, use_mmap, vocab_only));
+    std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname, use_mmap));
 
-    lctx.vocab = std::move(ml->file_loaders.at(0)->vocab);
-    auto & model = lctx.model;
-    model.hparams = ml->file_loaders.at(0)->hparams;
+    vocab = std::move(ml->file_loader->vocab);
+    model.hparams = ml->file_loader->hparams;
     model.n_gpu_layers = n_gpu_layers;
-    llama_file_version file_version = ml->file_loaders.at(0)->file_version;
+    llama_file_version file_version = ml->file_loader->file_version;
     auto & hparams = model.hparams;
 
     {
@@ -1089,7 +945,6 @@ static void llama_model_load_internal(
         fprintf(stderr, "%s: n_rot      = %u\n",  __func__, hparams.n_rot);
         fprintf(stderr, "%s: ftype      = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
         fprintf(stderr, "%s: n_ff       = %u\n",  __func__, n_ff);
-        fprintf(stderr, "%s: n_parts    = %zu\n", __func__, ml->file_loaders.size());
         fprintf(stderr, "%s: model size = %s\n",  __func__, llama_model_type_name(model.type));
     }
 
@@ -1122,15 +977,15 @@ static void llama_model_load_internal(
 
     // create the ggml context
     {
-        lctx.model.buf.resize(ctx_size);
+        model.buf.resize(ctx_size);
         if (use_mlock) {
-            lctx.model.mlock_buf.init(lctx.model.buf.addr);
-            lctx.model.mlock_buf.grow_to(lctx.model.buf.size);
+            model.mlock_buf.init(model.buf.addr);
+            model.mlock_buf.grow_to(model.buf.size);
         }
 
         struct ggml_init_params params = {
-            /*.mem_size   =*/ lctx.model.buf.size,
-            /*.mem_buffer =*/ lctx.model.buf.addr,
+            /*.mem_size   =*/ model.buf.size,
+            /*.mem_buffer =*/ model.buf.addr,
             /*.no_alloc   =*/ ml->use_mmap,
         };
 
@@ -1311,7 +1166,7 @@ static void llama_model_load_internal(
     }
 #endif
 
-    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
+    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
 
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);
@@ -1321,12 +1176,13 @@ static void llama_model_load_internal(
 
     // loading time will be recalculate after the first eval, so
     // we take page faults deferred by mmap() into consideration
-    lctx.t_load_us = ggml_time_us() - lctx.t_start_us;
+    model.t_load_us = ggml_time_us() - model.t_start_us;
 }
 
 static bool llama_model_load(
         const std::string & fname,
-        llama_context & lctx,
+        llama_model & model,
+        llama_vocab & vocab,
         int n_ctx,
         int n_batch,
         int n_gpu_layers,
@@ -1340,7 +1196,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
+        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1351,22 +1207,26 @@ static bool llama_model_load(
 
 // evaluate the transformer
 //
-//   - lctx:         llama context
-//   - tokens:       new batch of tokens to process
-//   - n_past:       the context size so far
-//   - n_threads:    number of threads to use
-//   - cgraph_fname: filename of the exported computation graph
+//   - lctx:      llama context
+//   - tokens:    new batch of tokens to process
+//   - embd       embeddings input
+//   - n_tokens   number of tokens
+//   - n_past:    the context size so far
+//   - n_threads: number of threads to use
 //
 static bool llama_eval_internal(
-        llama_context &  lctx,
-    const llama_token *  tokens,
-            const int    n_tokens,
-            const int    n_past,
-            const int    n_threads,
+         llama_context & lctx,
+     const llama_token * tokens,
+           const float * embd,
+             const int   n_tokens,
+             const int   n_past,
+             const int   n_threads,
             const char * cgraph_fname) {
 
+    LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
+
     // enforce that the first token is BOS
-    if (n_past == 0 && tokens[0] != llama_token_bos()) {
+    if (tokens && n_past == 0 && tokens[0] != llama_token_bos()) {
         fprintf(stderr, "%s: first token must be BOS\n", __func__);
         return false;
     }
@@ -1378,7 +1238,7 @@ static bool llama_eval_internal(
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
 
-    const auto & kv_self = model.kv_self;
+    const auto & kv_self = lctx.kv_self;
 
