llama : control-vector -> adapter

include/llama.h

@@ -385,6 +385,7 @@ extern "C" {
     } llama_chat_message;
 
     // lora adapter
+    // TODO: rename to llama_adapter_lora
     struct llama_lora_adapter;
 
     // Helpers for getting default parameters
@@ -501,14 +502,20 @@ extern "C" {
             const char * fname_out,
             const llama_model_quantize_params * params);
 
+    //
+    // Adapters
+    //
+
     // Load a LoRA adapter from file
     // The loaded adapter will be associated to the given model, and will be free when the model is deleted
+    // TODO: rename to llama_adapter_lora_init
     LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init(
             struct llama_model * model,
             const char * path_lora);
 
     // Add a loaded LoRA adapter to given context
     // This will not modify model's weight
+    // TODO: rename to llama_set_adapter_lora
     LLAMA_API int32_t llama_lora_adapter_set(
             struct llama_context * ctx,
             struct llama_lora_adapter * adapter,
@@ -516,16 +523,19 @@ extern "C" {
 
     // Remove a specific LoRA adapter from given context
     // Return -1 if the adapter is not present in the context
+    // TODO: rename to llama_rm_adapter_lora
     LLAMA_API int32_t llama_lora_adapter_remove(
             struct llama_context * ctx,
             struct llama_lora_adapter * adapter);
 
     // Remove all LoRA adapters from given context
+    // TODO: rename to llama_clear_adapter_lora
     LLAMA_API void llama_lora_adapter_clear(
             struct llama_context * ctx);
 
     // Manually free a LoRA adapter
     // Note: loaded adapters will be free when the associated model is deleted
+    // TODO: rename to llama_adapter_lora_free
     LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter);
 
     // Apply a loaded control vector to a llama_context, or if data is NULL, clear
@@ -534,6 +544,7 @@ extern "C" {
     // to an n_embd x n_layers buffer starting from layer 1.
     // il_start and il_end are the layer range the vector should apply to (both inclusive)
     // See llama_control_vector_load in common to load a control vector.
+    // TODO: rename to llama_adapter_vec_apply
     LLAMA_API int32_t llama_control_vector_apply(
             struct llama_context * lctx,
                      const float * data,

src/CMakeLists.txt

@@ -12,7 +12,7 @@ add_library(llama
             llama-arch.cpp
             llama-batch.cpp
             llama-context.cpp
-            llama-control-vector.cpp
+            llama-adapter.cpp
             llama-grammar.cpp
             llama-kv-cache.cpp
             llama-mmap.cpp

src/llama-adapter.cpp

@@ -0,0 +1,5 @@
+#include "llama-adapter.h"
+
+void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
+    delete adapter;
+}

