cui-llama.rn 1.4.0 → 1.4.1

package/cpp/log.h

@@ -1,121 +1,121 @@
-#pragma once
-
-#include "ggml.h" // for lm_ggml_log_level
-
-#ifndef __GNUC__
-#    define LOG_ATTRIBUTE_FORMAT(...)
-#elif defined(__MINGW32__)
-#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
-#else
-#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-
-#define LOG_DEFAULT_DEBUG 1
-#define LOG_DEFAULT_LLAMA 0
-
-// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
-// set via common_log_set_verbosity()
-extern int common_log_verbosity_thold;
-
-void common_log_set_verbosity_thold(int verbosity); // not thread-safe
-
-// the common_log uses an internal worker thread to print/write log messages
-// when the worker thread is paused, incoming log messages are discarded
-struct common_log;
-
-struct common_log * common_log_init();
-struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
-void                common_log_pause (struct common_log * log); // pause  the worker thread, not thread-safe
-void                common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
-void                common_log_free  (struct common_log * log);
-
-LOG_ATTRIBUTE_FORMAT(3, 4)
-void common_log_add(struct common_log * log, enum lm_ggml_log_level level, const char * fmt, ...);
-
-// defaults: file = NULL, colors = false, prefix = false, timestamps = false
-//
-// regular log output:
-//
-//   lm_ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
-//   llm_load_tensors: ggml ctx size =    0.27 MiB
-//   llm_load_tensors: offloading 32 repeating layers to GPU
-//   llm_load_tensors: offloading non-repeating layers to GPU
-//
-// with prefix = true, timestamps = true, the log output will look like this:
-//
-//   0.00.035.060 D lm_ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
-//   0.00.035.064 I llm_load_tensors: ggml ctx size =    0.27 MiB
-//   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
-//   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
-//
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
-// D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
-//
-
-void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
-void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
-void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
-void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
-
-// helper macros for logging
-// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
-//
-// for example:
-//
-//   LOG_DBG("this is a debug message: %d\n", expensive_function());
-//
-// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > common_log_verbosity_thold
-//
-
-
-
-#if defined(__ANDROID__)
-#include <android/log.h>
-#define LLAMA_ANDROID_LOG_TAG "RNLLAMA_LOG_ANDROID"
-
-#if defined(RNLLAMA_ANDROID_ENABLE_LOGGING)
-#define RNLLAMA_LOG_LEVEL 1
-#else
-#define RNLLAMA_LOG_LEVEL 0
-#endif
-
-#define LOG_TMPL(level, verbosity, ...) \
-    do { \
-        if ((verbosity) <= RNLLAMA_LOG_LEVEL) { \
-            int android_log_level = ANDROID_LOG_DEFAULT; \
-            switch (level) { \
-                case LM_GGML_LOG_LEVEL_INFO:  android_log_level = ANDROID_LOG_INFO; break; \
-                case LM_GGML_LOG_LEVEL_WARN:  android_log_level = ANDROID_LOG_WARN; break; \
-                case LM_GGML_LOG_LEVEL_ERROR: android_log_level = ANDROID_LOG_ERROR; break; \
-                default:             android_log_level = ANDROID_LOG_DEFAULT; \
-            } \
-            __android_log_print(android_log_level, LLAMA_ANDROID_LOG_TAG, __VA_ARGS__); \
-        } \
-    } while(0)
-#else
-
-#define LOG_TMPL(level, verbosity, ...) \
-    do { \
-        if ((verbosity) <= common_log_verbosity_thold) { \
-            common_log_add(common_log_main(), (level), __VA_ARGS__); \
-        } \
-    } while (0)
-
-#endif
-
-#define LOG(...)             LOG_TMPL(LM_GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
-#define LOGV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
-
-#define LOG_INF(...) LOG_TMPL(LM_GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(LM_GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(LM_GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(LM_GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(LM_GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
-
-#define LOG_INFV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
-#define LOG_WRNV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
-#define LOG_ERRV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
-#define LOG_DBGV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
-#define LOG_CNTV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)
+#pragma once
+
