whisper.rn 0.4.0-rc.8 → 0.4.0

package/cpp/ggml.h

@@ -176,25 +176,15 @@
 #ifdef WSP_GGML_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef WSP_GGML_BUILD
-#            define WSP_GGML_API __declspec(dllexport)
+#            define WSP_GGML_API __declspec(dllexport) extern
 #        else
-#            define WSP_GGML_API __declspec(dllimport)
+#            define WSP_GGML_API __declspec(dllimport) extern
 #        endif
 #    else
-#        define WSP_GGML_API __attribute__ ((visibility ("default")))
+#        define WSP_GGML_API __attribute__ ((visibility ("default"))) extern
 #    endif
 #else
-#    define WSP_GGML_API
-#endif
-
-#ifdef WSP_GGML_MULTIPLATFORM
-#    if defined(_WIN32)
-#        define WSP_GGML_CALL
-#    else
-#        define WSP_GGML_CALL __attribute__((__ms_abi__))
-#    endif
-#else
-#    define WSP_GGML_CALL
+#    define WSP_GGML_API extern
 #endif
 
 // TODO: support for clang
@@ -208,32 +198,36 @@
 
 #ifndef __GNUC__
 #    define WSP_GGML_ATTRIBUTE_FORMAT(...)
-#elif defined(__MINGW32__)
+#elif defined(__MINGW32__) && !defined(__clang__)
 #    define WSP_GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
 #else
 #    define WSP_GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#include <stdint.h>
-#include <stddef.h>
 #include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
 
 #define WSP_GGML_FILE_MAGIC   0x67676d6c // "ggml"
-#define WSP_GGML_FILE_VERSION 1
+#define WSP_GGML_FILE_VERSION 2
 
 #define WSP_GGML_QNT_VERSION        2    // bump this on quantization format changes
 #define WSP_GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define WSP_GGML_MAX_DIMS           4
 #define WSP_GGML_MAX_PARAMS         2048
-#define WSP_GGML_MAX_CONTEXTS       64
 #define WSP_GGML_MAX_SRC            10
+#define WSP_GGML_MAX_N_THREADS      512
+#define WSP_GGML_MAX_OP_PARAMS      64
+
 #ifndef WSP_GGML_MAX_NAME
-#define WSP_GGML_MAX_NAME           64
+#   define WSP_GGML_MAX_NAME        64
 #endif
-#define WSP_GGML_MAX_OP_PARAMS      64
+
 #define WSP_GGML_DEFAULT_N_THREADS  4
 #define WSP_GGML_DEFAULT_GRAPH_SIZE 2048
+
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define WSP_GGML_MEM_ALIGN 4
 #else
@@ -243,36 +237,35 @@
 #define WSP_GGML_EXIT_SUCCESS 0
 #define WSP_GGML_EXIT_ABORTED 1
 
-#define WSP_GGUF_MAGIC "GGUF"
-
-#define WSP_GGUF_VERSION 3
-
-#define WSP_GGUF_DEFAULT_ALIGNMENT 32
+#define WSP_GGML_ROPE_TYPE_NEOX   2
+#define WSP_GGML_ROPE_TYPE_MROPE  8
+#define WSP_GGML_ROPE_TYPE_VISION 24
 
 #define WSP_GGML_UNUSED(x) (void)(x)
 
 #define WSP_GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
-#define WSP_GGML_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fflush(stdout); \
-            fprintf(stderr, "WSP_GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            wsp_ggml_print_backtrace(); \
-            abort(); \
-        } \
-    } while (0)
-
 #ifndef NDEBUG
-#define WSP_GGML_UNREACHABLE() WSP_GGML_ASSERT(!"statement should not be reached")
+#   define WSP_GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
-#define WSP_GGML_UNREACHABLE() __builtin_unreachable()
+#   define WSP_GGML_UNREACHABLE() __builtin_unreachable()
+#elif defined(_MSC_VER)
+#   define WSP_GGML_UNREACHABLE() __assume(0)
+#else
+#   define WSP_GGML_UNREACHABLE() ((void) 0)
+#endif
+
+#ifdef __cplusplus
+#   define WSP_GGML_NORETURN [[noreturn]]
 #elif defined(_MSC_VER)
-#define WSP_GGML_UNREACHABLE() __assume(0)
+#   define WSP_GGML_NORETURN __declspec(noreturn)
 #else
-#define WSP_GGML_UNREACHABLE() ((void) 0)
+#   define WSP_GGML_NORETURN _Noreturn
 #endif
 
+#define WSP_GGML_ABORT(...) wsp_ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
+#define WSP_GGML_ASSERT(x) if (!(x)) WSP_GGML_ABORT("WSP_GGML_ASSERT(%s) failed", #x)
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -311,84 +304,125 @@
     WSP_GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
     WSP_GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
+#define WSP_GGML_TENSOR_BINARY_OP_LOCALS01 \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-#if defined(__ARM_NEON) && defined(__CUDACC__)
-    typedef half wsp_ggml_fp16_t;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-    typedef __fp16 wsp_ggml_fp16_t;
-#else
-    typedef uint16_t wsp_ggml_fp16_t;
-#endif
+    WSP_GGML_NORETURN WSP_GGML_ATTRIBUTE_FORMAT(3, 4)
+    WSP_GGML_API void wsp_ggml_abort(const char * file, int line, const char * fmt, ...);
+
+    enum wsp_ggml_status {
+        WSP_GGML_STATUS_ALLOC_FAILED = -2,
+        WSP_GGML_STATUS_FAILED = -1,
+        WSP_GGML_STATUS_SUCCESS = 0,
+        WSP_GGML_STATUS_ABORTED = 1,
+    };
 
-    // convert FP16 <-> FP32
-    WSP_GGML_API float       wsp_ggml_fp16_to_fp32(wsp_ggml_fp16_t x);
-    WSP_GGML_API wsp_ggml_fp16_t wsp_ggml_fp32_to_fp16(float x);
+    // get wsp_ggml_status name string
+    WSP_GGML_API const char * wsp_ggml_status_to_string(enum wsp_ggml_status status);
 
-    WSP_GGML_API void wsp_ggml_fp16_to_fp32_row(const wsp_ggml_fp16_t * x, float * y, int n);
-    WSP_GGML_API void wsp_ggml_fp32_to_fp16_row(const float * x, wsp_ggml_fp16_t * y, int n);
+    // ieee 754-2008 half-precision float16
+    // todo: make this not an integral type
+    typedef uint16_t wsp_ggml_fp16_t;
+    WSP_GGML_API float       wsp_ggml_fp16_to_fp32(wsp_ggml_fp16_t);
+    WSP_GGML_API wsp_ggml_fp16_t wsp_ggml_fp32_to_fp16(float);
+    WSP_GGML_API void        wsp_ggml_fp16_to_fp32_row(const wsp_ggml_fp16_t *, float *, int64_t);
+    WSP_GGML_API void        wsp_ggml_fp32_to_fp16_row(const float *, wsp_ggml_fp16_t *, int64_t);
+
+    // google brain half-precision bfloat16
+    typedef struct { uint16_t bits; } wsp_ggml_bf16_t;
+    WSP_GGML_API wsp_ggml_bf16_t wsp_ggml_fp32_to_bf16(float);
+    WSP_GGML_API float       wsp_ggml_bf16_to_fp32(wsp_ggml_bf16_t);  // consider just doing << 16
+    WSP_GGML_API void        wsp_ggml_bf16_to_fp32_row(const wsp_ggml_bf16_t *, float *, int64_t);
+    WSP_GGML_API void        wsp_ggml_fp32_to_bf16_row_ref(const float *, wsp_ggml_bf16_t *, int64_t);
+    WSP_GGML_API void        wsp_ggml_fp32_to_bf16_row(const float *, wsp_ggml_bf16_t *, int64_t);
 
     struct wsp_ggml_object;
     struct wsp_ggml_context;
+    struct wsp_ggml_cgraph;
 