src/llama-adapter.h

@@ -0,0 +1,354 @@
+#pragma once
+
+#include "llama-impl.h"
+#include "ggml-cpp.h"
+
+#include "llama-model.h" // TODO: need only hparams
+
+#include <vector>
+#include <map>
+
+//
+// llama_adapter_vec
+//
+
+// TODO: rename to llama_adapter_vec
+struct llama_control_vector {
+    std::vector<struct ggml_tensor *> tensors; // per layer
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    int32_t layer_start = -1;
+    int32_t layer_end   = -1;
+
+    struct ggml_tensor * tensor_for(int il) const {
+        if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
+            return nullptr;
+        }
+        return tensors[il];
+    }
+
+    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+        ggml_tensor * layer_dir = tensor_for(il);
+        if (layer_dir != nullptr) {
+            cur = ggml_add(ctx, cur, layer_dir);
+        }
+        return cur;
+    }
+};
+
+static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
+    GGML_ASSERT(cvec.tensors.empty());
+    GGML_ASSERT(cvec.ctxs.empty());
+    GGML_ASSERT(cvec.bufs.empty());
+
+    // create a context for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ model.hparams.n_layer*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * ctx = ggml_init(params);
+            if (!ctx) {
+                return nullptr;
+            }
+            ctx_map[buft] = ctx;
+            cvec.ctxs.emplace_back(ctx);
+            return ctx;
+        }
+        return it->second;
+    };
+
+    // make tensors
+    cvec.tensors.reserve(model.hparams.n_layer);
+    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
+    for (size_t il = 1; il < model.hparams.n_layer; il++) {
+        ggml_backend_buffer_type_t buft = select_buft(*model.dev_layer.at(il).buft_list,
+            [&](ggml_context * ctx) {
+                ggml_tensor * cur = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
+                ggml_tensor * layer_dir = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
+                return ggml_add(ctx, cur, layer_dir);
+            });
+        ggml_context * ctx = ctx_for_buft(buft);
+        if (!ctx) {
+            LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
+            return false;
+        }
+        ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
+        cvec.tensors.push_back(tensor);
+    }
+
+    // allocate tensors / buffers and zero
+    cvec.bufs.reserve(ctx_map.size());
+    for (auto it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx = it.second;
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        if (!buf) {
+            LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__);
+            return false;
+        }
+        ggml_backend_buffer_clear(buf, 0);
+        cvec.bufs.emplace_back(buf);
+    }
+
+    return true;
+}
+
+static int32_t llama_control_vector_apply(struct llama_control_vector & cvec, const llama_model & model, const float * data, size_t len, int32_t n_embd, int32_t il_start, int32_t il_end) {
+    if (data == nullptr) {
+        // disable the current control vector (but leave allocated for later)
+        cvec.layer_start = -1;
+        cvec.layer_end   = -1;
+        return 0;
+    }
+
+    if (n_embd != (int) model.hparams.n_embd) {
+        LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
+        return 1;
+    }
+
+    if (cvec.tensors.empty()) {
+        if (!llama_control_vector_init(cvec, model)) {
+            return 1;
+        }
+    }
+
+    cvec.layer_start = il_start;
+    cvec.layer_end   = il_end;
+
+    for (size_t il = 1; il < model.hparams.n_layer; il++) {
+        assert(cvec.tensors[il] != nullptr);
+
+        const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
+        if (off + n_embd <= len) {
+            ggml_backend_tensor_set(cvec.tensors[il], data + off, 0, n_embd * ggml_element_size(cvec.tensors[il]));
+        }
+    }
+
+    return 0;
+}
+
+//
+// llama_adapter_lora
+//
+
+// TODO: rename to llama_adapter_lora_weight
+struct llama_lora_weight {
+    struct ggml_tensor * a = nullptr;
+    struct ggml_tensor * b = nullptr;
+    llama_lora_weight() = default;
+    llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b): a(a), b(b) {}
+};
+
+// TODO: rename to llama_adapter_lora
+struct llama_lora_adapter {
+    struct llama_model * base_model;
+    // map tensor name to lora_a_b
+    std::unordered_map<std::string, struct llama_lora_weight> ab_map;
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    float alpha;
+
+    llama_lora_adapter(struct llama_model * base_model): base_model(base_model) {
+        base_model->lora_adapters.insert(this);
+    }
+
+    llama_lora_weight * get_weight(struct ggml_tensor * w) {
+        std::string name(w->name);
+        auto pos = ab_map.find(name);
+        if (ab_map.find(name) != ab_map.end()) {
+            return &pos->second;
+        }
+        return nullptr;
+    }
+
+    ~llama_lora_adapter() {
+        auto pos = base_model->lora_adapters.find(this);
+        if (pos != base_model->lora_adapters.end()) {
+            base_model->lora_adapters.erase(pos);
+        }
+    }
+};
+
+static struct ggml_tensor * llama_get_model_tensor(const struct llama_model * model, const char * name) {
+    auto it = std::find_if(model->tensors_by_name.begin(), model->tensors_by_name.end(),
+            [name](const std::pair<std::string, struct ggml_tensor *> & it) {
+                return it.first == name;
+            });
+    if (it == model->tensors_by_name.end()) {
+        return nullptr;
+    }
+    return it->second;
