feat: sync llama.cpp (#54)

Author: jhen0409

cpp/common.cpp

@@ -31,6 +31,8 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
+#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const *LLAMA_COMMIT;
@@ -103,7 +105,7 @@ struct gpt_params {
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model                = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model                = "";  // model path
     std::string model_draft          = "";  // draft model for speculative decoding
     std::string model_alias          = "unknown"; // model alias
     std::string model_url            = "";  // model url to download
@@ -144,7 +146,7 @@ struct gpt_params {
     bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
     size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
 
-    bool   kl_divergence   = false; // compute KL-divergence
+    bool   kl_divergence   = false; // compute KL divergence
 
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -159,6 +161,7 @@ struct gpt_params {
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
+    bool flash_attn        = false; // flash attention
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
@@ -172,15 +175,20 @@ struct gpt_params {
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
+    bool check_tensors     = false; // validate tensor data
 
     std::string cache_type_k = "f16"; // KV cache data type for the K
     std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
-    std::string mmproj = ""; // path to multimodal projector
-    std::string image  = ""; // path to an image file
+    std::string mmproj = "";        // path to multimodal projector
+    std::vector<std::string> image; // path to image file(s)
 };
 
+void gpt_params_handle_model_default(gpt_params & params);
+
+bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@@ -204,6 +212,7 @@ bool validate_file_name(const std::string & filename);
 std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
+std::string string_strip(const std::string & str);
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //

cpp/common.h

@@ -1784,12 +1784,14 @@ void lm_ggml_backend_sched_free(lm_ggml_backend_sched_t sched) {
 
 void lm_ggml_backend_sched_reset(lm_ggml_backend_sched_t sched) {
     // reset state for the next run
-    size_t hash_size = sched->hash_set.size;
-    memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
-    memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
-    memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
+    if (!sched->is_reset) {
+        size_t hash_size = sched->hash_set.size;
+        memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
+        memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
+        memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
 
-    sched->is_reset = true;
+        sched->is_reset = true;
+    }
     sched->is_alloc = false;
 }
 

cpp/ggml-backend.c

@@ -11,6 +11,12 @@
 #include <string.h> // memcpy
 #include <math.h>   // fabsf
 
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -307,7 +313,7 @@ inline static int32x4_t lm_ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t
 
 #endif // defined(__ARM_NEON)
 
-#if defined(__ARM_NEON) && !defined(__MSC_VER)
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
 
 #define LM_GGML_COMPUTE_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
 #define LM_GGML_COMPUTE_FP32_TO_FP16(x) lm_ggml_compute_fp32_to_fp16(x)