cui-llama.rn 1.6.0 → 1.7.0

package/ios/rnllama.xcframework/tvos-arm64/rnllama.framework/Headers/ggml.h

@@ -394,8 +394,8 @@ extern "C" {
 
     // precision
     enum lm_ggml_prec {
-        LM_GGML_PREC_DEFAULT,
-        LM_GGML_PREC_F32,
+        LM_GGML_PREC_DEFAULT =  0, // stored as lm_ggml_tensor.op_params, 0 by default
+        LM_GGML_PREC_F32     = 10,
     };
 
     // model file types
@@ -482,6 +482,7 @@ extern "C" {
         LM_GGML_OP_CONV_TRANSPOSE_1D,
         LM_GGML_OP_IM2COL,
         LM_GGML_OP_IM2COL_BACK,
+        LM_GGML_OP_CONV_2D_DW,
         LM_GGML_OP_CONV_TRANSPOSE_2D,
         LM_GGML_OP_POOL_1D,
         LM_GGML_OP_POOL_2D,
@@ -508,17 +509,12 @@ extern "C" {
 
         LM_GGML_OP_UNARY,
 
-        LM_GGML_OP_MAP_UNARY,
-        LM_GGML_OP_MAP_BINARY,
-
-        LM_GGML_OP_MAP_CUSTOM1_F32,
-        LM_GGML_OP_MAP_CUSTOM2_F32,
-        LM_GGML_OP_MAP_CUSTOM3_F32,
-
         LM_GGML_OP_MAP_CUSTOM1,
         LM_GGML_OP_MAP_CUSTOM2,
         LM_GGML_OP_MAP_CUSTOM3,
 
+        LM_GGML_OP_CUSTOM,
+
         LM_GGML_OP_CROSS_ENTROPY_LOSS,
         LM_GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         LM_GGML_OP_OPT_STEP_ADAMW,
@@ -541,6 +537,7 @@ extern "C" {
         LM_GGML_UNARY_OP_HARDSWISH,
         LM_GGML_UNARY_OP_HARDSIGMOID,
         LM_GGML_UNARY_OP_EXP,
+        LM_GGML_UNARY_OP_GELU_ERF,
 
         LM_GGML_UNARY_OP_COUNT,
     };
@@ -678,11 +675,18 @@ extern "C" {
     LM_GGML_API bool lm_ggml_is_3d        (const struct lm_ggml_tensor * tensor);
     LM_GGML_API int  lm_ggml_n_dims       (const struct lm_ggml_tensor * tensor); // returns 1 for scalars
 
+    // returns whether the tensor elements can be iterated over with a flattened index (no gaps, no permutation)
     LM_GGML_API bool lm_ggml_is_contiguous  (const struct lm_ggml_tensor * tensor);
     LM_GGML_API bool lm_ggml_is_contiguous_0(const struct lm_ggml_tensor * tensor); // same as lm_ggml_is_contiguous()
     LM_GGML_API bool lm_ggml_is_contiguous_1(const struct lm_ggml_tensor * tensor); // contiguous for dims >= 1
     LM_GGML_API bool lm_ggml_is_contiguous_2(const struct lm_ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok)
+    LM_GGML_API bool lm_ggml_is_contiguously_allocated(const struct lm_ggml_tensor * tensor);
+
+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    LM_GGML_API bool lm_ggml_is_contiguous_channels(const struct lm_ggml_tensor * tensor);
+
     LM_GGML_API bool lm_ggml_are_same_shape (const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1);
     LM_GGML_API bool lm_ggml_are_same_stride(const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1);
 
@@ -766,7 +770,7 @@ extern "C" {
     // Tensor flags
     LM_GGML_API void lm_ggml_set_input(struct lm_ggml_tensor * tensor);
     LM_GGML_API void lm_ggml_set_output(struct lm_ggml_tensor * tensor);
-    LM_GGML_API void lm_ggml_set_param(struct lm_ggml_context * ctx, struct lm_ggml_tensor * tensor);
+    LM_GGML_API void lm_ggml_set_param(struct lm_ggml_tensor * tensor);
     LM_GGML_API void lm_ggml_set_loss(struct lm_ggml_tensor * tensor);
 
     //
@@ -936,7 +940,7 @@ extern "C" {
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_repeat_back(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
-            struct lm_ggml_tensor  * b);
+            struct lm_ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
     // concat a and b along dim
     // used in stable-diffusion
@@ -1022,6 +1026,16 @@ extern "C" {
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a);
 