+}
+
+static void llama_lora_adapter_init_internal(struct llama_model * model, const char * path_lora, struct llama_lora_adapter & adapter) {
+    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
+
+    ggml_context * ctx_init;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ true,
+        /* .ctx      = */ &ctx_init,
+    };
+
+    gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) };
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
+    }
+
+    ggml_context_ptr ctx { ctx_init };
+
+    // check metadata
+    {
+        auto get_kv_str = [&](const std::string & key) -> std::string {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
+        };
+        auto get_kv_f32 = [&](const std::string & key) -> float {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
+        };
+        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
+
+        auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE));
+        if (general_type != "adapter") {
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
+
+        auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
+        auto general_arch = llm_arch_from_string(general_arch_str);
+        if (general_arch != model->arch) {
+            throw std::runtime_error("model arch and LoRA arch mismatch");
+        }
+
+        auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE));
+        if (adapter_type != "lora") {
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
+        }
+
+        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+    }
+
+    int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
+
+    // contexts for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            // add a new context
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * buft_ctx = ggml_init(params);
+            if (!buft_ctx) {
+                return nullptr;
+            }
+            ctx_map[buft] = buft_ctx;
+            adapter.ctxs.emplace_back(buft_ctx);
+            return buft_ctx;
+        };
+        return it->second;
+    };
+
+    // bundle lora_a and lora_b into pairs
+    std::map<std::string, llama_lora_weight> ab_map;
+    auto str_endswith = [](const std::string & str, const std::string & suffix) {
+        return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+    };
+    for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) {
+        std::string name(cur->name);
+        if (str_endswith(name, ".lora_a")) {
+            replace_all(name, ".lora_a", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(cur, nullptr);
+            } else {
+                ab_map[name].a = cur;
+            }
+        } else if (str_endswith(name, ".lora_b")) {
+            replace_all(name, ".lora_b", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(nullptr, cur);
+            } else {
+                ab_map[name].b = cur;
+            }
+        } else {
+            throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
+        }
+    }
+
+    // add tensors
+    for (auto & it : ab_map) {
+        const std::string & name = it.first;
+        llama_lora_weight & w = it.second;
+
+        if (!w.a || !w.b) {
+            throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
+        }
+
+        // device buft and device ctx
+        auto * model_tensor = llama_get_model_tensor(model, name.c_str());
+        if (!model_tensor) {
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
+        }
+        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        // validate tensor shape
+        if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
+            throw std::runtime_error("tensor '" + name + "' has incorrect shape");
+        }
+        if (w.a->ne[1] != w.b->ne[0]) {
+            throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+        }
+        // save tensor to adapter
+        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_set_name(tensor_a, w.a->name);
+        ggml_set_name(tensor_b, w.b->name);
+        adapter.ab_map[name] = llama_lora_weight(tensor_a, tensor_b);
+    }
+
+    // allocate tensors / buffers and zero
+    {
+        adapter.ctxs.reserve(ctx_map.size());
+        adapter.bufs.reserve(ctx_map.size());
+        for (auto & it : ctx_map) {
+            ggml_backend_buffer_type_t buft = it.first;
+            ggml_context * ctx_dev = it.second;
+            ggml_backend_buffer_ptr buf { ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft) };
+            if (!buf) {
+                throw std::runtime_error("failed to allocate buffer for lora adapter\n");
+            }
+            LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0);
+            adapter.bufs.emplace_back(std::move(buf));
+        }
+    }
+
+    // set tensor data
+    {
+        llama_file gguf_file(path_lora, "rb");
+        std::vector<uint8_t> read_buf;
+        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+            size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
+            size_t size = ggml_nbytes(orig);
+            read_buf.resize(size);
+            gguf_file.seek(offs, SEEK_SET);
+            gguf_file.read_raw(read_buf.data(), size);
+            ggml_backend_tensor_set(dev, read_buf.data(), 0, size);
+        };
+        for (auto & it : adapter.ab_map) {
+            auto orig = ab_map[it.first];
+            auto dev  = it.second;
+            set_tensor(orig.a, dev.a);
+            set_tensor(orig.b, dev.b);
+        }
+    }
+
+    LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
+}

