talk-llama : sync llama.cpp

Author: ggerganov

examples/talk-llama/llama-adapter.cpp

@@ -1,5 +1,7 @@
 #include "llama-adapter.h"
 
+#include "llama-impl.h"
+#include "llama-mmap.h"
 #include "llama-model.h"
 
 #include <algorithm>
@@ -9,15 +11,15 @@
 
 // vec
 
-struct ggml_tensor * llama_control_vector::tensor_for(int il) const {
+struct ggml_tensor * llama_adapter_cvec::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 * llama_control_vector::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+struct ggml_tensor * llama_adapter_cvec::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);
@@ -26,12 +28,12 @@ struct ggml_tensor * llama_control_vector::apply_to(struct ggml_context * ctx, s
     return cur;
 }
 
-static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
+bool llama_adapter_cvec::init(const llama_model & model) {
     const auto & hparams = model.hparams;
 
-    GGML_ASSERT(cvec.tensors.empty());
-    GGML_ASSERT(cvec.ctxs.empty());
-    GGML_ASSERT(cvec.bufs.empty());
+    GGML_ASSERT(tensors.empty());
+    GGML_ASSERT(ctxs.empty());
+    GGML_ASSERT(bufs.empty());
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -50,7 +52,7 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
             }
 
             ctx_map[buft] = ctx;
-            cvec.ctxs.emplace_back(ctx);
+            ctxs.emplace_back(ctx);
 
             return ctx;
         }
@@ -59,21 +61,21 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
     };
 
     // make tensors
-    cvec.tensors.reserve(hparams.n_layer);
-    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
+    tensors.reserve(hparams.n_layer);
+    tensors.push_back(nullptr); // there's never a tensor for layer 0
     for (size_t il = 1; il < hparams.n_layer; il++) {
-        ggml_backend_buffer_type_t buft = llama_model_select_buft(model, il);
+        ggml_backend_buffer_type_t buft = model.select_buft(il);
         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, hparams.n_embd);
-        cvec.tensors.push_back(tensor);
+        tensors.push_back(tensor);
     }
 
     // allocate tensors / buffers and zero
-    cvec.bufs.reserve(ctx_map.size());
+    bufs.reserve(ctx_map.size());
     for (auto it : ctx_map) {
         ggml_backend_buffer_type_t buft = it.first;
         ggml_context * ctx = it.second;
@@ -83,14 +85,13 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
             return false;
         }
         ggml_backend_buffer_clear(buf, 0);
-        cvec.bufs.emplace_back(buf);
+        bufs.emplace_back(buf);
     }
 
     return true;
 }
 
-int32_t llama_control_vector_apply(
-        struct llama_control_vector & cvec,
+int32_t llama_adapter_cvec::apply(
         const llama_model & model,
         const float * data,
         size_t len,
@@ -101,8 +102,8 @@ int32_t llama_control_vector_apply(
 
     if (data == nullptr) {
         // disable the current control vector (but leave allocated for later)
-        cvec.layer_start = -1;
-        cvec.layer_end   = -1;
+        layer_start = -1;
+        layer_end   = -1;
         return 0;
     }
 
@@ -111,21 +112,21 @@ int32_t llama_control_vector_apply(
         return 1;
     }
 
-    if (cvec.tensors.empty()) {
-        if (!llama_control_vector_init(cvec, model)) {
+    if (tensors.empty()) {
+        if (!init(model)) {
             return 1;
         }
     }
 
-    cvec.layer_start = il_start;
-    cvec.layer_end   = il_end;
+    layer_start = il_start;
+    layer_end   = il_end;
 
     for (size_t il = 1; il < hparams.n_layer; il++) {
-        assert(cvec.tensors[il] != nullptr);
+        assert(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]));
+            ggml_backend_tensor_set(tensors[il], data + off, 0, n_embd * ggml_element_size(tensors[il]));
         }
     }
 
@@ -134,7 +135,7 @@ int32_t llama_control_vector_apply(
 
 // lora
 
-llama_lora_weight * llama_lora_adapter::get_weight(struct ggml_tensor * w) {
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) {
     const std::string name(w->name);
 
     const auto pos = ab_map.find(name);
@@ -145,11 +146,7 @@ llama_lora_weight * llama_lora_adapter::get_weight(struct ggml_tensor * w) {
     return nullptr;
 }
 