+#include "ggml.h" // for lm_ggml_log_level
+
+#ifndef __GNUC__
+#    define LOG_ATTRIBUTE_FORMAT(...)
+#elif defined(__MINGW32__)
+#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+
+#define LOG_DEFAULT_DEBUG 1
+#define LOG_DEFAULT_LLAMA 0
+
+// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
+// set via common_log_set_verbosity()
+extern int common_log_verbosity_thold;
+
+void common_log_set_verbosity_thold(int verbosity); // not thread-safe
+
+// the common_log uses an internal worker thread to print/write log messages
+// when the worker thread is paused, incoming log messages are discarded
+struct common_log;
+
+struct common_log * common_log_init();
+struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
+void                common_log_pause (struct common_log * log); // pause  the worker thread, not thread-safe
+void                common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
+void                common_log_free  (struct common_log * log);
+
+LOG_ATTRIBUTE_FORMAT(3, 4)
+void common_log_add(struct common_log * log, enum lm_ggml_log_level level, const char * fmt, ...);
+
+// defaults: file = NULL, colors = false, prefix = false, timestamps = false
+//
+// regular log output:
+//
+//   lm_ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
+//   llm_load_tensors: ggml ctx size =    0.27 MiB
+//   llm_load_tensors: offloading 32 repeating layers to GPU
+//   llm_load_tensors: offloading non-repeating layers to GPU
+//
+// with prefix = true, timestamps = true, the log output will look like this:
+//
+//   0.00.035.060 D lm_ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
+//   0.00.035.064 I llm_load_tensors: ggml ctx size =    0.27 MiB
+//   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
+//   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
+//
+// I - info    (stdout, V = 0)
+// W - warning (stderr, V = 0)
+// E - error   (stderr, V = 0)
+// D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
+//
+
+void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
+void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
+void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
+void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+
+// helper macros for logging
+// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
+//
+// for example:
+//
+//   LOG_DBG("this is a debug message: %d\n", expensive_function());
+//
+// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > common_log_verbosity_thold
+//
+
+
+
+#if defined(__ANDROID__)
+#include <android/log.h>
+#define LLAMA_ANDROID_LOG_TAG "RNLLAMA_LOG_ANDROID"
+
+#if defined(RNLLAMA_ANDROID_ENABLE_LOGGING)
+#define RNLLAMA_LOG_LEVEL 1
+#else
+#define RNLLAMA_LOG_LEVEL 0
+#endif
+
+#define LOG_TMPL(level, verbosity, ...) \
+    do { \
+        if ((verbosity) <= RNLLAMA_LOG_LEVEL) { \
+            int android_log_level = ANDROID_LOG_DEFAULT; \
+            switch (level) { \
+                case LM_GGML_LOG_LEVEL_INFO:  android_log_level = ANDROID_LOG_INFO; break; \
+                case LM_GGML_LOG_LEVEL_WARN:  android_log_level = ANDROID_LOG_WARN; break; \
+                case LM_GGML_LOG_LEVEL_ERROR: android_log_level = ANDROID_LOG_ERROR; break; \
+                default:             android_log_level = ANDROID_LOG_DEFAULT; \
+            } \
+            __android_log_print(android_log_level, LLAMA_ANDROID_LOG_TAG, __VA_ARGS__); \
+        } \
+    } while(0)
+#else
+
+#define LOG_TMPL(level, verbosity, ...) \
+    do { \
+        if ((verbosity) <= common_log_verbosity_thold) { \
+            common_log_add(common_log_main(), (level), __VA_ARGS__); \
+        } \
+    } while (0)
+
+#endif
+
+#define LOG(...)             LOG_TMPL(LM_GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
+#define LOGV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
+
+#define LOG_INF(...) LOG_TMPL(LM_GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(LM_GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(LM_GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(LM_GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(LM_GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+
+#define LOG_INFV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
+#define LOG_WRNV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
+#define LOG_ERRV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
+#define LOG_DBGV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
+#define LOG_CNTV(verbosity, ...) LOG_TMPL(LM_GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)

package/cpp/rn-llama.hpp

@@ -219,7 +219,7 @@ struct llama_rn_context
     std::string stopping_word;
     bool incomplete = false;
 
