cui-llama.rn 1.7.4 → 1.7.6

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/json-partial.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include "nlohmann/json.hpp"
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/json-schema-to-grammar.h

@@ -1,9 +1,9 @@
 #pragma once
 
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to LM_GGML_ASSERT:
-#define JSON_ASSERT LM_GGML_ASSERT
-#include "json.hpp"
+#include "nlohmann/json_fwd.hpp"
+
+#include <functional>
+#include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);

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

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
+    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -45,6 +46,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
+    LLM_ARCH_GEMMA3N,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
@@ -76,6 +78,8 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -148,6 +152,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -190,13 +195,13 @@ enum llm_kv {
     LLM_KV_TOKENIZER_MASK_ID,
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_SEP,
     LLM_KV_TOKENIZER_ADD_PREFIX,
     LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
-    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -213,6 +218,8 @@ enum llm_kv {
     LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
     LLM_KV_CONVNEXT_BLOCK_COUNT,
 
+    LLM_KV_CLASSIFIER_OUTPUT_LABELS,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -263,6 +270,22 @@ enum llm_tensor {
     LLM_TENSOR_LAYER_OUT_NORM,
     LLM_TENSOR_POST_ATTN_NORM,
     LLM_TENSOR_POST_MLP_NORM,
+    LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
+    LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n
+    LLM_TENSOR_PER_LAYER_INP_GATE,   // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ,       // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ_NORM,  // gemma3n
+    LLM_TENSOR_PER_LAYER_POST_NORM,  // gemma3n
+    LLM_TENSOR_ALTUP_PROJ,           // gemma3n
+    LLM_TENSOR_ALTUP_UNEMBD_PROJ,    // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_SCALE,  // gemma3n
+    LLM_TENSOR_ALTUP_PREDICT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER,         // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER_NORM,    // gemma3n
+    LLM_TENSOR_LAUREL_L,             // gemma3n
+    LLM_TENSOR_LAUREL_R,             // gemma3n
+    LLM_TENSOR_LAUREL_POST_NORM,     // gemma3n
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -435,3 +458,6 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
+
+bool llm_arch_is_recurrent(const llm_arch & arch);
+bool llm_arch_is_hybrid   (const llm_arch & arch);

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

@@ -2,88 +2,146 @@
 
 #include "llama.h"
 
+#include "llama-cparams.h"
+
 #include <array>
 #include <vector>
+#include <set>
+#include <bitset>
+#include <unordered_map>
 
-// very similar to llama_batch,
-// but has more metadata about sequences
+// keep this struct lightweight
+// it points to data in `llama_batch_allocr`
 struct llama_ubatch {
     bool equal_seqs;
     // TODO: whole_seqs for embeddings?
 
-    uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs)
-    uint32_t n_seq_tokens; // tokens per sequence
-    uint32_t n_seqs;
-
-    llama_token  *  token;    // [n_tokens]
-    float        *  embd;     // [n_embd, n_tokens]
-    llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs]
-    llama_seq_id ** seq_id;   // [n_seqs]
-    int8_t       *  output;   // [n_tokens]
+    uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
+    uint32_t n_seq_tokens; // tokens per sequence set
+    uint32_t n_seqs;       // sequence sets in the ubatch
+    uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
+
+    // seq_id_unq: unique sequence ids in the ubatch
+    // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
+    //             used for extracting sequence pooled embeddings
+
+    //                          // size               | idx | val
+    llama_token  *  token;      // [n_tokens]         | i   | id, token
+    float        *  embd;       // [n_embd, n_tokens] | i   | embd
+    llama_pos    *  pos;        // [n_tokens]         | i   | pos
+    int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
+    llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
+    llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
+    int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
+    int8_t       *  output;     // [n_tokens]         | i   | -
 };
 
-struct llama_sbatch_seq {
-    int32_t n_seq_id;
+// a helper for sanitizing, fulfilling and splitting a batch
+class llama_batch_allocr {
+public:
+    llama_batch_allocr(uint32_t n_pos_per_embd);
 
