cui-llama.rn 1.5.0 → 1.6.0

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-context.h

@@ -0,0 +1,265 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-graph.h"
+#include "llama-adapter.h"
+
+#include "ggml-cpp.h"
+
+#include <map>
+#include <vector>
+
+struct llama_model;
+struct llama_kv_cache;
+
+class llama_io_read_i;
+class llama_io_write_i;
+
+struct llama_context {
+    // init scheduler and compute buffers, reserve worst-case graphs
+    llama_context(
+            const llama_model & model,
+                  llama_context_params params);
+
+    ~llama_context();
+
+    void synchronize();
+
+    const llama_model & get_model() const;
+
+    uint32_t n_ctx()         const;
+    uint32_t n_ctx_per_seq() const;
+    uint32_t n_batch()       const;
+    uint32_t n_ubatch()      const;
+    uint32_t n_seq_max()     const;
+
+    uint32_t n_threads()       const;
+    uint32_t n_threads_batch() const;
+
+          llama_kv_cache * get_kv_self();
+    const llama_kv_cache * get_kv_self() const;
+
+    void kv_self_update();
+
+    enum llama_pooling_type pooling_type() const;
+
+    float * get_logits();
+    float * get_logits_ith(int32_t i);
+
+    float * get_embeddings();
+    float * get_embeddings_ith(int32_t i);
+    float * get_embeddings_seq(llama_seq_id seq_id);
+
+    void attach_threadpool(
+            lm_ggml_threadpool_t threadpool,
+            lm_ggml_threadpool_t threadpool_batch);
+
+    void detach_threadpool();
+
+    void set_n_threads(int32_t n_threads, int32_t n_threads_batch);
+
+    void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
+
+    void set_embeddings (bool value);
+    void set_causal_attn(bool value);
+    void set_warmup(bool value);
+
+    void set_adapter_lora(
+            llama_adapter_lora * adapter,
+            float scale);
+
+    bool rm_adapter_lora(
+            llama_adapter_lora * adapter);
+
+    void clear_adapter_lora();
+
+    bool apply_adapter_cvec(
+            const float * data,
+                 size_t   len,
+                int32_t   n_embd,
+                int32_t   il_start,
+                int32_t   il_end);
+
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
+
+    //
+    // state save/load
+    //
+
+    size_t state_get_size();
+    size_t state_get_data(      uint8_t * dst, size_t size);
+    size_t state_set_data(const uint8_t * src, size_t size);
+
+    size_t state_seq_get_size(llama_seq_id seq_id);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+
+    bool state_load_file(
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    bool state_save_file(
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    size_t state_seq_load_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    size_t state_seq_save_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    //
+    // perf
+    //
+
+    llama_perf_context_data perf_get_data() const;
+    void perf_reset();
+
+private:
+    //
+    // output
+    //
+
+    // Make sure enough space is available for outputs.
+    // Returns max number of outputs for which space was reserved.
+    int32_t output_reserve(int32_t n_outputs);
+
+    // make the outputs have the same order they had in the user-provided batch
+    // TODO: maybe remove this
+    void output_reorder();
+
+    //
+    // graph
+    //
+
+    int32_t graph_max_nodes() const;
+
+    // zero-out inputs and create the ctx_compute for the compute graph
+    lm_ggml_cgraph * graph_init();
+
+    llm_graph_result_ptr graph_build(
+            lm_ggml_context * ctx,
+             lm_ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
+    // returns the result of lm_ggml_backend_sched_graph_compute_async execution
+    lm_ggml_status graph_compute(
+            lm_ggml_cgraph * gf,
+                   bool   batched);
+
+    llm_graph_cb graph_get_cb() const;
+
+    // used by kv_self_update()
+    lm_ggml_tensor * build_rope_shift(
+        lm_ggml_context * ctx0,
+        lm_ggml_tensor * cur,
+        lm_ggml_tensor * shift,
+        lm_ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale,
+        lm_ggml_backend_buffer * bbuf) const;
+
+    llm_graph_result_ptr build_kv_self_shift(
+            lm_ggml_context * ctx0,
+            lm_ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_kv_self_defrag(
+            lm_ggml_context * ctx0,
+            lm_ggml_cgraph * gf) const;
+
+    // TODO: read/write lora adapters and cvec
+    size_t state_write_data(llama_io_write_i & io);
+    size_t state_read_data (llama_io_read_i  & io);
+
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+
+    //
+    // members
+    //
+
+    const llama_model & model;
+
+    llama_cparams       cparams;
+    llama_adapter_cvec  cvec;
+    llama_adapter_loras loras;
+    llama_sbatch        sbatch;
+
+    llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
+
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;
+
+    // decode output (2-dimensional array: [n_outputs][n_vocab])
+    size_t  logits_size = 0; // capacity (of floats) for logits
+    float * logits      = nullptr;
+
+    // embeddings output (2-dimensional array: [n_outputs][n_embd])
+    // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
+    size_t  embd_size = 0; // capacity (of floats) for embeddings
+    float * embd      = nullptr;
+
+    // sequence embeddings output (map of [n_embd] vectors)
+    // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
+    std::map<llama_seq_id, std::vector<float>> embd_seq;
+
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
+
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
+
+    lm_ggml_backend_sched_ptr sched;
+
+    lm_ggml_backend_t backend_cpu = nullptr;
+    std::vector<lm_ggml_backend_ptr> backends;
+
+    lm_ggml_context_ptr ctx_compute;
+
+    lm_ggml_threadpool_t threadpool       = nullptr;
+    lm_ggml_threadpool_t threadpool_batch = nullptr;
+
+    lm_ggml_abort_callback abort_callback      = nullptr;
+    void *              abort_callback_data = nullptr;
+
+    std::vector<std::pair<lm_ggml_backend_t, lm_ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
+    // buffer types used for the compute buffer of each backend
+    std::vector<lm_ggml_backend_t>             backend_ptrs;
+    std::vector<lm_ggml_backend_buffer_type_t> backend_buft;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    // host buffer for the model output (logits and embeddings)
+    lm_ggml_backend_buffer_ptr buf_output;
+
+    bool has_evaluated_once = false;
+
+    // perf
+    mutable int64_t t_start_us  = 0;
+    mutable int64_t t_load_us   = 0;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens    = 0;
+
+    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    mutable int32_t n_eval   = 0; // number of eval calls
+};