-    std::vector<common_lora_adapter_info> lora;
+    std::vector<common_adapter_lora_info> lora;
 
     ~llama_rn_context()
     {
@@ -279,7 +279,7 @@ struct llama_rn_context
 
     bool validateModelChatTemplate() const {
         llama_chat_message chat[] = {{"user", "test"}};
-        int32_t chat_res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0);
+        int32_t chat_res = llama_chat_apply_template(llama_model_chat_template(model), chat, 1, true, nullptr, 0);
         return chat_res > 0;
     }
 
@@ -307,7 +307,7 @@ struct llama_rn_context
 
     void loadPrompt()
     {
-        std::vector<llama_token> prompt_tokens = ::common_tokenize(model, params.prompt, true, true);
+        std::vector<llama_token> prompt_tokens = ::common_tokenize(llama_model_get_vocab(model), params.prompt, true, true);
         num_prompt_tokens = prompt_tokens.size();
 
         // LOG tokens
@@ -439,14 +439,14 @@ struct llama_rn_context
         if (params.n_predict == 0)
         {
             has_next_token = false;
-            result.tok = llama_token_eos(model);
+            result.tok = llama_vocab_eos(llama_model_get_vocab(model));
             return result;
         }
 
         {
             // out of user input, sample next token
             std::vector<llama_token_data> candidates;
-            candidates.reserve(llama_n_vocab(model));
+            candidates.reserve(llama_vocab_n_tokens(llama_model_get_vocab(model)));
 
             result.tok = common_sampler_sample(ctx_sampling, ctx, -1);
 
@@ -479,7 +479,7 @@ struct llama_rn_context
         // decrement remaining sampling budget
         --n_remain;
 
-        if (!embd.empty() && embd.back() == llama_token_eos(model))
+        if (!embd.empty() && embd.back() == llama_vocab_eos(llama_model_get_vocab(model)))
         {
             // stopping_word = llama_token_to_piece(ctx, embd.back());
             has_next_token = false;
@@ -584,7 +584,7 @@ struct llama_rn_context
 
     std::vector<float> getEmbedding(common_params &embd_params)
     {
-        static const int n_embd = llama_n_embd(llama_get_model(ctx));
+        static const int n_embd = llama_model_n_embd(llama_get_model(ctx));
         if (!embd_params.embedding)
         {
             LOG_WARNING("embedding disabled, embedding: %s", embd_params.embedding);
@@ -716,25 +716,31 @@ struct llama_rn_context
             std::string("]");
     }
 
-    int applyLoraAdapters(std::vector<common_lora_adapter_info> lora) {
+    int applyLoraAdapters(std::vector<common_adapter_lora_info> lora) {
         for (auto &la : lora) {
-            la.ptr = llama_lora_adapter_init(model, la.path.c_str());
+            la.ptr = llama_adapter_lora_init(model, la.path.c_str());
             if (la.ptr == nullptr) {
                 LOG_ERROR("failed to apply lora adapter '%s'\n", la.path.c_str());
                 return -1;
             }
         }
         this->lora = lora;
-        common_lora_adapters_apply(ctx, lora);
+        for (auto &la : lora) {
+            llama_set_adapter_lora(ctx, la.ptr, 1);
+        }
+
         return 0;
     }
 
     void removeLoraAdapters() {
+        for (auto &la : this->lora) {
+            llama_adapter_lora_free(la.ptr);
+        }
         this->lora.clear();
-        common_lora_adapters_apply(ctx, this->lora); // apply empty list
+        llama_clear_adapter_lora(ctx);
     }
 
-    std::vector<common_lora_adapter_info> getLoadedLoraAdapters() {
+    std::vector<common_adapter_lora_info> getLoadedLoraAdapters() {
         return this->lora;
     }
 // Context Shifting from KoboldCpp <https://github.com/LostRuins/koboldcpp>