-    llama_seq_id * seq_id;
+    // sanitize and auto-gen missing data in the input batch
+    // memory is optional. if provided will be used to check for sequence continuity and to determine the positions
+    bool init(
+            const llama_batch & batch_inp,
+            const llama_vocab & vocab,
+            const llama_memory_i * memory,
+            uint32_t n_embd,
+            bool output_all);
 
-    size_t offset;
-    size_t length;
-};
+    const llama_batch & get_batch() const;
 
-// sequence-length-aware batch splitting
-struct llama_sbatch {
-    // tokens left in this batch
-    size_t n_tokens;
+    uint32_t get_n_tokens()  const;
+    uint32_t get_n_outputs() const;
 
-    size_t n_embd;
+    // the array of output indices in the order they were encountered during the ubatch splitting
+    std::vector<int32_t> & get_out_ids();
 
-    bool logits_all; // TODO: remove once lctx.logits_all is removed too
+    // min/max positions of each sequence in the current ubatch
+    llama_pos seq_pos_min(llama_seq_id seq_id) const;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const;
 
-    // sorted indices into the batch
-    std::vector<int64_t> ids;
-    // batch indices of the output
-    std::vector<int64_t> out_ids;
-    std::vector<llama_sbatch_seq> seq;
+    // call once before splitting the batch to reset the internal state
+    void split_reset();
 
-    const llama_batch * batch = nullptr;
+    // simple split, unknown number of sequence sets of unequal lengths
+    llama_ubatch split_simple(uint32_t n_ubatch);
 
-    // buffers for the ubatch
-    std::vector<llama_token>    ubatch_token;
-    std::vector<float>          ubatch_embd;
-    std::vector<llama_pos>      ubatch_pos;
-    std::vector<int32_t>        ubatch_n_seq_id;
-    std::vector<llama_seq_id *> ubatch_seq_id;
-    std::vector<int8_t>         ubatch_output;
+    // make ubatches of equal-length sequences sets
+    llama_ubatch split_equal(uint32_t n_ubatch);
 
-    llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);
+    // sequence-set-wise split - each ubatch contains a single sequence-set
+    llama_ubatch split_seq(uint32_t n_ubatch);
 
-    void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length);
+    // a helper method for creating a well-defined ubatch of tokens
+    // TODO: support embeddings if needed in the future
+    llama_ubatch ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs);
 
-    // simple split, unknown number of sequences of unequal lengths
-    llama_ubatch split_simple(size_t n_ubatch);
+private:
+    void clear();
 
-    // make batches of equal-length sequences
-    llama_ubatch split_equal(size_t n_ubatch);
+    // create the next ubatch based on the provided batch indices (idxs) and the number of sequence sets (n_seqs)
+    // return llama_ubatch.n_tokens == 0 if the entire batch was consumed
+    llama_ubatch ubatch_add(const std::vector<int32_t> & idxs, uint32_t n_seqs, bool equal_seqs);
 
-    // sequence-wise split
-    llama_ubatch split_seq(size_t n_ubatch);
+    // for debugging, start with LLAMA_BATCH_DEBUG=2
+    void ubatch_print(const llama_ubatch & ubatch, int debug);
 
-    llama_sbatch() = default;
-    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
-};
+    llama_batch batch;
+
+    // only for debugging purposes
+    const llama_vocab * vocab;
+
+    // TODO: this is more of a temporary solution until we have a better way to handle multiple positions per token/embd
+    //       ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
+    const uint32_t n_pos_per_embd;
 
-// temporary allocate memory for the input batch if needed
-struct llama_batch_allocr {
-    struct llama_batch batch;
+    uint32_t n_embd;
+    uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+
     std::vector<llama_pos>      pos;
     std::vector<int32_t>        n_seq_id;
     std::vector<llama_seq_id *> seq_id;
-    std::vector<int8_t>         logits;
+    std::vector<llama_seq_id>   seq_id_unq;
+    std::vector<int32_t>        seq_idx;
+    std::vector<int8_t>         output;
+
+    using pos_set_t = std::set<llama_pos>;
+    using seq_cpl_t = std::vector<bool>;
+
+    std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
+    std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
+
+    using idx_vec_t = std::vector<int32_t>;
+    using seq_set_t = std::bitset<LLAMA_MAX_SEQ>;
 