     LLAMA_ASSERT(!!kv_self.ctx);
 
@@ -1406,12 +1266,18 @@ static bool llama_eval_internal(
     ggml_cgraph gf = {};
     gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
 
-    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    ggml_set_name(embd, "embd");
-    memcpy(embd->data, tokens, N*ggml_element_size(embd));
-
     struct ggml_tensor * cur;
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+    struct ggml_tensor * inpL;
+
+    if (tokens) {
+        struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+        ggml_set_name(embd, "embd");
+        memcpy(embd->data, tokens, N*ggml_element_size(embd));
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+    } else {
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N);
+        memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
+    }
 
     const int i_gpu_start = n_layer - n_gpu_layers;
     (void) i_gpu_start;
@@ -1473,11 +1339,11 @@ static bool llama_eval_internal(
             offload_func_kq(tmpq);
             ggml_set_name(tmpq, "tmpq");
 
-            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
             offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
-            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
             offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
@@ -1726,7 +1592,7 @@ static bool llama_eval_internal(
     //memcpy(embd_w.data(), ggml_get_data(cur), sizeof(float)*n_vocab*N);
 
     // update kv token count
-    lctx.model.kv_self.n = n_past + N;
+    lctx.kv_self.n = n_past + N;
 
     // extract logits
     {
@@ -2005,9 +1871,10 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can
     for (size_t i = 0; i < candidates->size; ++i) {
         cum_sum += candidates->data[i].p;
 
-        // Check if the running sum is greater than p or if we have kept at least min_keep tokens
-        if (cum_sum > p && i >= min_keep) {
-            last_idx = i;
+        // Check if the running sum is at least p or if we have kept at least min_keep tokens
+        // we set the last index to i+1 to indicate that the current iterate should be included in the set
+        if (cum_sum >= p && i + 1 >= min_keep) {
+            last_idx = i + 1;
             break;
         }
     }
@@ -2432,9 +2299,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         nthread = std::thread::hardware_concurrency();
     }
 
-    std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false,
-                                                                            /*vocab_only*/ false));
-    llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), params->ftype);
+    std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false));
+    llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loader.get(), params->ftype);
 
 #ifdef GGML_USE_K_QUANTS
     int n_attention_wv    = 0;
@@ -2459,6 +2325,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     std::vector<std::thread> workers;
     std::mutex mutex;
 
+    auto use_more_bits = [] (int i_layer, int num_layers) -> bool {
+        return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
+    };
+
     size_t idx = 0;
     for (llama_load_tensor & tensor : model_loader->tensors_map.tensors) {
         llama_buffer read_data;
@@ -2513,15 +2383,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                         (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8 ||
-                         (i_attention_wv - n_attention_wv/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
+                        use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
+                else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
+                        (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
                 ++i_attention_wv;
             } else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                         (i_feed_forward_w2 < n_feed_forward_w2/8 || i_feed_forward_w2 >= 7*n_feed_forward_w2/8 ||
-                         (i_feed_forward_w2 - n_feed_forward_w2/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
+                         use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+                //else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < n_feed_forward_w2/8) new_type = GGML_TYPE_Q6_K;
                 ++i_feed_forward_w2;
             } else if (tensor.name.find("attention.wo.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
@@ -2630,18 +2501,47 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     }
 }
 
+
+
 //
 // interface implementation
 //
 
-struct llama_context * llama_init_from_file(
+struct llama_model * llama_load_model_from_file(
                              const char * path_model,
             struct llama_context_params   params) {
     ggml_time_init();
 
-    llama_context * ctx = new llama_context;
+    llama_model * model = new llama_model;
+
+    ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
+
+    if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers,
+                params.main_gpu, params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
+                params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
+        delete model;
+        fprintf(stderr, "%s: failed to load model\n", __func__);
+        return nullptr;
+    }
+
+    return model;
+}
 