+    // NOTE: always add types at the end of the enum to keep backward compatibility
     enum wsp_ggml_type {
-        WSP_GGML_TYPE_F32  = 0,
-        WSP_GGML_TYPE_F16  = 1,
-        WSP_GGML_TYPE_Q4_0 = 2,
-        WSP_GGML_TYPE_Q4_1 = 3,
+        WSP_GGML_TYPE_F32     = 0,
+        WSP_GGML_TYPE_F16     = 1,
+        WSP_GGML_TYPE_Q4_0    = 2,
+        WSP_GGML_TYPE_Q4_1    = 3,
         // WSP_GGML_TYPE_Q4_2 = 4, support has been removed
-        // WSP_GGML_TYPE_Q4_3 (5) support has been removed
-        WSP_GGML_TYPE_Q5_0 = 6,
-        WSP_GGML_TYPE_Q5_1 = 7,
-        WSP_GGML_TYPE_Q8_0 = 8,
-        WSP_GGML_TYPE_Q8_1 = 9,
-        // k-quantizations
-        WSP_GGML_TYPE_Q2_K = 10,
-        WSP_GGML_TYPE_Q3_K = 11,
-        WSP_GGML_TYPE_Q4_K = 12,
-        WSP_GGML_TYPE_Q5_K = 13,
-        WSP_GGML_TYPE_Q6_K = 14,
-        WSP_GGML_TYPE_Q8_K = 15,
+        // WSP_GGML_TYPE_Q4_3 = 5, support has been removed
+        WSP_GGML_TYPE_Q5_0    = 6,
+        WSP_GGML_TYPE_Q5_1    = 7,
+        WSP_GGML_TYPE_Q8_0    = 8,
+        WSP_GGML_TYPE_Q8_1    = 9,
+        WSP_GGML_TYPE_Q2_K    = 10,
+        WSP_GGML_TYPE_Q3_K    = 11,
+        WSP_GGML_TYPE_Q4_K    = 12,
+        WSP_GGML_TYPE_Q5_K    = 13,
+        WSP_GGML_TYPE_Q6_K    = 14,
+        WSP_GGML_TYPE_Q8_K    = 15,
         WSP_GGML_TYPE_IQ2_XXS = 16,
         WSP_GGML_TYPE_IQ2_XS  = 17,
-        WSP_GGML_TYPE_I8,
-        WSP_GGML_TYPE_I16,
-        WSP_GGML_TYPE_I32,
-        WSP_GGML_TYPE_COUNT,
+        WSP_GGML_TYPE_IQ3_XXS = 18,
+        WSP_GGML_TYPE_IQ1_S   = 19,
+        WSP_GGML_TYPE_IQ4_NL  = 20,
+        WSP_GGML_TYPE_IQ3_S   = 21,
+        WSP_GGML_TYPE_IQ2_S   = 22,
+        WSP_GGML_TYPE_IQ4_XS  = 23,
+        WSP_GGML_TYPE_I8      = 24,
+        WSP_GGML_TYPE_I16     = 25,
+        WSP_GGML_TYPE_I32     = 26,
+        WSP_GGML_TYPE_I64     = 27,
+        WSP_GGML_TYPE_F64     = 28,
+        WSP_GGML_TYPE_IQ1_M   = 29,
+        WSP_GGML_TYPE_BF16    = 30,
+        // WSP_GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
+        // WSP_GGML_TYPE_Q4_0_4_8 = 32,
+        // WSP_GGML_TYPE_Q4_0_8_8 = 33,
+        WSP_GGML_TYPE_TQ1_0   = 34,
+        WSP_GGML_TYPE_TQ2_0   = 35,
+        // WSP_GGML_TYPE_IQ4_NL_4_4 = 36,
+        // WSP_GGML_TYPE_IQ4_NL_4_8 = 37,
+        // WSP_GGML_TYPE_IQ4_NL_8_8 = 38,
+        WSP_GGML_TYPE_COUNT   = 39,
     };
 
     // precision
     enum wsp_ggml_prec {
-        WSP_GGML_PREC_DEFAULT,
-        WSP_GGML_PREC_F32,
-    };
-
-    enum wsp_ggml_backend_type {
-        WSP_GGML_BACKEND_CPU = 0,
-        WSP_GGML_BACKEND_GPU = 10,
-        WSP_GGML_BACKEND_GPU_SPLIT = 20,
+        WSP_GGML_PREC_DEFAULT =  0, // stored as wsp_ggml_tensor.op_params, 0 by default
+        WSP_GGML_PREC_F32     = 10,
     };
 
     // model file types
     enum wsp_ggml_ftype {
-        WSP_GGML_FTYPE_UNKNOWN     = -1,
-        WSP_GGML_FTYPE_ALL_F32     = 0,
-        WSP_GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        WSP_GGML_FTYPE_UNKNOWN        = -1,
+        WSP_GGML_FTYPE_ALL_F32        = 0,
+        WSP_GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
         WSP_GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        WSP_GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
-        WSP_GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
         WSP_GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         WSP_GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ1_S   = 18, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ4_NL  = 19, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ3_S   = 20, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
     };
 
     // available tensor operations:
@@ -405,10 +439,13 @@ extern "C" {
         WSP_GGML_OP_SQR,
         WSP_GGML_OP_SQRT,
         WSP_GGML_OP_LOG,
+        WSP_GGML_OP_SIN,
+        WSP_GGML_OP_COS,
         WSP_GGML_OP_SUM,
         WSP_GGML_OP_SUM_ROWS,
         WSP_GGML_OP_MEAN,
         WSP_GGML_OP_ARGMAX,
+        WSP_GGML_OP_COUNT_EQUAL,
         WSP_GGML_OP_REPEAT,
         WSP_GGML_OP_REPEAT_BACK,
         WSP_GGML_OP_CONCAT,
@@ -417,6 +454,7 @@ extern "C" {
         WSP_GGML_OP_RMS_NORM,
         WSP_GGML_OP_RMS_NORM_BACK,
         WSP_GGML_OP_GROUP_NORM,
+        WSP_GGML_OP_L2_NORM,
 
         WSP_GGML_OP_MUL_MAT,
         WSP_GGML_OP_MUL_MAT_ID,
@@ -439,41 +477,47 @@ extern "C" {
         WSP_GGML_OP_SOFT_MAX_BACK,
         WSP_GGML_OP_ROPE,
         WSP_GGML_OP_ROPE_BACK,
-        WSP_GGML_OP_ALIBI,
         WSP_GGML_OP_CLAMP,
         WSP_GGML_OP_CONV_TRANSPOSE_1D,
         WSP_GGML_OP_IM2COL,
+        WSP_GGML_OP_IM2COL_BACK,
+        WSP_GGML_OP_CONV_2D_DW,
         WSP_GGML_OP_CONV_TRANSPOSE_2D,
         WSP_GGML_OP_POOL_1D,
         WSP_GGML_OP_POOL_2D,
+        WSP_GGML_OP_POOL_2D_BACK,
         WSP_GGML_OP_UPSCALE, // nearest interpolate
         WSP_GGML_OP_PAD,
+        WSP_GGML_OP_PAD_REFLECT_1D,
+        WSP_GGML_OP_ROLL,
+        WSP_GGML_OP_ARANGE,
+        WSP_GGML_OP_TIMESTEP_EMBEDDING,
         WSP_GGML_OP_ARGSORT,
         WSP_GGML_OP_LEAKY_RELU,
 
-        WSP_GGML_OP_FLASH_ATTN,
-        WSP_GGML_OP_FLASH_FF,
+        WSP_GGML_OP_FLASH_ATTN_EXT,
         WSP_GGML_OP_FLASH_ATTN_BACK,
+        WSP_GGML_OP_SSM_CONV,
+        WSP_GGML_OP_SSM_SCAN,
         WSP_GGML_OP_WIN_PART,
         WSP_GGML_OP_WIN_UNPART,
         WSP_GGML_OP_GET_REL_POS,
         WSP_GGML_OP_ADD_REL_POS,
+        WSP_GGML_OP_RWKV_WKV6,
+        WSP_GGML_OP_GATED_LINEAR_ATTN,
+        WSP_GGML_OP_RWKV_WKV7,
 
         WSP_GGML_OP_UNARY,
 
-        WSP_GGML_OP_MAP_UNARY,
-        WSP_GGML_OP_MAP_BINARY,
-
-        WSP_GGML_OP_MAP_CUSTOM1_F32,
-        WSP_GGML_OP_MAP_CUSTOM2_F32,
-        WSP_GGML_OP_MAP_CUSTOM3_F32,
-
         WSP_GGML_OP_MAP_CUSTOM1,
         WSP_GGML_OP_MAP_CUSTOM2,
         WSP_GGML_OP_MAP_CUSTOM3,
 
+        WSP_GGML_OP_CUSTOM,
+
         WSP_GGML_OP_CROSS_ENTROPY_LOSS,
         WSP_GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+        WSP_GGML_OP_OPT_STEP_ADAMW,
 
         WSP_GGML_OP_COUNT,
     };
@@ -486,44 +530,51 @@ extern "C" {
         WSP_GGML_UNARY_OP_TANH,
         WSP_GGML_UNARY_OP_ELU,
         WSP_GGML_UNARY_OP_RELU,
+        WSP_GGML_UNARY_OP_SIGMOID,
         WSP_GGML_UNARY_OP_GELU,
         WSP_GGML_UNARY_OP_GELU_QUICK,
         WSP_GGML_UNARY_OP_SILU,
+        WSP_GGML_UNARY_OP_HARDSWISH,
+        WSP_GGML_UNARY_OP_HARDSIGMOID,
+        WSP_GGML_UNARY_OP_EXP,
+        WSP_GGML_UNARY_OP_GELU_ERF,
 
         WSP_GGML_UNARY_OP_COUNT,
     };
 
     enum wsp_ggml_object_type {
-        WSP_GGML_OBJECT_TENSOR,
-        WSP_GGML_OBJECT_GRAPH,
-        WSP_GGML_OBJECT_WORK_BUFFER
+        WSP_GGML_OBJECT_TYPE_TENSOR,
+        WSP_GGML_OBJECT_TYPE_GRAPH,
+        WSP_GGML_OBJECT_TYPE_WORK_BUFFER
     };
 
     enum wsp_ggml_log_level {
-        WSP_GGML_LOG_LEVEL_ERROR = 2,
-        WSP_GGML_LOG_LEVEL_WARN = 3,
-        WSP_GGML_LOG_LEVEL_INFO = 4,
-        WSP_GGML_LOG_LEVEL_DEBUG = 5
+        WSP_GGML_LOG_LEVEL_NONE  = 0,
+        WSP_GGML_LOG_LEVEL_DEBUG = 1,
+        WSP_GGML_LOG_LEVEL_INFO  = 2,
+        WSP_GGML_LOG_LEVEL_WARN  = 3,
+        WSP_GGML_LOG_LEVEL_ERROR = 4,
+        WSP_GGML_LOG_LEVEL_CONT  = 5, // continue previous log
     };
 