+    // GELU using erf (error function) when possible
+    // some backends may fallback to approximation based on Abramowitz and Stegun formula
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_gelu_erf(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a);
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_gelu_erf_inplace(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a);
+
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_gelu_quick(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a);
@@ -1666,7 +1680,7 @@ extern "C" {
             struct lm_ggml_tensor  * a,
             struct lm_ggml_tensor  * b);
 
-    // depthwise
+    // depthwise (via im2col and mul_mat)
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_2d_dw(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,  // convolution kernel
@@ -1678,6 +1692,22 @@ extern "C" {
             int                  d0,  // dilation dimension 0
             int                  d1); // dilation dimension 1
 
+    // Depthwise 2D convolution
+    // may be faster than lm_ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_2d_dw_direct(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+            struct lm_ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_2d_p0(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
@@ -1723,24 +1753,29 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    // nearest interpolate
+    enum lm_ggml_scale_mode {
+        LM_GGML_SCALE_MODE_NEAREST  = 0,
+        LM_GGML_SCALE_MODE_BILINEAR = 1,
+    };
+
+    // interpolate
     // multiplies ne0 and ne1 by scale factor
-    // used in stable-diffusion
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_upscale(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
-            int                   scale_factor);
+            int                   scale_factor,
+            enum lm_ggml_scale_mode  mode);
 
-    // nearest interpolate
-    // nearest interpolate to specified dimensions
-    // used in tortoise.cpp
+    // interpolate
+    // interpolate scale to specified dimensions
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_upscale_ext(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
-            int                   ne3);
+            int                   ne3,
+            enum lm_ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_pad(
@@ -1917,83 +1952,6 @@ extern "C" {
 
     // custom operators
 
-    typedef void (*lm_ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*lm_ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*lm_ggml_custom1_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-    typedef void (*lm_ggml_custom2_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-    typedef void (*lm_ggml_custom3_op_f32_t)(struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *, const struct lm_ggml_tensor *);
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_unary_f32(
-            struct lm_ggml_context        * ctx,
-            struct lm_ggml_tensor         * a,
-                   lm_ggml_unary_op_f32_t   fun),
-        "use lm_ggml_map_custom1 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_unary_inplace_f32(
-            struct lm_ggml_context        * ctx,
-            struct lm_ggml_tensor         * a,
-                   lm_ggml_unary_op_f32_t   fun),
-        "use lm_ggml_map_custom1_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_binary_f32(
-            struct lm_ggml_context         * ctx,
-            struct lm_ggml_tensor          * a,
-            struct lm_ggml_tensor          * b,
-                   lm_ggml_binary_op_f32_t   fun),
-        "use lm_ggml_map_custom2 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_binary_inplace_f32(
-            struct lm_ggml_context         * ctx,
-            struct lm_ggml_tensor          * a,
-            struct lm_ggml_tensor          * b,
-                   lm_ggml_binary_op_f32_t   fun),
-        "use lm_ggml_map_custom2_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom1_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-                   lm_ggml_custom1_op_f32_t   fun),
-        "use lm_ggml_map_custom1 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom1_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-                   lm_ggml_custom1_op_f32_t   fun),
-        "use lm_ggml_map_custom1_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom2_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-                   lm_ggml_custom2_op_f32_t   fun),
-        "use lm_ggml_map_custom2 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom2_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-                   lm_ggml_custom2_op_f32_t   fun),
-        "use lm_ggml_map_custom2_inplace instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom3_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-            struct lm_ggml_tensor           * c,
-                   lm_ggml_custom3_op_f32_t   fun),
-        "use lm_ggml_map_custom3 instead");
-
-    LM_GGML_DEPRECATED(LM_GGML_API struct lm_ggml_tensor * lm_ggml_map_custom3_inplace_f32(
-            struct lm_ggml_context          * ctx,
-            struct lm_ggml_tensor           * a,
-            struct lm_ggml_tensor           * b,
-            struct lm_ggml_tensor           * c,
-                   lm_ggml_custom3_op_f32_t   fun),
-        "use lm_ggml_map_custom3_inplace instead");
-
-    // custom operators v2
-
     typedef void (*lm_ggml_custom1_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, int ith, int nth, void * userdata);
     typedef void (*lm_ggml_custom2_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, const struct lm_ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*lm_ggml_custom3_op_t)(struct lm_ggml_tensor * dst , const struct lm_ggml_tensor * a, const struct lm_ggml_tensor * b, const struct lm_ggml_tensor * c, int ith, int nth, void * userdata);
@@ -2049,6 +2007,30 @@ extern "C" {
             int                     n_tasks,
             void                  * userdata);
 
