llama_cpp 0.2.1 → 0.3.0

data/ext/llama_cpp/src/llama.cpp

@@ -19,6 +19,15 @@
 #ifdef GGML_USE_METAL
 #include "ggml-metal.h"
 #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>
@@ -40,6 +49,10 @@
 #include <sstream>
 #include <numeric>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
 
@@ -173,6 +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;
 
@@ -189,10 +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;
 
@@ -206,6 +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);
@@ -224,24 +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;
@@ -252,8 +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;
 
@@ -752,7 +774,7 @@ 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_loaders.at(0)->file, prefetch_size, ggml_is_numa()));
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
@@ -882,6 +904,7 @@ static bool kv_cache_init(
     const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+    cache.n = 0;
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size;
@@ -900,6 +923,7 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+    (void) n_gpu_layers;
 #ifdef GGML_USE_CUBLAS
     if (n_gpu_layers > n_layer + 1) {
         ggml_cuda_assign_buffers_no_scratch(cache.v);
@@ -914,21 +938,21 @@ static bool kv_cache_init(
 
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
+        /*.seed                        =*/ -1,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
         /*.low_vram                    =*/ false,
-        /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.embedding                   =*/ false,
-        /*.progress_callback           =*/ nullptr,
-        /*.progress_callback_user_data =*/ nullptr,
     };
 
     return result;
@@ -953,7 +977,7 @@ 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
@@ -962,6 +986,10 @@ void llama_init_backend() {
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
+
+    if (numa) {
+        ggml_numa_init();
+    }
 }
 
 int64_t llama_time_us() {
@@ -1022,7 +1050,8 @@ 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,
@@ -1036,12 +1065,11 @@ 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));
 
-    lctx.vocab = std::move(ml->file_loaders.at(0)->vocab);
-    auto & model = lctx.model;
+    vocab = std::move(ml->file_loaders.at(0)->vocab);
     model.hparams = ml->file_loaders.at(0)->hparams;
     model.n_gpu_layers = n_gpu_layers;
     llama_file_version file_version = ml->file_loaders.at(0)->file_version;
@@ -1111,15 +1139,15 @@ 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,
         };
 
@@ -1249,7 +1277,7 @@ static void llama_model_load_internal(
             vram_scratch = n_batch * MB;
             ggml_cuda_set_scratch_size(vram_scratch);
             if (n_gpu_layers > 0) {
-                fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+                fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",
                         __func__, vram_scratch / MB);
             }
         }
@@ -1300,7 +1328,7 @@ 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);
@@ -1310,12 +1338,13 @@ 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,
@@ -1329,7 +1358,7 @@ 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) {
@@ -1367,7 +1396,7 @@ 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);
 
@@ -1462,11 +1491,11 @@ 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");
 
@@ -1609,7 +1638,7 @@ static bool llama_eval_internal(
                     model.layers[il].w1,
                     cur);
             offload_func(cur);
-            ggml_set_name(cur, "result_w2");
+            ggml_set_name(cur, "result_w1");
 
             // SILU activation
             cur = ggml_silu(ctx0, cur);
@@ -1646,15 +1675,11 @@ static bool llama_eval_internal(
     {
         cur = ggml_rms_norm(ctx0, inpL);
         offload_func_nr(cur);
-        ggml_set_name(cur, "rms_norm_inpL");
-
-        cur = ggml_rms_norm(ctx0, cur);
-        offload_func_nr(cur);
-        ggml_set_name(cur, "rms_norm_after");
+        ggml_set_name(cur, "rms_norm_2");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
-        offload_func_nr(cur);
+        // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
         ggml_set_name(cur, "result_norm");
 