-    if (params.seed < 0) {
+void llama_free_model(struct llama_model * model) {
+    delete model;
+}
+
+struct llama_context * llama_new_context_with_model(
+                             struct llama_model * model,
+            struct llama_context_params   params) {
+
+    if (!model) {
+        return nullptr;
+    }
+
+    llama_context * ctx = new llama_context(*model, model->vocab);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
@@ -2667,24 +2567,16 @@ struct llama_context * llama_init_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers, params.main_gpu,
-                params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
-                params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
-        fprintf(stderr, "%s: failed to load model\n", __func__);
-        llama_free(ctx);
-        return nullptr;
-    }
-
     // reserve memory for context buffers
     if (!params.vocab_only) {
-        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
+        if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
         }
 
         {
-            const size_t memory_size = ggml_nbytes(ctx->model.kv_self.k) + ggml_nbytes(ctx->model.kv_self.v);
+            const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v);
             fprintf(stderr, "%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
         }
 
@@ -2736,8 +2628,8 @@ struct llama_context * llama_init_from_file(
 
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr,       ctx->buf_compute.size,       0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.addr, ctx->kv_self.buf.size, 0));
 
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size, 0));
         LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
@@ -2748,7 +2640,23 @@ struct llama_context * llama_init_from_file(
     return ctx;
 }
 
+struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params) {
+
+    struct llama_model * model = llama_load_model_from_file(path_model, params);
+    if (!model) {
+        return nullptr;
+    }
+    struct llama_context * ctx = llama_new_context_with_model(model, params);
+    ctx->model_owner = true;
+    return ctx;
+}
+
 void llama_free(struct llama_context * ctx) {
+    if (ctx->model_owner) {
+        delete &ctx->model;
+    }
     delete ctx;
 }
 
@@ -2765,11 +2673,9 @@ int llama_model_quantize(
     }
 }
 
-int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) {
     fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
 
-    auto & model = ctx->model;
-
     const int64_t t_start_lora_us = ggml_time_us();
 
     auto fin = std::ifstream(path_lora, std::ios::binary);
@@ -2817,7 +2723,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
     // create a name -> tensor map of the model to accelerate lookups
     std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
-    for (auto & kv: model.tensors_by_name) {
+    for (const auto & kv: model.tensors_by_name) {
         model_tensors.insert(kv);
     }
 
@@ -2828,7 +2734,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
     llama_buffer base_buf;
     if (path_base_model) {
         fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model);
-        model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*vocab_only*/ false));
+        model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
 
         size_t ctx_size;
         size_t mmapped_size;
@@ -2846,7 +2752,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
         // maybe this should in llama_model_loader
         if (model_loader->use_mmap) {
-            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ 0));
+            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loader->file, /* prefetch */ 0, ggml_is_numa()));
         }
     }
 
@@ -2907,7 +2813,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
                         return false;
                     }
         }
-        ggml_tensor* lora_tensor;
+        ggml_tensor * lora_tensor;
         if (n_dims == 2) {
             lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
         }
@@ -2915,6 +2821,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
             fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims);
             return 1;
         }
+        ggml_set_name(lora_tensor, "lora_tensor");
 
         // load tensor data
         size_t offset = fin.tellg();
@@ -2930,6 +2837,21 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
             lora_tensors.find(base_name + ".loraB") != lora_tensors.end()) {
 
             ggml_tensor * dest_t = model_tensors[base_name];
+
+            offload_func_t offload_func = llama_nop;
+            offload_func_t offload_func_force_inplace = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+            if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) {
+                if (dest_t->type != GGML_TYPE_F16) {
+                    throw std::runtime_error(format(
+                        "%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models", __func__));
+                }
+                offload_func = ggml_cuda_assign_buffers;
+                offload_func_force_inplace = ggml_cuda_assign_buffers_force_inplace;
+            }
+#endif // GGML_USE_CUBLAS
+
             ggml_tensor * base_t;
             if (model_loader) {
                 // load from base model
@@ -2957,7 +2879,12 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
             }
 
             ggml_tensor * loraA = lora_tensors[base_name + ".loraA"];
+            GGML_ASSERT(loraA->type == GGML_TYPE_F32);
+            ggml_set_name(loraA, "loraA");
+
             ggml_tensor * loraB = lora_tensors[base_name + ".loraB"];
+            GGML_ASSERT(loraB->type == GGML_TYPE_F32);
+            ggml_set_name(loraB, "loraB");
 