+    typedef void (*lm_ggml_custom_op_t)(struct lm_ggml_tensor * dst , int ith, int nth, void * userdata);
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_custom_4d(
+            struct lm_ggml_context * ctx,
+            enum lm_ggml_type        type,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            struct lm_ggml_tensor ** args,
+            int                   n_args,
+            lm_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_custom_inplace(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+            struct lm_ggml_tensor ** args,
+            int                   n_args,
+            lm_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
     // loss function
 
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_cross_entropy_loss(
@@ -2079,15 +2061,14 @@ extern "C" {
 
     LM_GGML_API void lm_ggml_build_forward_expand(struct lm_ggml_cgraph * cgraph, struct lm_ggml_tensor * tensor);
     LM_GGML_API void lm_ggml_build_backward_expand(
-        struct lm_ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-        struct lm_ggml_context * ctx_compute, // context for gradient computation
-        struct lm_ggml_cgraph  * cgraph,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+        struct lm_ggml_context *  ctx,        // context for gradient computation
+        struct lm_ggml_cgraph  *  cgraph,
+        struct lm_ggml_tensor  ** grad_accs);
 
     // graph allocation in a context
     LM_GGML_API struct lm_ggml_cgraph * lm_ggml_new_graph       (struct lm_ggml_context * ctx); // size = LM_GGML_DEFAULT_GRAPH_SIZE, grads = false
     LM_GGML_API struct lm_ggml_cgraph * lm_ggml_new_graph_custom(struct lm_ggml_context * ctx, size_t size, bool grads);
-    LM_GGML_API struct lm_ggml_cgraph * lm_ggml_graph_dup       (struct lm_ggml_context * ctx, struct lm_ggml_cgraph * cgraph);
+    LM_GGML_API struct lm_ggml_cgraph * lm_ggml_graph_dup       (struct lm_ggml_context * ctx, struct lm_ggml_cgraph * cgraph, bool force_grads);
     LM_GGML_API void                 lm_ggml_graph_cpy       (struct lm_ggml_cgraph * src, struct lm_ggml_cgraph * dst);
     LM_GGML_API void                 lm_ggml_graph_reset     (struct lm_ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
     LM_GGML_API void                 lm_ggml_graph_clear     (struct lm_ggml_cgraph * cgraph);

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

@@ -23,6 +23,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -58,6 +59,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -109,6 +111,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -143,6 +146,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -256,6 +261,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -303,6 +310,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,

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

@@ -70,7 +70,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

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

@@ -14,6 +14,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3,
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN,
     LLM_CHAT_TEMPLATE_PHI_3,
     LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,
@@ -29,8 +30,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
@@ -41,6 +42,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -7,6 +7,7 @@
 #include "llama-adapter.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
@@ -27,7 +28,12 @@ struct llama_context {
 
     void synchronize();
 
-    const llama_model & get_model() const;
+    const llama_model   & get_model()   const;
+    const llama_cparams & get_cparams() const;
+
+    lm_ggml_backend_sched_t get_sched() const;
+
+    lm_ggml_context * get_ctx_compute() const;
 
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
@@ -128,6 +134,32 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    //
+    // training
+    //
+
+    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
+
+    void opt_epoch(
+            lm_ggml_opt_dataset_t      dataset,
+            lm_ggml_opt_result_t       result_train,
+            lm_ggml_opt_result_t       result_eval,
+            int64_t                 idata_split,
+            lm_ggml_opt_epoch_callback callback_train,
+            lm_ggml_opt_epoch_callback callback_eval);
+
+    void opt_epoch_iter(
+            lm_ggml_opt_dataset_t               dataset,
+            lm_ggml_opt_result_t                result,
+            const std::vector<llama_token> & tokens,
+            const std::vector<llama_token> & labels_sparse,
+            llama_batch                    & batch,
+            lm_ggml_opt_epoch_callback          callback,
+            bool                             train,
+            int64_t                          idata_in_loop,
+            int64_t                          ndata_in_loop,
+            int64_t                          t_loop_start);
+
 private:
     //
     // output
@@ -137,50 +169,30 @@ private:
     // 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
     //
 
+public:
     int32_t graph_max_nodes() const;
 
     // zero-out inputs and create the ctx_compute for the compute graph
     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);
+
+private:
     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);
@@ -197,14 +209,10 @@ private:
     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;
+    std::unique_ptr<llama_memory_i> memory;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
@@ -231,6 +239,9 @@ private:
 
     lm_ggml_context_ptr ctx_compute;
 
+    // training
+    lm_ggml_opt_context_t opt_ctx = nullptr;
+
     lm_ggml_threadpool_t threadpool       = nullptr;
     lm_ggml_threadpool_t threadpool_batch = nullptr;
 

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

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool warmup;
+    bool op_offload;
 
     enum llama_pooling_type pooling_type;