         embeddings = cur;
@@ -1719,7 +1744,7 @@ 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
     {
@@ -1998,9 +2023,10 @@ 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;
         }
     }
@@ -2452,6 +2478,10 @@ 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;
@@ -2485,21 +2515,37 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         } else {
             new_type = quantized_type;
 #ifdef GGML_USE_K_QUANTS
+            if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K ||
+                quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) {
+                int nx = tensor.ne.at(0);
+                int ny = tensor.ne.at(1);
+                if (nx % QK_K != 0 || ny % QK_K != 0) {
+                    fprintf(stderr, "\n\n========================= Tensor sizes %d x %d are not divisible by %d\n",nx,ny,QK_K);
+                    fprintf(stderr, "This is required to be able to use k-quants for now!\n");
+                    fprintf(stderr, "========================================================================================\n\n");
+                    throw std::runtime_error("Unsupported tensor size encountered\n");
+                }
+            }
             if (tensor.name == "output.weight") {
-               new_type = GGML_TYPE_Q6_K;
+                int nx = tensor.ne.at(0);
+                int ny = tensor.ne.at(1);
+                if (nx % QK_K == 0 && ny % QK_K == 0) {
+                    new_type = GGML_TYPE_Q6_K;
+                }
             } else if (tensor.name.find("attention.wv.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_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;
@@ -2612,12 +2658,39 @@ 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;
+}
+
+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 < 0) {
         params.seed = time(NULL);
@@ -2645,24 +2718,16 @@ 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);
         }
 
@@ -2690,16 +2755,21 @@ struct llama_context * llama_init_from_file(
         // this allocates all Metal resources and memory buffers
         ctx->ctx_metal = ggml_metal_init();
 
-        void *data_ptr = NULL;
+        void * data_ptr  = NULL;
         size_t data_size = 0;
+
         if (params.use_mmap) {
-            data_ptr = ctx->model.mapping->addr;
-            data_size= ctx->model.mapping->size;
+            data_ptr  = ctx->model.mapping->addr;
+            data_size = ctx->model.mapping->size;
         } else {
-            data_ptr = ggml_get_mem_buffer(ctx->model.ctx);
-            data_size= ggml_get_mem_size(ctx->model.ctx);
+            data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
+            data_size = ggml_get_mem_size  (ctx->model.ctx);
         }
 
+        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+
+        printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+
 #define LLAMA_METAL_CHECK_BUF(result)                                          \
     if (!(result)) {                                                           \
         fprintf(stderr, "%s: failed to add buffer\n", __func__);               \
@@ -2707,12 +2777,13 @@ struct llama_context * llama_init_from_file(
         return NULL;                                                           \
     }
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr,    ctx->buf_scratch[0].size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr,    ctx->buf_scratch[1].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->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));
 #undef LLAMA_METAL_CHECK_BUF
     }
 #endif
@@ -2720,7 +2791,23 @@ 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;
 }
 
@@ -2737,11 +2824,9 @@ 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);
@@ -2818,7 +2903,7 @@ 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_loaders.at(0)->file, /* prefetch */ 0, ggml_is_numa()));
         }
     }
 
@@ -2984,7 +3069,16 @@ 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;
@@ -2992,7 +3086,7 @@ 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)
@@ -3017,7 +3111,7 @@ 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
@@ -3083,7 +3177,7 @@ 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;
@@ -3098,9 +3192,7 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
         if (kv_size) {
             const size_t elt_size = ggml_element_size(kv_self.k);
 
-            char buffer[4096];
-
-            ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
             gf.n_threads = 1;
 
@@ -3189,7 +3281,7 @@ 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;
@@ -3206,9 +3298,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
             const size_t elt_size = ggml_element_size(kv_self.k);
 
-            char buffer[4096];
-
-            ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
             gf.n_threads = 1;
 
@@ -3235,7 +3325,7 @@ 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;
@@ -3443,9 +3533,12 @@ void llama_print_timings(struct llama_context * ctx) {
 
     fprintf(stderr, "\n");
     fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0);
-    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample);
-    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval);
-    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval);
+    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample, 1e6 / ctx->t_sample_us * n_sample);
+    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval, 1e6 / ctx->t_p_eval_us * n_p_eval);
+    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval,   1e6 / ctx->t_eval_us   * n_eval);
     fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0);
 }
 
@@ -3479,6 +3572,6 @@ 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;
 }