             if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
                 fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
@@ -2967,19 +2894,32 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
             // w = w + BA*s
             ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+            offload_func(BA);
+            ggml_set_name(BA, "BA");
 
             if (scaling != 1.0f) {
                 ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
+                ggml_set_name(scale_tensor, "scale_tensor");
+
                 BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
+                offload_func(BA);
+                ggml_set_name(BA, "BA_scaled");
             }
 
             ggml_tensor * r;
             if (base_t == dest_t) {
                 r = ggml_add_inplace(lora_ctx, dest_t, BA);
+                offload_func_force_inplace(r);
+                ggml_set_name(r, "r_add_inplace");
             }
             else {
                 r = ggml_add(lora_ctx, base_t, BA);
+                offload_func(r);
+                ggml_set_name(r, "r_add");
+
                 r = ggml_cpy(lora_ctx, r, dest_t);
+                offload_func(r);
+                ggml_set_name(r, "r_cpy");
             }
 
             struct ggml_cgraph gf = ggml_build_forward(r);
@@ -3012,7 +2952,16 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
 int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
     try {
-        return llama_apply_lora_from_file_internal(ctx, path_lora, path_base_model, n_threads);
+        return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads);
+    } catch (const std::exception & err) {
+        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what());
+        return 1;
+    }
+}
+
+int llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, const char * path_base_model, int n_threads) {
+    try {
+        return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads);
     } catch (const std::exception & err) {
         fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
@@ -3020,13 +2969,13 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor
 }
 
 int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
-    return ctx->model.kv_self.n;
+    return ctx->kv_self.n;
 }
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
-void llama_set_rng_seed(struct llama_context * ctx, int seed) {
-    if (seed < 0) {
+void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
+    if (seed == LLAMA_DEFAULT_SEED) {
         seed = time(NULL);
     }
     ctx->rng.seed(seed);
@@ -3045,7 +2994,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     const size_t s_embedding       = ctx->embedding.size() * sizeof(float);
     const size_t s_kv_size         = sizeof(size_t);
     const size_t s_kv_ntok         = sizeof(int);
-    const size_t s_kv              = ctx->model.kv_self.buf.size;
+    const size_t s_kv              = ctx->kv_self.buf.size;
 
     const size_t s_total = (
         + s_rng_size
@@ -3111,7 +3060,7 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
 
     // copy kv cache
     {
-        const auto & kv_self = ctx->model.kv_self;
+        const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
         const int    n_layer = hparams.n_layer;
         const int    n_embd  = hparams.n_embd;
@@ -3215,7 +3164,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
     // set kv cache
     {
-        const auto & kv_self = ctx->model.kv_self;
+        const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
         const int    n_layer = hparams.n_layer;
         const int    n_embd  = hparams.n_embd;
@@ -3259,7 +3208,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
             ggml_free(cpy_ctx);
         }
 
-        ctx->model.kv_self.n = kv_ntok;
+        ctx->kv_self.n = kv_ntok;
     }
 
     const size_t nread    = inp - src;
@@ -3355,7 +3304,29 @@ int llama_eval(
                          int   n_tokens,
                          int   n_past,
                          int   n_threads) {
-    if (!llama_eval_internal(*ctx, tokens, n_tokens, n_past, n_threads, nullptr)) {
+    if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) {
+        fprintf(stderr, "%s: failed to eval\n", __func__);
+        return 1;
+    }
+
+    // get a more accurate load time, upon first eval
+    // TODO: fix this
+    if (!ctx->has_evaluated_once) {
+        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
+        ctx->has_evaluated_once = true;
+    }
+
+    return 0;
+}
+
+
+int llama_eval_embd(
+            struct llama_context * ctx,
+                     const float * embd,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads) {
+    if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) {
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
@@ -3376,7 +3347,7 @@ int llama_eval_export(struct llama_context * ctx, const char * fname) {
 
     const std::vector<llama_token> tmp(n_batch, llama_token_bos());
 
-    if (!llama_eval_internal(*ctx, tmp.data(), tmp.size(), n_ctx, 1, fname)) {
+    if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) {
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
@@ -3506,6 +3477,6 @@ const char * llama_print_system_info(void) {
 }
 
 // For internal test use
-std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
+const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
     return ctx->model.tensors_by_name;
 }