-    // ggml object
-    struct wsp_ggml_object {
-        size_t offs;
-        size_t size;
-
-        struct wsp_ggml_object * next;
-
-        enum wsp_ggml_object_type type;
-
-        char padding[4];
+    // this tensor...
+    enum wsp_ggml_tensor_flag {
+        WSP_GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
+        WSP_GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
+        WSP_GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
+        WSP_GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
-    static const size_t WSP_GGML_OBJECT_SIZE = sizeof(struct wsp_ggml_object);
+    struct wsp_ggml_init_params {
+        // memory pool
+        size_t mem_size;   // bytes
+        void * mem_buffer; // if NULL, memory will be allocated internally
+        bool   no_alloc;   // don't allocate memory for the tensor data
+    };
 
     // n-dimensional tensor
     struct wsp_ggml_tensor {
-        enum wsp_ggml_type         type;
-        enum wsp_ggml_backend_type backend;
+        enum wsp_ggml_type type;
 
         struct wsp_ggml_backend_buffer * buffer;
 
@@ -539,16 +590,11 @@ extern "C" {
         // op params - allocated as int32_t for alignment
         int32_t op_params[WSP_GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 
-        bool is_param;
+        int32_t flags;
 
-        struct wsp_ggml_tensor * grad;
         struct wsp_ggml_tensor * src[WSP_GGML_MAX_SRC];
 
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
-
+        // source tensor and offset for views
         struct wsp_ggml_tensor * view_src;
         size_t               view_offs;
 
@@ -563,85 +609,21 @@ extern "C" {
 
     static const size_t WSP_GGML_TENSOR_SIZE = sizeof(struct wsp_ggml_tensor);
 
-    // the compute plan that needs to be prepared for wsp_ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
-    struct wsp_ggml_cplan {
-        size_t    work_size; // size of work buffer, calculated by `wsp_ggml_graph_plan()`
-        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `wsp_ggml_graph_compute()`
+    // Abort callback
+    // If not NULL, called before ggml computation
+    // If it returns true, the computation is aborted
+    typedef bool (*wsp_ggml_abort_callback)(void * data);
 
-        int n_threads;
 
-        // abort wsp_ggml_graph_compute when true
-        bool (*abort_callback)(void * data);
-        void * abort_callback_data;
-    };
-
-    enum wsp_ggml_cgraph_eval_order {
-        WSP_GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-        WSP_GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-        WSP_GGML_CGRAPH_EVAL_ORDER_COUNT
-    };
-
-    struct wsp_ggml_hash_set {
-        size_t size;
-        struct wsp_ggml_tensor ** keys;
-    };
-
-    // computation graph
-    struct wsp_ggml_cgraph {
-        int size;
-        int n_nodes;
-        int n_leafs;
-
-        struct wsp_ggml_tensor ** nodes;
-        struct wsp_ggml_tensor ** grads;
-        struct wsp_ggml_tensor ** leafs;
-
-        struct wsp_ggml_hash_set visited_hash_table;
-
-        enum wsp_ggml_cgraph_eval_order order;
-
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
-    };
-
-    // scratch buffer
-    struct wsp_ggml_scratch {
-        size_t offs;
-        size_t size;
-        void * data;
-    };
-
-    struct wsp_ggml_init_params {
-        // memory pool
-        size_t mem_size;   // bytes
-        void * mem_buffer; // if NULL, memory will be allocated internally
-        bool   no_alloc;   // don't allocate memory for the tensor data
-    };
-
-
-    // compute types
-
-    // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
-    // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
-    enum wsp_ggml_task_type {
-        WSP_GGML_TASK_INIT = 0,
-        WSP_GGML_TASK_COMPUTE,
-        WSP_GGML_TASK_FINALIZE,
-    };
-
-    struct wsp_ggml_compute_params {
-        enum wsp_ggml_task_type type;
+    //
+    // GUID
+    //
 
-        // ith = thread index, nth = number of threads
-        int ith, nth;
+    // GUID types
+    typedef uint8_t wsp_ggml_guid[16];
+    typedef wsp_ggml_guid * wsp_ggml_guid_t;
 
-        // work buffer for all threads
-        size_t wsize;
-        void * wdata;
-    };
+    WSP_GGML_API bool wsp_ggml_guid_matches(wsp_ggml_guid_t guid_a, wsp_ggml_guid_t guid_b);
 
     // misc
 
@@ -651,63 +633,78 @@ extern "C" {
     WSP_GGML_API int64_t wsp_ggml_cycles(void);
     WSP_GGML_API int64_t wsp_ggml_cycles_per_ms(void);
 
-    WSP_GGML_API void    wsp_ggml_print_backtrace(void);
-
-    WSP_GGML_API void    wsp_ggml_numa_init(void); // call once for better performance on NUMA systems
-    WSP_GGML_API bool    wsp_ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+    // accepts a UTF-8 path, even on Windows
+    WSP_GGML_API FILE *  wsp_ggml_fopen(const char * fname, const char * mode);
 
     WSP_GGML_API void    wsp_ggml_print_object (const struct wsp_ggml_object * obj);
     WSP_GGML_API void    wsp_ggml_print_objects(const struct wsp_ggml_context * ctx);
 
-    WSP_GGML_API WSP_GGML_CALL int64_t wsp_ggml_nelements   (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API WSP_GGML_CALL int64_t wsp_ggml_nrows       (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API WSP_GGML_CALL size_t  wsp_ggml_nbytes      (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           size_t  wsp_ggml_nbytes_pad  (const struct wsp_ggml_tensor * tensor); // same as wsp_ggml_nbytes() but padded to WSP_GGML_MEM_ALIGN
+    WSP_GGML_API int64_t wsp_ggml_nelements (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API int64_t wsp_ggml_nrows     (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API size_t  wsp_ggml_nbytes    (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API size_t  wsp_ggml_nbytes_pad(const struct wsp_ggml_tensor * tensor); // same as wsp_ggml_nbytes() but padded to WSP_GGML_MEM_ALIGN
 
-    WSP_GGML_API WSP_GGML_CALL int    wsp_ggml_blck_size(enum wsp_ggml_type type);
-    WSP_GGML_API WSP_GGML_CALL size_t wsp_ggml_type_size(enum wsp_ggml_type type);             // size in bytes for all elements in a block
-    WSP_GGML_API WSP_GGML_CALL size_t wsp_ggml_row_size (enum wsp_ggml_type type, int64_t ne); // size in bytes for all elements in a row
+    WSP_GGML_API int64_t wsp_ggml_blck_size(enum wsp_ggml_type type);
+    WSP_GGML_API size_t  wsp_ggml_type_size(enum wsp_ggml_type type);             // size in bytes for all elements in a block
+    WSP_GGML_API size_t  wsp_ggml_row_size (enum wsp_ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
     WSP_GGML_DEPRECATED(
     WSP_GGML_API double wsp_ggml_type_sizef(enum wsp_ggml_type type), // wsp_ggml_type_size()/wsp_ggml_blck_size() as float
     "use wsp_ggml_row_size() instead");
 
-    WSP_GGML_API WSP_GGML_CALL const char * wsp_ggml_type_name(enum wsp_ggml_type type);
-    WSP_GGML_API WSP_GGML_CALL const char * wsp_ggml_op_name  (enum wsp_ggml_op   op);
-    WSP_GGML_API           const char * wsp_ggml_op_symbol(enum wsp_ggml_op   op);
+    WSP_GGML_API const char * wsp_ggml_type_name(enum wsp_ggml_type type);
+    WSP_GGML_API const char * wsp_ggml_op_name  (enum wsp_ggml_op   op);
+    WSP_GGML_API const char * wsp_ggml_op_symbol(enum wsp_ggml_op   op);
 
-    WSP_GGML_API           const char * wsp_ggml_unary_op_name(enum wsp_ggml_unary_op op);
-    WSP_GGML_API WSP_GGML_CALL const char * wsp_ggml_op_desc(const struct wsp_ggml_tensor * t); // unary or op name
+    WSP_GGML_API const char * wsp_ggml_unary_op_name(enum wsp_ggml_unary_op op);
+    WSP_GGML_API const char * wsp_ggml_op_desc(const struct wsp_ggml_tensor * t); // unary or op name
 
-    WSP_GGML_API WSP_GGML_CALL size_t  wsp_ggml_element_size(const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API size_t  wsp_ggml_element_size(const struct wsp_ggml_tensor * tensor);
 
-    WSP_GGML_API WSP_GGML_CALL bool    wsp_ggml_is_quantized(enum wsp_ggml_type type);
+    WSP_GGML_API bool    wsp_ggml_is_quantized(enum wsp_ggml_type type);
 
     // TODO: temporary until model loading of ggml examples is refactored
     WSP_GGML_API enum wsp_ggml_type wsp_ggml_ftype_to_wsp_ggml_type(enum wsp_ggml_ftype ftype);
 
-    WSP_GGML_API WSP_GGML_CALL bool wsp_ggml_is_transposed(const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API WSP_GGML_CALL bool wsp_ggml_is_contiguous(const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API WSP_GGML_CALL bool wsp_ggml_is_permuted  (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           bool wsp_ggml_is_scalar    (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           bool wsp_ggml_is_vector    (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           bool wsp_ggml_is_matrix    (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           bool wsp_ggml_is_3d        (const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API           int  wsp_ggml_n_dims       (const struct wsp_ggml_tensor * tensor); // returns 1 for scalars
+    WSP_GGML_API bool wsp_ggml_is_transposed(const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_permuted  (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_empty     (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_scalar    (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_vector    (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_matrix    (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_3d        (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API int  wsp_ggml_n_dims       (const struct wsp_ggml_tensor * tensor); // returns 1 for scalars
 
