Merge branch 'master' into compilade/refactor-kv-cache

common/common.h

@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -58,13 +67,18 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
+struct cpu_params {
+    int      n_threads                   = -1;
+    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
+    bool     mask_valid                  = false;   // Default: any CPU
+    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
+    bool     strict_cpu                  = false;   // Use strict CPU placement
+    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
+};
+
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = cpu_get_num_math();
-    int32_t n_threads_draft       =    -1;
-    int32_t n_threads_batch       =    -1; // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft =    -1;
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@@ -91,6 +105,11 @@ struct gpt_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+    struct cpu_params draft_cpuparams;
+    struct cpu_params draft_cpuparams_batch;
+
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
@@ -99,6 +118,7 @@ struct gpt_params {
     enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -107,6 +127,7 @@ struct gpt_params {
     std::string model_draft          = ""; // draft model for speculative decoding
     std::string model_alias          = "unknown"; // model alias
     std::string model_url            = ""; // model url to download
+    std::string hf_token             = ""; // HF token
     std::string hf_repo              = ""; // HF repo
     std::string hf_file              = ""; // HF file
     std::string prompt               = "";
@@ -124,9 +145,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
-    std::string lora_base  = "";                              // base model path for the lora adapter
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -194,7 +214,7 @@ struct gpt_params {
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";
@@ -253,8 +273,11 @@ struct gpt_params {
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
+
+    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
 };
 
+void gpt_params_parse_from_env(gpt_params & params);
 void gpt_params_handle_model_default(gpt_params & params);
 
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
@@ -264,6 +287,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
 
 std::string gpt_params_get_system_info(const gpt_params & params);
 
+bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
+bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
+bool set_process_priority(enum ggml_sched_priority prio);
+
 //
 // String utils
 //
@@ -273,6 +301,8 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -304,14 +334,23 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
 
-struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
-struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
+struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
+struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
+
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
 
 // Batch utils
 
@@ -349,25 +388,13 @@ std::string llama_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
-// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
-//       that takes into account the tokenizer type and decides how to handle the leading space
-//
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
-// removes the leading space from the first non-BOS token
-std::string llama_detokenize_spm(
+// optionally renders special/control tokens
+std::string llama_detokenize(
                          llama_context * ctx,
-        const std::vector<llama_token> & tokens);
-
-// detokenizes a vector of tokens into a string
-// should work similar to Python's `tokenizer.decode`
-std::string llama_detokenize_bpe(
-                         llama_context * ctx,
-        const std::vector<llama_token> & tokens);
-
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
+        const std::vector<llama_token> & tokens,
+                                  bool   special = true);
 
 //
 // Chat template utils

common/grammar-parser.cpp

@@ -369,6 +369,9 @@ namespace grammar_parser {
             }
             // Validate the state to ensure that all rules are defined
             for (const auto & rule : state.rules) {
+                if (rule.empty()) {
+                    throw std::runtime_error("Undefined rule");
+                }
                 for (const auto & elem : rule) {
                     if (elem.type == LLAMA_GRETYPE_RULE_REF) {
                         // Ensure that the rule at that location exists

common/log.h

@@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
     buf << "[ ";
 
     bool first = true;
-    for (const auto &token : tokens)
+    for (const auto & token : tokens)
     {
         if (!first) {
             buf << ", ";

common/ngram-cache.h

@@ -37,11 +37,18 @@ struct llama_ngram {
     }
 };
 
+struct llama_token_hash_function {
+    size_t operator()(const llama_token token) const {
+        // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
+        return token * 11400714819323198485llu;
+    }
+};
+
 struct llama_ngram_hash_function {
     size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = 0;
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
+        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
+        for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
         }
         return hash;
     }

common/sampling.cpp

@@ -282,8 +282,6 @@ static llama_token llama_sampling_sample_impl(
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
 
     if (temp < 0.0) {
         // greedy sampling, with probs
@@ -324,12 +322,15 @@ static llama_token llama_sampling_sample_impl(
     }
 
     if (ctx_sampling->grammar != NULL && !is_resampling) {
+        // Get a pointer to the logits
+        float * logits = llama_get_logits_ith(ctx_main, idx);
+
         // Create an array with a single token data element for the sampled id
         llama_token_data single_token_data = {id, logits[id], 0.0f};
         llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
 
         // Apply grammar constraints to the single token
-        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
+        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
 
         // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
         bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
@@ -377,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
     if (ctx_sampling->grammar != NULL && !apply_grammar) {
         GGML_ASSERT(original_logits != NULL);
         // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
-        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+        *original_logits = {logits, logits + n_vocab};
     }
 
     // apply params.logit_bias map
@@ -390,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl(
         llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
     }
 
-    cur.clear();
+    cur.resize(n_vocab);
 
     for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
     }
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), false };
@@ -420,7 +421,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
 
     // apply grammar checks before sampling logic
     if (apply_grammar && ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
+        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
     }
 
     return cur_p;
@@ -454,6 +455,6 @@ void llama_sampling_accept(
     ctx_sampling->prev.push_back(id);
 
     if (ctx_sampling->grammar != NULL && apply_grammar) {
-        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
+        llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id);
     }
 }