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-cparams.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+
+struct llama_cparams {
+    uint32_t n_ctx;           // context size used during inference
+    uint32_t n_batch;
+    uint32_t n_ubatch;
+    uint32_t n_seq_max;
+    int      n_threads;       // number of threads to use for generation
+    int      n_threads_batch; // number of threads to use for batch processing
+
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    uint32_t n_ctx_orig_yarn;
+    // These hyperparameters are not exposed in GGUF, because all
+    // existing YaRN models use the same values for them.
+    float yarn_ext_factor;
+    float yarn_attn_factor;
+    float yarn_beta_fast;
+    float yarn_beta_slow;
+    float defrag_thold;
+
+    bool embeddings;
+    bool causal_attn;
+    bool offload_kqv;
+    bool flash_attn;
+    bool no_perf;
+    bool warmup;
+
+    enum llama_pooling_type pooling_type;
+
+    lm_ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-cpp.h

@@ -0,0 +1,30 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include <memory>
+
+#include "llama.h"
+
+struct llama_model_deleter {
+    void operator()(llama_model * model) { llama_model_free(model); }
+};
+
+struct llama_context_deleter {
+    void operator()(llama_context * context) { llama_free(context); }
+};
+
+struct llama_sampler_deleter {
+    void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
+};
+
+struct llama_adapter_lora_deleter {
+    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
+};
+
+typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
+typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
+typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;
+typedef std::unique_ptr<llama_adapter_lora, llama_adapter_lora_deleter> llama_adapter_lora_ptr;