-    WSP_GGML_API bool wsp_ggml_are_same_shape(const struct wsp_ggml_tensor * t0, const struct wsp_ggml_tensor * t1);
+    // returns whether the tensor elements can be iterated over with a flattened index (no gaps, no permutation)
+    WSP_GGML_API bool wsp_ggml_is_contiguous  (const struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API bool wsp_ggml_is_contiguous_0(const struct wsp_ggml_tensor * tensor); // same as wsp_ggml_is_contiguous()
+    WSP_GGML_API bool wsp_ggml_is_contiguous_1(const struct wsp_ggml_tensor * tensor); // contiguous for dims >= 1
+    WSP_GGML_API bool wsp_ggml_is_contiguous_2(const struct wsp_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)
+    WSP_GGML_API bool wsp_ggml_is_contiguously_allocated(const struct wsp_ggml_tensor * tensor);
+
+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    WSP_GGML_API bool wsp_ggml_is_contiguous_channels(const struct wsp_ggml_tensor * tensor);
+
+    WSP_GGML_API bool wsp_ggml_are_same_shape (const struct wsp_ggml_tensor * t0, const struct wsp_ggml_tensor * t1);
+    WSP_GGML_API bool wsp_ggml_are_same_stride(const struct wsp_ggml_tensor * t0, const struct wsp_ggml_tensor * t1);
+
+    WSP_GGML_API bool wsp_ggml_can_repeat(const struct wsp_ggml_tensor * t0, const struct wsp_ggml_tensor * t1);
 
     // use this to compute the memory overhead of a tensor
     WSP_GGML_API size_t wsp_ggml_tensor_overhead(void);
 
+    WSP_GGML_API bool wsp_ggml_validate_row_data(enum wsp_ggml_type type, const void * data, size_t nbytes);
+
     // main
 
-    WSP_GGML_API struct wsp_ggml_context * wsp_ggml_init(struct wsp_ggml_init_params params);
-    WSP_GGML_API void                  wsp_ggml_free(struct wsp_ggml_context * ctx);
+    WSP_GGML_API struct wsp_ggml_context * wsp_ggml_init (struct wsp_ggml_init_params params);
+    WSP_GGML_API void                  wsp_ggml_reset(struct wsp_ggml_context * ctx);
+    WSP_GGML_API void                  wsp_ggml_free (struct wsp_ggml_context * ctx);
 
     WSP_GGML_API size_t  wsp_ggml_used_mem(const struct wsp_ggml_context * ctx);
 
-    WSP_GGML_API size_t  wsp_ggml_set_scratch (struct wsp_ggml_context * ctx, struct wsp_ggml_scratch scratch);
     WSP_GGML_API bool    wsp_ggml_get_no_alloc(struct wsp_ggml_context * ctx);
     WSP_GGML_API void    wsp_ggml_set_no_alloc(struct wsp_ggml_context * ctx, bool no_alloc);
 
@@ -747,8 +744,7 @@ extern "C" {
             int64_t ne2,
             int64_t ne3);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_new_i32(struct wsp_ggml_context * ctx, int32_t value);
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_new_f32(struct wsp_ggml_context * ctx, float value);
+    WSP_GGML_API void * wsp_ggml_new_buffer(struct wsp_ggml_context * ctx, size_t nbytes);
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_dup_tensor (struct wsp_ggml_context * ctx, const struct wsp_ggml_tensor * src);
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_view_tensor(struct wsp_ggml_context * ctx, struct wsp_ggml_tensor * src);
@@ -758,35 +754,25 @@ extern "C" {
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_get_next_tensor (const struct wsp_ggml_context * ctx, struct wsp_ggml_tensor * tensor);
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_get_tensor(struct wsp_ggml_context * ctx, const char * name);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_zero(struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_i32 (struct wsp_ggml_tensor * tensor, int32_t value);
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_f32 (struct wsp_ggml_tensor * tensor, float value);
-
     // Converts a flat index into coordinates
-    WSP_GGML_API void    wsp_ggml_unravel_index(const struct wsp_ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
-
-    WSP_GGML_API int32_t wsp_ggml_get_i32_1d(const struct wsp_ggml_tensor * tensor, int i);
-    WSP_GGML_API void    wsp_ggml_set_i32_1d(const struct wsp_ggml_tensor * tensor, int i, int32_t value);
-
-    WSP_GGML_API int32_t wsp_ggml_get_i32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    WSP_GGML_API void    wsp_ggml_set_i32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+    WSP_GGML_API void wsp_ggml_unravel_index(const struct wsp_ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
-    WSP_GGML_API float   wsp_ggml_get_f32_1d(const struct wsp_ggml_tensor * tensor, int i);
-    WSP_GGML_API void    wsp_ggml_set_f32_1d(const struct wsp_ggml_tensor * tensor, int i, float value);
-
-    WSP_GGML_API float   wsp_ggml_get_f32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    WSP_GGML_API void    wsp_ggml_set_f32_nd(const struct wsp_ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    WSP_GGML_API enum wsp_ggml_unary_op wsp_ggml_get_unary_op(const struct wsp_ggml_tensor * tensor);
 
     WSP_GGML_API void *  wsp_ggml_get_data    (const struct wsp_ggml_tensor * tensor);
     WSP_GGML_API float * wsp_ggml_get_data_f32(const struct wsp_ggml_tensor * tensor);
 
-    WSP_GGML_API WSP_GGML_CALL enum wsp_ggml_unary_op wsp_ggml_get_unary_op(const struct wsp_ggml_tensor * tensor);
-
     WSP_GGML_API const char *         wsp_ggml_get_name   (const struct wsp_ggml_tensor * tensor);
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_name   (      struct wsp_ggml_tensor * tensor, const char * name);
     WSP_GGML_ATTRIBUTE_FORMAT(2, 3)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_format_name(      struct wsp_ggml_tensor * tensor, const char * fmt, ...);
 
+    // Tensor flags
+    WSP_GGML_API void wsp_ggml_set_input(struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API void wsp_ggml_set_output(struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API void wsp_ggml_set_param(struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API void wsp_ggml_set_loss(struct wsp_ggml_tensor * tensor);
+
     //
     // operations on tensors with backpropagation
     //
@@ -901,6 +887,22 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_sin(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_sin_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cos(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cos_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
     // return scalar
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_sum(
             struct wsp_ggml_context * ctx,
@@ -921,6 +923,12 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    // count number of equal elements in a and b
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_count_equal(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b);
+
     // if a is the same shape as b, and a is not parameter, return a
     // otherwise, return a new tensor: repeat(a) to fit in b
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_repeat(
@@ -928,18 +936,28 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b);
 
+    // repeat a to the specified shape
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_repeat_4d(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+                       int64_t    ne0,
+                       int64_t    ne1,
+                       int64_t    ne2,
+                       int64_t    ne3);
+
     // sums repetitions in a into shape of b
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_repeat_back(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
-    // concat a and b on dim 2
+    // concat a and b along dim
     // used in stable-diffusion
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_concat(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * b,
+            int                   dim);
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_abs(
             struct wsp_ggml_context * ctx,
@@ -1001,6 +1019,14 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_sigmoid(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_sigmoid_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_gelu(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
@@ -1009,6 +1035,16 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    // GELU using erf (error function) when possible
+    // some backends may fallback to approximation based on Abramowitz and Stegun formula
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_gelu_erf(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_gelu_erf_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_gelu_quick(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
@@ -1032,6 +1068,24 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b);
 
+    // hardswish(x) = x * relu6(x + 3) / 6
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_hardswish(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    // hardsigmoid(x) = relu6(x + 3) / 6
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_hardsigmoid(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_exp(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_exp_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
     // normalize along rows
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_norm(
             struct wsp_ggml_context * ctx,
@@ -1055,16 +1109,29 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_group_norm(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_group_norm_inplace(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
+
+    // l2 normalize along rows
+    // used in rwkv v7
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_l2_norm(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            float                 eps);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_l2_norm_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1089,14 +1156,11 @@ extern "C" {
             enum wsp_ggml_prec       prec);
 
     // indirect matrix multiplication
-    //  wsp_ggml_mul_mat_id(ctx, as, ids, id, b) ~= wsp_ggml_mul_mat(as[ids[id]], b)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_mul_mat_id(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * const as[],
-            int                   n_as,
-            struct wsp_ggml_tensor  * ids,
-            int                   id,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * as,
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * ids);
 
     // A: m columns, n rows,
     // B: p columns, n rows,
@@ -1129,7 +1193,7 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_inplace(
@@ -1139,19 +1203,19 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_1d(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_1d_inplace(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_2d(
@@ -1159,7 +1223,7 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_2d_inplace(
@@ -1167,7 +1231,7 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // a -> b, return view(b)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cpy(
@@ -1302,14 +1366,14 @@ extern "C" {
     // supports 3D: a->ne[2] == b->ne[1]
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_get_rows(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * a,  // data
+            struct wsp_ggml_tensor  * b); // row indices
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_get_rows_back(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
-            struct wsp_ggml_tensor  * c);
+            struct wsp_ggml_tensor  * a,  // gradients of wsp_ggml_get_rows result
+            struct wsp_ggml_tensor  * b,  // row indices
+            struct wsp_ggml_tensor  * c); // data for wsp_ggml_get_rows, only used for its shape
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_diag(
         struct wsp_ggml_context     * ctx,
@@ -1348,29 +1412,34 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask)
+    // fused soft_max(a*scale + mask*(ALiBi slope))
     // mask is optional
+    // max_bias = 0.0f for no ALiBi
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * mask,
-            float                 scale);
+            float                 scale,
+            float                 max_bias);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_back(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_ext_back(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * b,
+            float                 scale,
+            float                 max_bias);
 