-void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
-    delete adapter;
-}
-
-static void llama_lora_adapter_init_impl(struct llama_model & model, const char * path_lora, struct llama_lora_adapter & adapter) {
+static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
     ggml_context * ctx_init;
@@ -221,7 +218,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     };
 
     // bundle lora_a and lora_b into pairs
-    std::map<std::string, llama_lora_weight> ab_map;
+    std::map<std::string, llama_adapter_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;
     };
@@ -231,17 +228,21 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
         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);
+                ab_map[name] = llama_adapter_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);
+                ab_map[name] = llama_adapter_lora_weight(nullptr, cur);
             } else {
                 ab_map[name].b = cur;
             }
+        } else if (str_endswith(name, "_norm.weight")) {
+            // TODO: add support for norm vector
+            // for now, we don't really care because most adapters still work fine without it
+            continue;
         } else {
             throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
         }
@@ -250,33 +251,41 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     // add tensors
     for (auto & it : ab_map) {
         const std::string & name = it.first;
-        llama_lora_weight & w = it.second;
+        llama_adapter_lora_weight & w = it.second;
+        bool is_token_embd = str_endswith(name, "token_embd.weight");
 
         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_model_get_tensor(model, name.c_str());
+        const auto * model_tensor = model.get_tensor(name.c_str());
         if (!model_tensor) {
-            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong 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)");
+        if (is_token_embd) {
+            // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
+            if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) {
+                throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
+            }
+        } else {
+            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 (hint: maybe wrong base model?)");
+            }
+            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);
+        adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
     }
 
     // allocate tensors / buffers and zero
@@ -318,11 +327,11 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
-struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model, const char * path_lora) {
-    struct llama_lora_adapter * adapter = new llama_lora_adapter();
+struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) {
+    struct llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
-        llama_lora_adapter_init_impl(*model, path_lora, *adapter);
+        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
         return adapter;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -332,3 +341,7 @@ struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model,
 
     return nullptr;
 }
+
+void llama_adapter_lora_free(struct llama_adapter_lora * adapter) {
+    delete adapter;
+}

examples/talk-llama/llama-adapter.h

@@ -1,66 +1,74 @@
 #pragma once
 
-#include "llama-impl.h"
-#include "llama-hparams.h"
+#include "llama.h"
 
 #include "ggml-cpp.h"
 
+#include <string>
 #include <unordered_map>
 #include <vector>
 
+// TODO: pimpl
+
 //
 // llama_adapter_cvec
 //
 
-// TODO: rename to llama_adapter_cvec
-struct llama_control_vector {
-    std::vector<ggml_context_ptr> ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+struct llama_adapter_cvec {
+    struct ggml_tensor * tensor_for(int il) const;
 
-    std::vector<struct ggml_tensor *> tensors; // per layer
+    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+
+    int32_t apply(
+            const llama_model & model,
+            const float * data,
+            size_t len,
+            int32_t n_embd,
+            int32_t il_start,
+            int32_t il_end);
+
+private:
+    bool init(const llama_model & model);
 
     int32_t layer_start = -1;
     int32_t layer_end   = -1;
 
-    struct ggml_tensor * tensor_for(int il) const;
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
 
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+    std::vector<struct ggml_tensor *> tensors; // per layer
 };
 
-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);
-
 //
 // llama_adapter_lora
 //
 
-// TODO: rename to llama_adapter_lora_weight
-struct llama_lora_weight {
+struct llama_adapter_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) {}
+    // get actual scale based on rank and alpha
+    float get_scale(float alpha, float adapter_scale) const {
+        const float rank  = (float) b->ne[0];
+        const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
+        return scale;
+    }
+
+    llama_adapter_lora_weight() = default;
+    llama_adapter_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_adapter_lora {
     // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_lora_weight> ab_map;
+    std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
 
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
     float alpha;
 
-    llama_lora_adapter() = default;
-    ~llama_lora_adapter() = default;
+    llama_adapter_lora() = default;
+    ~llama_adapter_lora() = default;
 
-    llama_lora_weight * get_weight(struct ggml_tensor * w);
+    llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
 };