src/llama-context.h

@@ -4,7 +4,7 @@
 #include "llama-batch.h"
 #include "llama-model.h"
 #include "llama-kv-cache.h"
-#include "llama-control-vector.h"
+#include "llama-adapter.h"
 
 #include "ggml-cpp.h"
 
@@ -54,7 +54,7 @@ struct llama_context {
     const struct llama_model & model;
 
     struct llama_cparams        cparams;
-    struct llama_sbatch         sbatch;
+    struct llama_sbatch         sbatch;  // TODO: revisit if needed
     struct llama_kv_cache       kv_self;
     struct llama_control_vector cvec;
 

src/llama-control-vector.cpp

@@ -1 +0,0 @@
-#include "llama-control-vector.h"

src/llama-control-vector.h

@@ -1,130 +0,0 @@
-#pragma once
-
-#include "llama-impl.h"
-#include "ggml-cpp.h"
-
-#include "llama-model.h" // TODO: need only hparams
-
-#include <vector>
-#include <map>
-
-struct llama_control_vector {
-    std::vector<struct ggml_tensor *> tensors; // per layer
-    std::vector<ggml_context_ptr> ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
-
-    int32_t layer_start = -1;
-    int32_t layer_end   = -1;
-
-    struct ggml_tensor * tensor_for(int il) const {
-        if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
-            return nullptr;
-        }
-        return tensors[il];
-    }
-
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
-        ggml_tensor * layer_dir = tensor_for(il);
-        if (layer_dir != nullptr) {
-            cur = ggml_add(ctx, cur, layer_dir);
-        }
-        return cur;
-    }
-};
-
-static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
-    GGML_ASSERT(cvec.tensors.empty());
-    GGML_ASSERT(cvec.ctxs.empty());
-    GGML_ASSERT(cvec.bufs.empty());
-
-    // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
-    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
-        auto it = ctx_map.find(buft);
-        if (it == ctx_map.end()) {
-            struct ggml_init_params params = {
-                /*.mem_size   =*/ model.hparams.n_layer*ggml_tensor_overhead(),
-                /*.mem_buffer =*/ NULL,
-                /*.no_alloc   =*/ true,
-            };
-            ggml_context * ctx = ggml_init(params);
-            if (!ctx) {
-                return nullptr;
-            }
-            ctx_map[buft] = ctx;
-            cvec.ctxs.emplace_back(ctx);
-            return ctx;
-        }
-        return it->second;
-    };
-
-    // make tensors
-    cvec.tensors.reserve(model.hparams.n_layer);
-    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
-    for (size_t il = 1; il < model.hparams.n_layer; il++) {
-        ggml_backend_buffer_type_t buft = select_buft(*model.dev_layer.at(il).buft_list,
-            [&](ggml_context * ctx) {
-                ggml_tensor * cur = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
-                ggml_tensor * layer_dir = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
-                return ggml_add(ctx, cur, layer_dir);
-            });
-        ggml_context * ctx = ctx_for_buft(buft);
-        if (!ctx) {
-            LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
-            return false;
-        }
-        ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
-        cvec.tensors.push_back(tensor);
-    }
-
-    // allocate tensors / buffers and zero
-    cvec.bufs.reserve(ctx_map.size());
-    for (auto it : ctx_map) {
-        ggml_backend_buffer_type_t buft = it.first;
-        ggml_context * ctx = it.second;
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
-        if (!buf) {
-            LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__);
-            return false;
-        }
-        ggml_backend_buffer_clear(buf, 0);
-        cvec.bufs.emplace_back(buf);
-    }
-
-    return true;
-}
-
-static int32_t llama_control_vector_apply(struct llama_control_vector & cvec, const llama_model & model, const float * data, size_t len, int32_t n_embd, int32_t il_start, int32_t il_end) {
-    if (data == nullptr) {
-        // disable the current control vector (but leave allocated for later)
-        cvec.layer_start = -1;
-        cvec.layer_end   = -1;
-        return 0;
-    }
-
-    if (n_embd != (int) model.hparams.n_embd) {
-        LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
-        return 1;
-    }
-
-    if (cvec.tensors.empty()) {
-        if (!llama_control_vector_init(cvec, model)) {
-            return 1;
-        }
-    }
-
-    cvec.layer_start = il_start;
-    cvec.layer_end   = il_end;
-
-    for (size_t il = 1; il < model.hparams.n_layer; il++) {
-        assert(cvec.tensors[il] != nullptr);
-
-        const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
-        if (off + n_embd <= len) {
-            ggml_backend_tensor_set(cvec.tensors[il], data + off, 0, n_embd * ggml_element_size(cvec.tensors[il]));
-        }
-    }
-
-    return 0;
-}
-