     // in-place, returns view(a)
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_back_inplace(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_ext_back_inplace(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b);
+            struct wsp_ggml_tensor  * b,
+            float                 scale,
+            float                 max_bias);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
-    // if mode & 2 == 1, GPT-NeoX style
-    // if mode & 4 == 1, ChatGLM style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & WSP_GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope(
@@ -1378,8 +1447,7 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // in-place, returns view(a)
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_inplace(
@@ -1387,18 +1455,34 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // custom RoPE
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_custom(
+    // c is freq factors (e.g. phi3-128k), (optional)
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_multi(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1407,14 +1491,14 @@ extern "C" {
             float                 beta_slow);
 
     // in-place, returns view(a)
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_custom_inplace(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_ext_inplace(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1422,46 +1506,73 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
-    // compute correction dims for YaRN RoPE scaling
-    WSP_GGML_CALL void wsp_ggml_rope_yarn_corr_dims(
-        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
+    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_custom(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use wsp_ggml_rope_ext instead");
 
-    // xPos RoPE, in-place, returns view(a)
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_xpos_inplace(
+    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_custom_inplace(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b,
             int                   n_dims,
-            float                 base,
-            bool                  down);
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use wsp_ggml_rope_ext_inplace instead");
+
+    // compute correction dims for YaRN RoPE scaling
+    WSP_GGML_API void wsp_ggml_rope_yarn_corr_dims(
+        int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_back(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_ext_back(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * a, // gradients of wsp_ggml_rope result
+            struct wsp_ggml_tensor  * b, // positions
+            struct wsp_ggml_tensor  * c, // freq factors
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
             float                 attn_factor,
             float                 beta_fast,
-            float                 beta_slow,
-            float                 xpos_base,
-            bool                  xpos_down);
+            float                 beta_slow);
 
-    // alibi position embedding
-    // in-place, returns view(a)
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_alibi(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_multi_back(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            int                   n_past,
-            int                   n_head,
-            float                 bias_max);
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
 
     // clamp
     // in-place, returns view(a)
@@ -1471,22 +1582,38 @@ extern "C" {
             float                 min,
             float                 max);
 
+    // im2col
+    // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_im2col(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
-            int                  s0,
-            int                  s1,
-            int                  p0,
-            int                  p1,
-            int                  d0,
-            int                  d1,
-            bool                 is_2D);
+            struct wsp_ggml_tensor  * a,  // convolution kernel
+            struct wsp_ggml_tensor  * b,  // data
+            int                   s0, // stride dimension 0
+            int                   s1, // stride dimension 1
+            int                   p0, // padding dimension 0
+            int                   p1, // padding dimension 1
+            int                   d0, // dilation dimension 0
+            int                   d1, // dilation dimension 1
+            bool                  is_2D,
+            enum wsp_ggml_type        dst_type);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_im2col_back(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,  // convolution kernel
+        struct wsp_ggml_tensor  * b,  // gradient of im2col output
+        int64_t             * ne, // shape of im2col input
+        int                   s0, // stride dimension 0
+        int                   s1, // stride dimension 1
+        int                   p0, // padding dimension 0
+        int                   p1, // padding dimension 1
+        int                   d0, // dilation dimension 0
+        int                   d1, // dilation dimension 1
+        bool                  is_2D);
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * a,   // convolution kernel
+            struct wsp_ggml_tensor  * b,   // data
             int                   s0,  // stride
             int                   p0,  // padding
             int                   d0); // dilation
@@ -1495,30 +1622,46 @@ extern "C" {
     // alias for wsp_ggml_conv_1d(a, b, s, a->ne[0]/2, d)
     WSP_GGML_API struct wsp_ggml_tensor* wsp_ggml_conv_1d_ph(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
-            int                   s,
-            int                   d);
+            struct wsp_ggml_tensor  * a,  // convolution kernel
+            struct wsp_ggml_tensor  * b,  // data
+            int                   s,  // stride
+            int                   d); // dilation
+
+    // depthwise
+    // TODO: this is very likely wrong for some cases! - needs more testing
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d_dw(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,   // convolution kernel
+            struct wsp_ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d_dw_ph(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,   // convolution kernel
+            struct wsp_ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   d0); // dilation
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_transpose_1d(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
-            int                   s0,
-            int                   p0,
-            int                   d0);
+            struct wsp_ggml_tensor  * a,   // convolution kernel
+            struct wsp_ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_2d(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b,
-            int                   s0,
-            int                   s1,
-            int                   p0,
-            int                   p1,
-            int                   d0,
-            int                   d1);
-
+            struct wsp_ggml_tensor  * a,   // convolution kernel
+            struct wsp_ggml_tensor  * b,   // data
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
 
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
@@ -1546,6 +1689,34 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b);
 
+    // depthwise (via im2col and mul_mat)
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_2d_dw(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,  // convolution kernel
+            struct wsp_ggml_tensor  * b,  // data
+            int                  s0,  // stride dimension 0
+            int                  s1,  // stride dimension 1
+            int                  p0,  // padding dimension 0
+            int                  p1,  // padding dimension 1
+            int                  d0,  // dilation dimension 0
+            int                  d1); // dilation dimension 1
+
+    // Depthwise 2D convolution
+    // may be faster than wsp_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
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_2d_dw_direct(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_transpose_2d_p0(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -1579,12 +1750,41 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    // nearest interpolate
-    // used in stable-diffusion
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pool_2d_back(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * af, // "a"/input used in forward pass
+            enum wsp_ggml_op_pool     op,
+            int                   k0,
+            int                   k1,
+            int                   s0,
+            int                   s1,
+            float                 p0,
+            float                 p1);
+
+    enum wsp_ggml_scale_mode {
+        WSP_GGML_SCALE_MODE_NEAREST  = 0,
+        WSP_GGML_SCALE_MODE_BILINEAR = 1,
+    };
+
+    // interpolate
+    // multiplies ne0 and ne1 by scale factor
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_upscale(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
-            int                   scale_factor);
+            int                   scale_factor,
+            enum wsp_ggml_scale_mode  mode);
+
+    // interpolate
+    // interpolate scale to specified dimensions
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_upscale_ext(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3,
+            enum wsp_ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad(
@@ -1595,10 +1795,37 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_reflect_1d(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                   p0,
+            int                   p1);
+
+    // Move tensor elements by an offset given for each dimension. Elements that
+    // are shifted beyond the last position are wrapped around to the beginning.
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_roll(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                   shift0,
+            int                   shift1,
+            int                   shift2,
+            int                   shift3);
+
+
+    // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+    // timesteps: [N,]
+    // return: [N, dim]
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_timestep_embedding(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * timesteps,
+            int                   dim,
+            int                   max_period);
+
     // sort rows
     enum wsp_ggml_sort_order {
-        WSP_GGML_SORT_ASC,
-        WSP_GGML_SORT_DESC,
+        WSP_GGML_SORT_ORDER_ASC,
+        WSP_GGML_SORT_ORDER_DESC,
     };
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_argsort(
@@ -1606,19 +1833,43 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             enum wsp_ggml_sort_order  order);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_arange(
+            struct wsp_ggml_context * ctx,
+            float                 start,
+            float                 stop,
+            float                 step);
+
     // top k elements per row
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_top_k(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             int                   k);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_attn(
+#define WSP_GGML_KQ_MASK_PAD 64
+
+    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = WSP_GGML_PAD(n_batch, WSP_GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_attn_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * q,
             struct wsp_ggml_tensor  * k,
             struct wsp_ggml_tensor  * v,
-            bool                  masked);
+            struct wsp_ggml_tensor  * mask,
+            float                 scale,
+            float                 max_bias,
+            float                 logit_softcap);
 
+    WSP_GGML_API void wsp_ggml_flash_attn_ext_set_prec(
+            struct wsp_ggml_tensor * a,
+            enum wsp_ggml_prec       prec);
+
+    WSP_GGML_API enum wsp_ggml_prec wsp_ggml_flash_attn_ext_get_prec(
+            const struct wsp_ggml_tensor * a);
+
+    // TODO: needs to be adapted to wsp_ggml_flash_attn_ext
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_attn_back(
            struct wsp_ggml_context * ctx,
            struct wsp_ggml_tensor  * q,
@@ -1627,13 +1878,19 @@ extern "C" {
            struct wsp_ggml_tensor  * d,
            bool                  masked);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_ff(
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_ssm_conv(
             struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * a,
-            struct wsp_ggml_tensor  * b0,
-            struct wsp_ggml_tensor  * b1,
-            struct wsp_ggml_tensor  * c0,
-            struct wsp_ggml_tensor  * c1);
+            struct wsp_ggml_tensor  * sx,
+            struct wsp_ggml_tensor  * c);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_ssm_scan(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * s,
+            struct wsp_ggml_tensor  * x,
+            struct wsp_ggml_tensor  * dt,
+            struct wsp_ggml_tensor  * A,
+            struct wsp_ggml_tensor  * B,
+            struct wsp_ggml_tensor  * C);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -1685,90 +1942,42 @@ extern "C" {
             struct wsp_ggml_tensor  * pw,
             struct wsp_ggml_tensor  * ph);
 
-    // custom operators
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rwkv_wkv6(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * k,
+            struct wsp_ggml_tensor  * v,
+            struct wsp_ggml_tensor  * r,
+            struct wsp_ggml_tensor  * tf,
+            struct wsp_ggml_tensor  * td,
+            struct wsp_ggml_tensor  * state);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_gated_linear_attn(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * k,
+            struct wsp_ggml_tensor  * v,
+            struct wsp_ggml_tensor  * q,
+            struct wsp_ggml_tensor  * g,
+            struct wsp_ggml_tensor  * state,
+            float scale);
 