-    // optionally fulfill the batch returned by llama_batch_get_one
-    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
+    std::vector<seq_set_t> seq_set; // seq_set[i]: the sequence set of token i
+
+    std::unordered_map<seq_set_t, idx_vec_t> seq_set_map; // the indices at which the sequence set appears
+
+    // batch indices of the output
+    std::vector<int32_t> out_ids;
+
+    // used[i] indicates if token i has already been used in a previous ubatch
+    std::vector<bool> used;
+
+    // llama_ubatch points to this data:
+    struct ubatch {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id>   seq_id_unq;
+        std::vector<int32_t>        seq_idx;
+        std::vector<int8_t>         output;
+    };
+
+    // current splitting state:
+    std::vector<ubatch> ubatches;
+
+    int debug;
 };

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

@@ -43,6 +43,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
+    LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -1,7 +1,6 @@
 #pragma once
 
 #include "llama.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -13,11 +12,14 @@
 #include <vector>
 
 struct llama_model;
-struct llama_kv_cache;
+class llama_batch_allocr;
 
 class llama_io_read_i;
 class llama_io_write_i;
 
+struct llama_memory_i;
+struct llama_memory_context_i;
+
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
     llama_context(
@@ -44,10 +46,12 @@ struct llama_context {
     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;
+    llama_memory_t get_memory() const;
 
-    void kv_self_update();
+    // return true of the KV cache was updated
+    // TODO: remove
+    bool kv_self_update(bool optimize);
+    void kv_self_defrag_sched();
 
     enum llama_pooling_type pooling_type() const;
 
@@ -88,8 +92,18 @@ struct llama_context {
                 int32_t   il_start,
                 int32_t   il_end);
 
-    int encode(llama_batch & inp_batch);
-    int decode(llama_batch & inp_batch);
+    // process a single ubatch with a specific graph type
+    // if memory_context is provided, it will be applied first to the context's memory
+    // ret contains the status of the graph computation
+    // returns nullptr only if ret != LM_GGML_STATUS_SUCCESS
+    llm_graph_result_ptr process_ubatch(
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+            llama_memory_context_i * mctx,
+                       lm_ggml_status & ret);
+
+    int encode(const llama_batch & batch_inp);
+    int decode(const llama_batch & batch_inp);
 
     //
     // state save/load
@@ -167,7 +181,7 @@ private:
 
     // 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);
+    uint32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -180,16 +194,18 @@ public:
     lm_ggml_cgraph * graph_init();
 
     // returns the result of lm_ggml_backend_sched_graph_compute_async execution
-    lm_ggml_status graph_compute(
-            lm_ggml_cgraph * gf,
-                   bool   batched);
+    lm_ggml_status graph_compute(lm_ggml_cgraph * gf, bool batched);
+
+    // reserve a graph with a dummy ubatch of the specified size
+    lm_ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
 
 private:
     llm_graph_result_ptr graph_build(
-            lm_ggml_context * ctx,
-             lm_ggml_cgraph * gf,
-      const llama_ubatch & ubatch,
-          llm_graph_type   gtype);
+                      lm_ggml_context * ctx,
+                       lm_ggml_cgraph * gf,
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+      const llama_memory_context_i * mctx);
 
     llm_graph_cb graph_get_cb() const;
 
@@ -214,6 +230,9 @@ private:
 
     std::unique_ptr<llama_memory_i> memory;
 
+    // TODO: temporary, until the llama_kv_self_defrag() API is removed
+    bool memory_force_optimize = false;
+
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
@@ -227,8 +246,10 @@ private:
     // 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
+    // reuse the batch_allocr to avoid unnecessary memory allocations
+    std::unique_ptr<llama_batch_allocr> balloc;
+
+    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 

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

@@ -4,6 +4,8 @@
 
 #include <cstdint>
 
+#define LLAMA_MAX_SEQ 64
+
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
     uint32_t n_batch;