src/llama.cpp

@@ -230,58 +230,10 @@ static const size_t kiB = 1024;
 static const size_t MiB = 1024*kiB;
 static const size_t GiB = 1024*MiB;
 
-struct llama_lora_weight {
-    struct ggml_tensor * a = nullptr;
-    struct ggml_tensor * b = nullptr;
-    llama_lora_weight() = default;
-    llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b): a(a), b(b) {}
-};
-
-struct llama_lora_adapter {
-    struct llama_model * base_model;
-    // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_lora_weight> ab_map;
-    std::vector<ggml_context_ptr> ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
-
-    float alpha;
-
-    llama_lora_adapter(struct llama_model * base_model): base_model(base_model) {
-        base_model->lora_adapters.insert(this);
-    }
-
-    llama_lora_weight * get_weight(struct ggml_tensor * w) {
-        std::string name(w->name);
-        auto pos = ab_map.find(name);
-        if (ab_map.find(name) != ab_map.end()) {
-            return &pos->second;
-        }
-        return nullptr;
-    }
-
-    ~llama_lora_adapter() {
-        auto pos = base_model->lora_adapters.find(this);
-        if (pos != base_model->lora_adapters.end()) {
-            base_model->lora_adapters.erase(pos);
-        }
-    }
-};
-
 static int llama_get_device_count(const llama_model & model) {
     return (int) model.devices.size();
 }
 
-static struct ggml_tensor * llama_get_model_tensor(const struct llama_model * model, const char * name) {
-    auto it = std::find_if(model->tensors_by_name.begin(), model->tensors_by_name.end(),
-            [name](const std::pair<std::string, struct ggml_tensor *> & it) {
-                return it.first == name;
-            });
-    if (it == model->tensors_by_name.end()) {
-        return nullptr;
-    }
-    return it->second;
-}
-
 //
 // model loading and saving
 //
@@ -15884,172 +15836,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     }
 }
 