-    typedef void (*wsp_ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*wsp_ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*wsp_ggml_custom1_op_f32_t)(struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *);
-    typedef void (*wsp_ggml_custom2_op_f32_t)(struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *);
-    typedef void (*wsp_ggml_custom3_op_f32_t)(struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *, const struct wsp_ggml_tensor *);
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_unary_f32(
-            struct wsp_ggml_context        * ctx,
-            struct wsp_ggml_tensor         * a,
-                   wsp_ggml_unary_op_f32_t   fun),
-        "use wsp_ggml_map_custom1 instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_unary_inplace_f32(
-            struct wsp_ggml_context        * ctx,
-            struct wsp_ggml_tensor         * a,
-                   wsp_ggml_unary_op_f32_t   fun),
-        "use wsp_ggml_map_custom1_inplace instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_binary_f32(
-            struct wsp_ggml_context         * ctx,
-            struct wsp_ggml_tensor          * a,
-            struct wsp_ggml_tensor          * b,
-                   wsp_ggml_binary_op_f32_t   fun),
-        "use wsp_ggml_map_custom2 instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_binary_inplace_f32(
-            struct wsp_ggml_context         * ctx,
-            struct wsp_ggml_tensor          * a,
-            struct wsp_ggml_tensor          * b,
-                   wsp_ggml_binary_op_f32_t   fun),
-        "use wsp_ggml_map_custom2_inplace instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom1_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-                   wsp_ggml_custom1_op_f32_t   fun),
-        "use wsp_ggml_map_custom1 instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom1_inplace_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-                   wsp_ggml_custom1_op_f32_t   fun),
-        "use wsp_ggml_map_custom1_inplace instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom2_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-            struct wsp_ggml_tensor           * b,
-                   wsp_ggml_custom2_op_f32_t   fun),
-        "use wsp_ggml_map_custom2 instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom2_inplace_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-            struct wsp_ggml_tensor           * b,
-                   wsp_ggml_custom2_op_f32_t   fun),
-        "use wsp_ggml_map_custom2_inplace instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom3_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-            struct wsp_ggml_tensor           * b,
-            struct wsp_ggml_tensor           * c,
-                   wsp_ggml_custom3_op_f32_t   fun),
-        "use wsp_ggml_map_custom3 instead");
-
-    WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom3_inplace_f32(
-            struct wsp_ggml_context          * ctx,
-            struct wsp_ggml_tensor           * a,
-            struct wsp_ggml_tensor           * b,
-            struct wsp_ggml_tensor           * c,
-                   wsp_ggml_custom3_op_f32_t   fun),
-        "use wsp_ggml_map_custom3_inplace instead");
-
-    // custom operators v2
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rwkv_wkv7(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * r,
+            struct wsp_ggml_tensor  * w,
+            struct wsp_ggml_tensor  * k,
+            struct wsp_ggml_tensor  * v,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * state);
+
+    // custom operators
 
     typedef void (*wsp_ggml_custom1_op_t)(struct wsp_ggml_tensor * dst , const struct wsp_ggml_tensor * a, int ith, int nth, void * userdata);
     typedef void (*wsp_ggml_custom2_op_t)(struct wsp_ggml_tensor * dst , const struct wsp_ggml_tensor * a, const struct wsp_ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*wsp_ggml_custom3_op_t)(struct wsp_ggml_tensor * dst , const struct wsp_ggml_tensor * a, const struct wsp_ggml_tensor * b, const struct wsp_ggml_tensor * c, int ith, int nth, void * userdata);
 
-    #define WSP_GGML_N_TASKS_MAX -1
+#define WSP_GGML_N_TASKS_MAX (-1)
+    // n_tasks == WSP_GGML_N_TASKS_MAX means to use max number of tasks
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_map_custom1(
             struct wsp_ggml_context   * ctx,
@@ -1818,56 +2027,85 @@ extern "C" {
             int                     n_tasks,
             void                  * userdata);
 
+    typedef void (*wsp_ggml_custom_op_t)(struct wsp_ggml_tensor * dst , int ith, int nth, void * userdata);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_custom_4d(
+            struct wsp_ggml_context * ctx,
+            enum wsp_ggml_type        type,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            struct wsp_ggml_tensor ** args,
+            int                   n_args,
+            wsp_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_custom_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor ** args,
+            int                   n_args,
+            wsp_ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
     // loss function
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cross_entropy_loss(
-            struct wsp_ggml_context         * ctx,
-            struct wsp_ggml_tensor          * a,
-            struct wsp_ggml_tensor          * b);
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,  // logits
+            struct wsp_ggml_tensor  * b); // labels
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cross_entropy_loss_back(
-            struct wsp_ggml_context         * ctx,
-            struct wsp_ggml_tensor          * a,
-            struct wsp_ggml_tensor          * b,
-            struct wsp_ggml_tensor          * c);
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,  // logits
+            struct wsp_ggml_tensor  * b,  // labels
+            struct wsp_ggml_tensor  * c); // gradients of cross_entropy_loss result
+
+    // AdamW optimizer step
+    // Paper: https://arxiv.org/pdf/1711.05101v3.pdf
+    // PyTorch: https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_opt_step_adamw(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * grad,
+            struct wsp_ggml_tensor  * m,
+            struct wsp_ggml_tensor  * v,
+            struct wsp_ggml_tensor  * adamw_params); // parameters such a the learning rate
 
     //
     // automatic differentiation
     //
 
-    WSP_GGML_API void wsp_ggml_set_param(
-            struct wsp_ggml_context * ctx,
-            struct wsp_ggml_tensor  * tensor);
+    WSP_GGML_API void wsp_ggml_build_forward_expand(struct wsp_ggml_cgraph * cgraph, struct wsp_ggml_tensor * tensor);
+    WSP_GGML_API void wsp_ggml_build_backward_expand(
+        struct wsp_ggml_context *  ctx,        // context for gradient computation
+        struct wsp_ggml_cgraph  *  cgraph,
+        struct wsp_ggml_tensor  ** grad_accs);
 
+    // graph allocation in a context
+    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_new_graph       (struct wsp_ggml_context * ctx); // size = WSP_GGML_DEFAULT_GRAPH_SIZE, grads = false
+    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_new_graph_custom(struct wsp_ggml_context * ctx, size_t size, bool grads);
+    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_graph_dup       (struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * cgraph, bool force_grads);
+    WSP_GGML_API void                 wsp_ggml_graph_cpy       (struct wsp_ggml_cgraph * src, struct wsp_ggml_cgraph * dst);
+    WSP_GGML_API void                 wsp_ggml_graph_reset     (struct wsp_ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
+    WSP_GGML_API void                 wsp_ggml_graph_clear     (struct wsp_ggml_cgraph * cgraph);
 
-    WSP_GGML_API void wsp_ggml_build_forward_expand (struct wsp_ggml_cgraph * cgraph, struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API void wsp_ggml_build_backward_expand(struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * gf, struct wsp_ggml_cgraph * gb, bool keep);
+    WSP_GGML_API int                   wsp_ggml_graph_size   (struct wsp_ggml_cgraph * cgraph);
+    WSP_GGML_API struct wsp_ggml_tensor *  wsp_ggml_graph_node   (struct wsp_ggml_cgraph * cgraph, int i); // if i < 0, returns nodes[n_nodes + i]
+    WSP_GGML_API struct wsp_ggml_tensor ** wsp_ggml_graph_nodes  (struct wsp_ggml_cgraph * cgraph);
+    WSP_GGML_API int                   wsp_ggml_graph_n_nodes(struct wsp_ggml_cgraph * cgraph);
 
-    // graph allocation in a context
-    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_new_graph         (struct wsp_ggml_context * ctx); // size = WSP_GGML_DEFAULT_GRAPH_SIZE, grads = false
-    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_new_graph_custom  (struct wsp_ggml_context * ctx, size_t size, bool grads);
-    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_graph_dup         (struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * cgraph);
-    WSP_GGML_API struct wsp_ggml_cgraph   wsp_ggml_graph_view        (struct wsp_ggml_cgraph * cgraph, int i0, int i1);
-    WSP_GGML_API void                 wsp_ggml_graph_cpy         (struct wsp_ggml_cgraph * src, struct wsp_ggml_cgraph * dst);
-    WSP_GGML_API void                 wsp_ggml_graph_reset       (struct wsp_ggml_cgraph * cgraph);  // zero grads
-    WSP_GGML_API void                 wsp_ggml_graph_clear       (struct wsp_ggml_cgraph * cgraph);
+    WSP_GGML_API void   wsp_ggml_graph_add_node(struct wsp_ggml_cgraph * cgraph, struct wsp_ggml_tensor * tensor);
 
     WSP_GGML_API size_t wsp_ggml_graph_overhead(void);
     WSP_GGML_API size_t wsp_ggml_graph_overhead_custom(size_t size, bool grads);
 
-    // wsp_ggml_graph_plan() has to be called before wsp_ggml_graph_compute()
-    // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    WSP_GGML_API struct wsp_ggml_cplan wsp_ggml_graph_plan   (const struct wsp_ggml_cgraph * cgraph, int n_threads /*= WSP_GGML_DEFAULT_N_THREADS*/);
-    WSP_GGML_API int               wsp_ggml_graph_compute(      struct wsp_ggml_cgraph * cgraph, struct wsp_ggml_cplan * cplan);
-
-    // same as wsp_ggml_graph_compute() but the work data is allocated as a part of the context
-    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    WSP_GGML_API void wsp_ggml_graph_compute_with_ctx(struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * cgraph, int n_threads);
-
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_graph_get_tensor(struct wsp_ggml_cgraph * cgraph, const char * name);
-
-    WSP_GGML_API void                 wsp_ggml_graph_export(const struct wsp_ggml_cgraph * cgraph, const char * fname);
-    WSP_GGML_API struct wsp_ggml_cgraph * wsp_ggml_graph_import(const char * fname, struct wsp_ggml_context ** ctx_data, struct wsp_ggml_context ** ctx_eval);
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_graph_get_tensor  (const struct wsp_ggml_cgraph * cgraph, const char * name);
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_graph_get_grad    (const struct wsp_ggml_cgraph * cgraph, const struct wsp_ggml_tensor * node);
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_graph_get_grad_acc(const struct wsp_ggml_cgraph * cgraph, const struct wsp_ggml_tensor * node);
 