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-grammar.h

@@ -0,0 +1,173 @@
+#pragma once
+
+#include "llama.h"
+
+#include <map>
+#include <regex>
+#include <string>
+#include <vector>
+
+struct llama_vocab;
+
+// grammar element type
+enum llama_gretype {
+    // end of rule definition
+    LLAMA_GRETYPE_END            = 0,
+
+    // start of alternate definition for rule
+    LLAMA_GRETYPE_ALT            = 1,
+
+    // non-terminal element: reference to rule
+    LLAMA_GRETYPE_RULE_REF       = 2,
+
+    // terminal element: character (code point)
+    LLAMA_GRETYPE_CHAR           = 3,
+
+    // inverse char(s) ([^a], [^a-b] [^abc])
+    LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+    // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+    // be an inclusive range ([a-z])
+    LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+    // modifies a preceding LLAMA_GRETYPE_CHAR or
+    // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+    LLAMA_GRETYPE_CHAR_ALT       = 6,
+
+    // any character (.)
+    LLAMA_GRETYPE_CHAR_ANY       = 7,
+};
+
+typedef struct llama_grammar_element {
+    enum llama_gretype type;
+    uint32_t           value; // Unicode code point or rule ID
+} llama_grammar_element;
+
+struct llama_partial_utf8 {
+    uint32_t value;    // bit value so far (unshifted)
+    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
+};
+
+struct llama_grammar_candidate {
+    size_t               index;
+    const uint32_t     * code_points;
+    llama_partial_utf8   partial_utf8;
+};
+
+using llama_grammar_rule  = std::vector<      llama_grammar_element>;
+using llama_grammar_stack = std::vector<const llama_grammar_element *>;
+
+using llama_grammar_rules      = std::vector<llama_grammar_rule>;
+using llama_grammar_stacks     = std::vector<llama_grammar_stack>;
+using llama_grammar_candidates = std::vector<llama_grammar_candidate>;
+
+// TODO: remove, needed for tests atm
+const llama_grammar_rules  & llama_grammar_get_rules (const struct llama_grammar * grammar);
+      llama_grammar_stacks & llama_grammar_get_stacks(      struct llama_grammar * grammar);
+
+// takes a set of possible pushdown stacks on a grammar, which are required to
+// be positioned at a character range (see `llama_grammar_advance_stack`), and
+// produces the N possible stacks if the given char is accepted at those
+// positions
+void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr);
+
+std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
+        const llama_grammar_rules      & rules,
+        const llama_grammar_stack      & stack,
+        const llama_grammar_candidates & candidates);
+
+struct llama_grammar_parser {
+    std::map<std::string, uint32_t> symbol_ids;
+
+    llama_grammar_rules rules;
+
+    llama_grammar_stack c_rules() const;
+
+    uint32_t get_symbol_id(const char * src, size_t len);
+    uint32_t generate_symbol_id(const std::string & base_name);
+
+    void add_rule(uint32_t rule_id, const llama_grammar_rule & rule);
+
+    const char * parse_alternates(
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            const char         * src,
+            const std::string  & rule_name,
+            llama_grammar_rule & rule,
+            bool               is_nested);
+
+    const char * parse_rule(const char * src);
+
+    bool parse(const char * src);
+    void print(FILE * file);
+};
+
+struct llama_grammar_trigger_pattern {
+    std::string pattern;
+    std::regex  regex;
+};
+
+struct llama_grammar {
+    // note: allow null vocab for testing (not great)
+    const llama_vocab * vocab;
+
+    const llama_grammar_rules  rules;  // TODO: shared ptr
+          llama_grammar_stacks stacks;
+
+    // buffer for partially generated UTF-8 sequence from accepted tokens
+    llama_partial_utf8 partial_utf8;
+
+    // lazy grammars wait for trigger words or tokens before constraining the sampling.
+    // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // (useful e.g. for tool_choice=required)
+    bool                     lazy             = false;
+    bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
+    std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
+    std::vector<llama_grammar_trigger_pattern>
+                             trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
+                                                       // string, and the grammar will be given the string from the first match group onwards.
+
+};
+
+//
+// internal API
+//
+
+// note: needed for tests (not great)
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+        const llama_grammar_element ** rules,
+        size_t n_rules,
+        size_t start_rule_index);
+
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
+void llama_grammar_free_impl(struct llama_grammar * grammar);
+
+struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar);
+
+// TODO: move the API below as member functions of llama_grammar
+void llama_grammar_apply_impl(
+        const struct llama_grammar & grammar,
+            llama_token_data_array * cur_p);
+
+void llama_grammar_accept_impl(
+              struct llama_grammar & grammar,
+                       llama_token   token);
+
+void llama_grammar_accept_str(
+              struct llama_grammar & grammar,
+                 const std::string & piece);