-static void llama_lora_adapter_init_internal(struct llama_model * model, const char * path_lora, struct llama_lora_adapter & adapter) {
-    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
-
-    ggml_context * ctx_init;
-    struct gguf_init_params meta_gguf_params = {
-        /* .no_alloc = */ true,
-        /* .ctx      = */ &ctx_init,
-    };
-
-    gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) };
-    if (!ctx_gguf) {
-        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
-    }
-
-    ggml_context_ptr ctx { ctx_init };
-
-    // check metadata
-    {
-        auto get_kv_str = [&](const std::string & key) -> std::string {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
-        };
-        auto get_kv_f32 = [&](const std::string & key) -> float {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
-        };
-        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
-
-        auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE));
-        if (general_type != "adapter") {
-            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
-        }
-
-        auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
-        auto general_arch = llm_arch_from_string(general_arch_str);
-        if (general_arch != model->arch) {
-            throw std::runtime_error("model arch and LoRA arch mismatch");
-        }
-
-        auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE));
-        if (adapter_type != "lora") {
-            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
-        }
-
-        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
-    }
-
-    int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
-
-    // contexts for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
-    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
-        auto it = ctx_map.find(buft);
-        if (it == ctx_map.end()) {
-            // add a new context
-            struct ggml_init_params params = {
-                /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
-                /*.mem_buffer =*/ NULL,
-                /*.no_alloc   =*/ true,
-            };
-            ggml_context * buft_ctx = ggml_init(params);
-            if (!buft_ctx) {
-                return nullptr;
-            }
-            ctx_map[buft] = buft_ctx;
-            adapter.ctxs.emplace_back(buft_ctx);
-            return buft_ctx;
-        };
-        return it->second;
-    };
-
-    // bundle lora_a and lora_b into pairs
-    std::map<std::string, llama_lora_weight> ab_map;
-    auto str_endswith = [](const std::string & str, const std::string & suffix) {
-        return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-    };
-    for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) {
-        std::string name(cur->name);
-        if (str_endswith(name, ".lora_a")) {
-            replace_all(name, ".lora_a", "");
-            if (ab_map.find(name) == ab_map.end()) {
-                ab_map[name] = llama_lora_weight(cur, nullptr);
-            } else {
-                ab_map[name].a = cur;
-            }
-        } else if (str_endswith(name, ".lora_b")) {
-            replace_all(name, ".lora_b", "");
-            if (ab_map.find(name) == ab_map.end()) {
-                ab_map[name] = llama_lora_weight(nullptr, cur);
-            } else {
-                ab_map[name].b = cur;
-            }
-        } else {
-            throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
-        }
-    }
-
-    // add tensors
-    for (auto & it : ab_map) {
-        const std::string & name = it.first;
-        llama_lora_weight & w = it.second;
-
-        if (!w.a || !w.b) {
-            throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
-        }
-
-        // device buft and device ctx
-        auto * model_tensor = llama_get_model_tensor(model, name.c_str());
-        if (!model_tensor) {
-            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
-        }
-        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
-        // validate tensor shape
-        if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
-            throw std::runtime_error("tensor '" + name + "' has incorrect shape");
-        }
-        if (w.a->ne[1] != w.b->ne[0]) {
-            throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
-        }
-        // save tensor to adapter
-        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
-        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
-        ggml_set_name(tensor_a, w.a->name);
-        ggml_set_name(tensor_b, w.b->name);
-        adapter.ab_map[name] = llama_lora_weight(tensor_a, tensor_b);
-    }
-
-    // allocate tensors / buffers and zero
-    {
-        adapter.ctxs.reserve(ctx_map.size());
-        adapter.bufs.reserve(ctx_map.size());
-        for (auto & it : ctx_map) {
-            ggml_backend_buffer_type_t buft = it.first;
-            ggml_context * ctx_dev = it.second;
-            ggml_backend_buffer_ptr buf { ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft) };
-            if (!buf) {
-                throw std::runtime_error("failed to allocate buffer for lora adapter\n");
-            }
-            LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0);
-            adapter.bufs.emplace_back(std::move(buf));
-        }
-    }
-
-    // set tensor data
-    {
-        llama_file gguf_file(path_lora, "rb");
-        std::vector<uint8_t> read_buf;
-        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
-            size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
-            size_t size = ggml_nbytes(orig);
-            read_buf.resize(size);
-            gguf_file.seek(offs, SEEK_SET);
-            gguf_file.read_raw(read_buf.data(), size);
-            ggml_backend_tensor_set(dev, read_buf.data(), 0, size);
-        };
-        for (auto & it : adapter.ab_map) {
-            auto orig = ab_map[it.first];
-            auto dev  = it.second;
-            set_tensor(orig.a, dev.a);
-            set_tensor(orig.b, dev.b);
-        }
-    }
-
-    LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
-}
-
 int32_t llama_lora_adapter_set(
             struct llama_context * ctx,
             struct llama_lora_adapter * adapter,
@@ -16077,10 +15863,6 @@ void llama_lora_adapter_clear(struct llama_context * ctx) {
     ctx->lora_adapters.clear();
 }
 
-void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
-    delete adapter;
-}
-
 // TODO: tmp
 int32_t llama_control_vector_apply(
         struct llama_context * lctx,