     // print info and performance information for the graph
     WSP_GGML_API void wsp_ggml_graph_print(const struct wsp_ggml_cgraph * cgraph);
@@ -1875,191 +2113,14 @@ extern "C" {
     // dump the graph into a file using the dot format
     WSP_GGML_API void wsp_ggml_graph_dump_dot(const struct wsp_ggml_cgraph * gb, const struct wsp_ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    WSP_GGML_API void wsp_ggml_build_backward_gradient_checkpointing(
-            struct wsp_ggml_context   * ctx,
-            struct wsp_ggml_cgraph    * gf,
-            struct wsp_ggml_cgraph    * gb,
-            struct wsp_ggml_cgraph    * gb_tmp,
-            struct wsp_ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum wsp_ggml_opt_type {
-        WSP_GGML_OPT_ADAM,
-        WSP_GGML_OPT_LBFGS,
-    };
-
-    // linesearch methods
-    enum wsp_ggml_linesearch {
-        WSP_GGML_LINESEARCH_DEFAULT = 1,
-
-        WSP_GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        WSP_GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        WSP_GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum wsp_ggml_opt_result {
-        WSP_GGML_OPT_OK = 0,
-        WSP_GGML_OPT_DID_NOT_CONVERGE,
-        WSP_GGML_OPT_NO_CONTEXT,
-        WSP_GGML_OPT_INVALID_WOLFE,
-        WSP_GGML_OPT_FAIL,
-        WSP_GGML_OPT_CANCEL,
-
-        WSP_GGML_LINESEARCH_FAIL = -128,
-        WSP_GGML_LINESEARCH_MINIMUM_STEP,
-        WSP_GGML_LINESEARCH_MAXIMUM_STEP,
-        WSP_GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        WSP_GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*wsp_ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*wsp_ggml_log_callback)(enum wsp_ggml_log_level level, const char * text, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (wsp_ggml_opt_default_params) for default values
-    //
-    struct wsp_ggml_opt_params {
-        enum wsp_ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum wsp_ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct wsp_ggml_opt_context {
-        struct wsp_ggml_context * ctx;
-        struct wsp_ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct wsp_ggml_tensor * g;  // current gradient
-            struct wsp_ggml_tensor * m;  // first moment
-            struct wsp_ggml_tensor * v;  // second moment
-            struct wsp_ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct wsp_ggml_tensor * x;    // current parameters
-            struct wsp_ggml_tensor * xp;   // previous parameters
-            struct wsp_ggml_tensor * g;    // current gradient
-            struct wsp_ggml_tensor * gp;   // previous gradient
-            struct wsp_ggml_tensor * d;    // search direction
-            struct wsp_ggml_tensor * pf;   // past function values
-            struct wsp_ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct wsp_ggml_tensor * lmys; // the L-BFGS memory ys
-            struct wsp_ggml_tensor * lms;  // the L-BFGS memory s
-            struct wsp_ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
-    WSP_GGML_API struct wsp_ggml_opt_params wsp_ggml_opt_default_params(enum wsp_ggml_opt_type type);
+    // Set callback for all future logging events.
+    // If this is not called, or NULL is supplied, everything is output on stderr.
+    WSP_GGML_API void wsp_ggml_log_set(wsp_ggml_log_callback log_callback, void * user_data);
 
-    // optimize the function defined by the tensor f
-    WSP_GGML_API enum wsp_ggml_opt_result wsp_ggml_opt(
-            struct wsp_ggml_context * ctx,
-            struct wsp_ggml_opt_params params,
-            struct wsp_ggml_tensor * f);
-
-    // initialize optimizer context
-    WSP_GGML_API void wsp_ggml_opt_init(
-            struct wsp_ggml_context     * ctx,
-            struct wsp_ggml_opt_context * opt,
-            struct wsp_ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    WSP_GGML_API enum wsp_ggml_opt_result wsp_ggml_opt_resume(
-            struct wsp_ggml_context * ctx,
-            struct wsp_ggml_opt_context * opt,
-            struct wsp_ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    WSP_GGML_API enum wsp_ggml_opt_result wsp_ggml_opt_resume_g(
-            struct wsp_ggml_context * ctx,
-            struct wsp_ggml_opt_context * opt,
-            struct wsp_ggml_tensor * f,
-            struct wsp_ggml_cgraph * gf,
-            struct wsp_ggml_cgraph * gb,
-            wsp_ggml_opt_callback callback,
-            void * callback_data);
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_zero(struct wsp_ggml_tensor * tensor);
 
     //
     // quantization
@@ -2077,201 +2138,83 @@ extern "C" {
     WSP_GGML_API void wsp_ggml_wsp_quantize_init(enum wsp_ggml_type type);
     WSP_GGML_API void wsp_ggml_wsp_quantize_free(void);
 
-    // TODO: these would probably get removed in favor of the more general wsp_ggml_wsp_quantize_chunk
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
-
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-
     // some quantization type cannot be used without an importance matrix
     WSP_GGML_API bool wsp_ggml_wsp_quantize_requires_imatrix(enum wsp_ggml_type type);
 
     // calls wsp_ggml_wsp_quantize_init internally (i.e. can allocate memory)
-    WSP_GGML_API size_t wsp_ggml_wsp_quantize_chunk(enum wsp_ggml_type type, const float * src, void * dst,
-            int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-
-    //
-    // gguf
-    //
-
-    enum wsp_gguf_type {
-        WSP_GGUF_TYPE_UINT8   = 0,
-        WSP_GGUF_TYPE_INT8    = 1,
-        WSP_GGUF_TYPE_UINT16  = 2,
-        WSP_GGUF_TYPE_INT16   = 3,
-        WSP_GGUF_TYPE_UINT32  = 4,
-        WSP_GGUF_TYPE_INT32   = 5,
-        WSP_GGUF_TYPE_FLOAT32 = 6,
-        WSP_GGUF_TYPE_BOOL    = 7,
-        WSP_GGUF_TYPE_STRING  = 8,
-        WSP_GGUF_TYPE_ARRAY   = 9,
-        WSP_GGUF_TYPE_UINT64  = 10,
-        WSP_GGUF_TYPE_INT64   = 11,
-        WSP_GGUF_TYPE_FLOAT64 = 12,
-        WSP_GGUF_TYPE_COUNT,       // marks the end of the enum
+    WSP_GGML_API size_t wsp_ggml_wsp_quantize_chunk(
+            enum wsp_ggml_type   type,
+               const float * src,
+                      void * dst,
+                   int64_t   start,
+                   int64_t   nrows,
+                   int64_t   n_per_row,
+               const float * imatrix);
+
+#ifdef __cplusplus
+    // restrict not standard in C++
+#    if defined(__GNUC__)
+#        define WSP_GGML_RESTRICT __restrict__
+#    elif defined(__clang__)
+#        define WSP_GGML_RESTRICT __restrict
+#    elif defined(_MSC_VER)
+#        define WSP_GGML_RESTRICT __restrict
+#    else
+#        define WSP_GGML_RESTRICT
+#    endif
+#else
+#    if defined (_MSC_VER) && (__STDC_VERSION__ < 201112L)
+#        define WSP_GGML_RESTRICT __restrict
+#    else
+#        define WSP_GGML_RESTRICT restrict
+#    endif
+#endif
+    typedef void (*wsp_ggml_to_float_t)  (const void  * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+    typedef void (*wsp_ggml_from_float_t)(const float * WSP_GGML_RESTRICT x, void  * WSP_GGML_RESTRICT y, int64_t k);
+
+    struct wsp_ggml_type_traits {
+        const char             * type_name;
+        int64_t                  blck_size;
+        int64_t                  blck_size_interleave; // interleave elements in blocks
+        size_t                   type_size;
+        bool                     is_quantized;
+        wsp_ggml_to_float_t          to_float;
+        wsp_ggml_from_float_t        from_float_ref;
     };
 
-    struct wsp_gguf_context;
+    WSP_GGML_API const struct wsp_ggml_type_traits * wsp_ggml_get_type_traits(enum wsp_ggml_type type);
 
-    struct wsp_gguf_init_params {
-        bool no_alloc;
+    // ggml threadpool
+    // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend
+    // the goal should be to create an API that other backends can use move everything to the ggml base
 
-        // if not NULL, create a wsp_ggml_context and allocate the tensor data in it
-        struct wsp_ggml_context ** ctx;
+    // scheduling priorities
+    enum wsp_ggml_sched_priority {
+        WSP_GGML_SCHED_PRIO_LOW = -1,
+        WSP_GGML_SCHED_PRIO_NORMAL,
+        WSP_GGML_SCHED_PRIO_MEDIUM,
+        WSP_GGML_SCHED_PRIO_HIGH,
+        WSP_GGML_SCHED_PRIO_REALTIME
     };
 
-    WSP_GGML_API struct wsp_gguf_context * wsp_gguf_init_empty(void);
-    WSP_GGML_API struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp_gguf_init_params params);
-    //WSP_GGML_API struct wsp_gguf_context * wsp_gguf_init_from_buffer(..);
-
-    WSP_GGML_API void wsp_gguf_free(struct wsp_gguf_context * ctx);
-
-    WSP_GGML_API const char * wsp_gguf_type_name(enum wsp_gguf_type type);
-
-    WSP_GGML_API int    wsp_gguf_get_version    (const struct wsp_gguf_context * ctx);
-    WSP_GGML_API size_t wsp_gguf_get_alignment  (const struct wsp_gguf_context * ctx);
-    WSP_GGML_API size_t wsp_gguf_get_data_offset(const struct wsp_gguf_context * ctx);
-    WSP_GGML_API void * wsp_gguf_get_data       (const struct wsp_gguf_context * ctx);
-
-    WSP_GGML_API int          wsp_gguf_get_n_kv(const struct wsp_gguf_context * ctx);
-    WSP_GGML_API int          wsp_gguf_find_key(const struct wsp_gguf_context * ctx, const char * key);
-    WSP_GGML_API const char * wsp_gguf_get_key (const struct wsp_gguf_context * ctx, int key_id);
-
-    WSP_GGML_API enum wsp_gguf_type wsp_gguf_get_kv_type (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API enum wsp_gguf_type wsp_gguf_get_arr_type(const struct wsp_gguf_context * ctx, int key_id);
-
-    // will abort if the wrong type is used for the key
-    WSP_GGML_API uint8_t      wsp_gguf_get_val_u8  (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API int8_t       wsp_gguf_get_val_i8  (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API uint16_t     wsp_gguf_get_val_u16 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API int16_t      wsp_gguf_get_val_i16 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API uint32_t     wsp_gguf_get_val_u32 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API int32_t      wsp_gguf_get_val_i32 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API float        wsp_gguf_get_val_f32 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API uint64_t     wsp_gguf_get_val_u64 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API int64_t      wsp_gguf_get_val_i64 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API double       wsp_gguf_get_val_f64 (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API bool         wsp_gguf_get_val_bool(const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API const char * wsp_gguf_get_val_str (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API const void * wsp_gguf_get_val_data(const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API int          wsp_gguf_get_arr_n   (const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API const void * wsp_gguf_get_arr_data(const struct wsp_gguf_context * ctx, int key_id);
-    WSP_GGML_API const char * wsp_gguf_get_arr_str (const struct wsp_gguf_context * ctx, int key_id, int i);
-
-    WSP_GGML_API int            wsp_gguf_get_n_tensors    (const struct wsp_gguf_context * ctx);
-    WSP_GGML_API int            wsp_gguf_find_tensor      (const struct wsp_gguf_context * ctx, const char * name);
-    WSP_GGML_API size_t         wsp_gguf_get_tensor_offset(const struct wsp_gguf_context * ctx, int i);
-    WSP_GGML_API char *         wsp_gguf_get_tensor_name  (const struct wsp_gguf_context * ctx, int i);
-    WSP_GGML_API enum wsp_ggml_type wsp_gguf_get_tensor_type  (const struct wsp_gguf_context * ctx, int i);
-
-    // overrides existing values or adds a new one
-    WSP_GGML_API void wsp_gguf_set_val_u8  (struct wsp_gguf_context * ctx, const char * key, uint8_t  val);
-    WSP_GGML_API void wsp_gguf_set_val_i8  (struct wsp_gguf_context * ctx, const char * key, int8_t   val);
-    WSP_GGML_API void wsp_gguf_set_val_u16 (struct wsp_gguf_context * ctx, const char * key, uint16_t val);
-    WSP_GGML_API void wsp_gguf_set_val_i16 (struct wsp_gguf_context * ctx, const char * key, int16_t  val);
-    WSP_GGML_API void wsp_gguf_set_val_u32 (struct wsp_gguf_context * ctx, const char * key, uint32_t val);
-    WSP_GGML_API void wsp_gguf_set_val_i32 (struct wsp_gguf_context * ctx, const char * key, int32_t  val);
-    WSP_GGML_API void wsp_gguf_set_val_f32 (struct wsp_gguf_context * ctx, const char * key, float    val);
-    WSP_GGML_API void wsp_gguf_set_val_u64 (struct wsp_gguf_context * ctx, const char * key, uint64_t val);
-    WSP_GGML_API void wsp_gguf_set_val_i64 (struct wsp_gguf_context * ctx, const char * key, int64_t  val);
-    WSP_GGML_API void wsp_gguf_set_val_f64 (struct wsp_gguf_context * ctx, const char * key, double   val);
-    WSP_GGML_API void wsp_gguf_set_val_bool(struct wsp_gguf_context * ctx, const char * key, bool     val);
-    WSP_GGML_API void wsp_gguf_set_val_str (struct wsp_gguf_context * ctx, const char * key, const char * val);
-    WSP_GGML_API void wsp_gguf_set_arr_data(struct wsp_gguf_context * ctx, const char * key, enum wsp_gguf_type type, const void * data, int n);
-    WSP_GGML_API void wsp_gguf_set_arr_str (struct wsp_gguf_context * ctx, const char * key, const char ** data, int n);
-
-    // set or add KV pairs from another context
-    WSP_GGML_API void wsp_gguf_set_kv(struct wsp_gguf_context * ctx, struct wsp_gguf_context * src);
-
-    // manage tensor info
-    WSP_GGML_API void wsp_gguf_add_tensor(struct wsp_gguf_context * ctx, const struct wsp_ggml_tensor * tensor);
-    WSP_GGML_API void wsp_gguf_set_tensor_type(struct wsp_gguf_context * ctx, const char * name, enum wsp_ggml_type type);
-    WSP_GGML_API void wsp_gguf_set_tensor_data(struct wsp_gguf_context * ctx, const char * name, const void * data, size_t size);
-
-    // writing gguf files can be done in 2 ways:
-    //
-    // - write the entire wsp_gguf_context to a binary file in a single pass:
-    //
-    //   wsp_gguf_write_to_file(ctx, fname);
-    //
-    // - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
-    //
-    //   FILE * f = fopen(fname, "wb");
-    //   fseek(f, wsp_gguf_get_meta_size(ctx), SEEK_SET);
-    //   fwrite(f, ...);
-    //   void * data = wsp_gguf_meta_get_meta_data(ctx);
-    //   fseek(f, 0, SEEK_SET);
-    //   fwrite(f, data, wsp_gguf_get_meta_size(ctx));
-    //   free(data);
-    //   fclose(f);
-    //
-
-    // write the entire context to a binary file
-    WSP_GGML_API void wsp_gguf_write_to_file(const struct wsp_gguf_context * ctx, const char * fname, bool only_meta);
-
-    // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
-    WSP_GGML_API size_t wsp_gguf_get_meta_size(const struct wsp_gguf_context * ctx);
-    WSP_GGML_API void   wsp_gguf_get_meta_data(const struct wsp_gguf_context * ctx, void * data);
+    // threadpool params
+    // Use wsp_ggml_threadpool_params_default() or wsp_ggml_threadpool_params_init() to populate the defaults
+    struct wsp_ggml_threadpool_params {
+        bool                cpumask[WSP_GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum wsp_ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
 
-    //
-    // system info
-    //
+    struct wsp_ggml_threadpool;     // forward declaration, see ggml.c
 
-    WSP_GGML_API int wsp_ggml_cpu_has_avx        (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_avx_vnni   (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_avx2       (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_avx512     (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_avx512_vbmi(void);
-    WSP_GGML_API int wsp_ggml_cpu_has_avx512_vnni(void);
-    WSP_GGML_API int wsp_ggml_cpu_has_fma        (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_neon       (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_arm_fma    (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_metal      (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_f16c       (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_fp16_va    (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_wasm_simd  (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_blas       (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_cublas     (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_clblast    (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_gpublas    (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_sse3       (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_ssse3      (void);
-    WSP_GGML_API int wsp_ggml_cpu_has_vsx        (void);
+    typedef struct wsp_ggml_threadpool * wsp_ggml_threadpool_t;
 
-    //
-    // Internal types and functions exposed for tests and benchmarks
-    //
-
-#ifdef  __cplusplus
-// restrict not standard in C++
-#define WSP_GGML_RESTRICT
-#else
-#define WSP_GGML_RESTRICT restrict
-#endif
-    typedef void (*wsp_ggml_to_float_t)  (const void  * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int k);
-    typedef void (*wsp_ggml_from_float_t)(const float * WSP_GGML_RESTRICT x, void  * WSP_GGML_RESTRICT y, int k);
-    typedef void (*wsp_ggml_vec_dot_t)   (const int n, float * WSP_GGML_RESTRICT s, const void * WSP_GGML_RESTRICT x, const void * WSP_GGML_RESTRICT y);
-
-    typedef struct {
-        const char      * type_name;
-        int               blck_size;
-        size_t            type_size;
-        bool              is_quantized;
-        wsp_ggml_to_float_t   to_float;
-        wsp_ggml_from_float_t from_float;
-        wsp_ggml_from_float_t from_float_reference;
-        wsp_ggml_vec_dot_t    vec_dot;
-        enum wsp_ggml_type    vec_dot_type;
-    } wsp_ggml_type_traits_t;
-
-    WSP_GGML_API wsp_ggml_type_traits_t wsp_ggml_internal_get_type_traits(enum wsp_ggml_type type);
+    WSP_GGML_API struct wsp_ggml_threadpool_params wsp_ggml_threadpool_params_default(int n_threads);
+    WSP_GGML_API void                          wsp_ggml_threadpool_params_init   (struct wsp_ggml_threadpool_params * p, int n_threads);
+    WSP_GGML_API bool                          wsp_ggml_threadpool_params_match  (const struct wsp_ggml_threadpool_params * p0, const struct wsp_ggml_threadpool_params * p1);
 
 #ifdef  __cplusplus
 }