llama_cpp 0.7.0 → 0.7.1

data/ext/llama_cpp/src/ggml.c

@@ -162,40 +162,16 @@ typedef void * thread_ret_t;
 
 #define GGML_PRINT(...) printf(__VA_ARGS__)
 
+//
+// end of logging block
+//
+
 #ifdef GGML_USE_ACCELERATE
 // uncomment to use vDSP for soft max computation
 // note: not sure if it is actually faster
 //#define GGML_SOFT_MAX_ACCELERATE
 #endif
 
-//
-// logging
-//
-
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_10(...)
-#endif
-
-#define GGML_PRINT(...) printf(__VA_ARGS__)
-
-//
-// end of logging block
-//
-
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
 #define GGML_ALIGNED_FREE(ptr)    _aligned_free(ptr)
@@ -4951,6 +4927,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     *result = (struct ggml_tensor) {
         /*.type         =*/ type,
         /*.backend      =*/ GGML_BACKEND_CPU,
+        /*.buffer       =*/ NULL,
         /*.n_dims       =*/ n_dims,
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
@@ -5517,6 +5494,39 @@ struct ggml_tensor * ggml_view_tensor(
     return result;
 }
 
+struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
+    struct ggml_object * obj = ctx->objects_begin;
+
+    char * const mem_buffer = ctx->mem_buffer;
+
+    while (obj != NULL) {
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            return (struct ggml_tensor *)(mem_buffer + obj->offs);
+        }
+
+        obj = obj->next;
+    }
+
+    return NULL;
+}
+
+struct ggml_tensor * ggml_get_next_tensor(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+    struct ggml_object * obj = (struct ggml_object *) ((char *)tensor - GGML_OBJECT_SIZE);
+    obj = obj->next;
+
+    char * const mem_buffer = ctx->mem_buffer;
+
+    while (obj != NULL) {
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            return (struct ggml_tensor *)(mem_buffer + obj->offs);
+        }
+
+        obj = obj->next;
+    }
+
+    return NULL;
+}
+
 struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name) {
     struct ggml_object * obj = ctx->objects_begin;
 
@@ -8670,6 +8680,7 @@ void ggml_set_param(
 
     GGML_ASSERT(tensor->grad == NULL);
     tensor->grad = ggml_dup_tensor(ctx, tensor);
+    ggml_format_name(tensor->grad, "%s (grad)", tensor->name);
 }
 
 // ggml_compute_forward_dup
@@ -11256,7 +11267,7 @@ static void ggml_compute_forward_silu_f32(
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
             UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
@@ -13082,24 +13093,22 @@ static void ggml_compute_forward_alibi_f32(
         return;
     }
 
-    const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    assert(n_past >= 0);
-
-    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int ne1 = src0->ne[1]; // seq_len_without_past
-    const int ne2 = src0->ne[2]; // n_head -> this is k
-    //const int ne3 = src0->ne[3]; // 1 -> bsz
+    const int64_t ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
+    const int64_t ne1 = src0->ne[1]; // seq_len_without_past
+    const int64_t ne2 = src0->ne[2]; // n_head -> this is k
+    //const int64_t ne3 = src0->ne[3]; // 1 -> bsz
 
-    const int n  = ggml_nrows(src0);
-    const int ne2_ne3 = n/ne1; // ne2*ne3
+    const int64_t n  = ggml_nrows(src0);
+    const int64_t ne2_ne3 = n/ne1; // ne2*ne3
 
-    const int nb0 = src0->nb[0];
-    const int nb1 = src0->nb[1];
-    const int nb2 = src0->nb[2];
+    const size_t nb0 = src0->nb[0];
+    const size_t nb1 = src0->nb[1];
+    const size_t nb2 = src0->nb[2];
     //const int nb3 = src0->nb[3];
 
     GGML_ASSERT(nb0 == sizeof(float));
@@ -13111,9 +13120,9 @@ static void ggml_compute_forward_alibi_f32(
     const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
-    for (int i = 0; i < ne0; i++) {
-        for (int j = 0; j < ne1; j++) {
-            for (int k = 0; k < ne2_ne3; k++) {
+    for (int64_t i = 0; i < ne0; i++) {
+        for (int64_t j = 0; j < ne1; j++) {
+            for (int64_t k = 0; k < ne2_ne3; k++) {
                 float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
                 float *      pdst = (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
 
@@ -13128,7 +13137,6 @@ static void ggml_compute_forward_alibi_f32(
                 }
 
                 pdst[0] = i * m_k + src[0];
-
             }
         }
     }
@@ -14454,7 +14462,7 @@ static void ggml_compute_forward_conv_2d_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -20203,6 +20211,10 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         ggml_vec_cpy_f32(nx, xp, x);
         ggml_vec_cpy_f32(nx, gp, g);
 
+        // TODO: instead of passing &cancel here, use the return code of the linesearch
+        //       to determine if the optimization should be cancelled
+        //       this is a simple change, but not doing this atm, since I don't have a nice
+        //       way to test and don't want to break something with so many changes lined up
         ls = linesearch_backtracking(&params, nx, x, &fx, g, d, step, xp, f, gb, &cplan, np, ps, &cancel, callback, callback_data);
         if (cancel) {
             return GGML_OPT_CANCEL;

data/ext/llama_cpp/src/ggml.h

@@ -326,7 +326,7 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
-    enum ggml_backend {
+    enum ggml_backend_type {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_GPU = 10,
         GGML_BACKEND_GPU_SPLIT = 20,
@@ -479,8 +479,10 @@ extern "C" {
 
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type    type;
-        enum ggml_backend backend;
+        enum ggml_type         type;
+        enum ggml_backend_type backend;
+
+        struct ggml_backend_buffer * buffer;
 
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
@@ -514,7 +516,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[4];
+        char padding[12];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -702,6 +704,9 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
+    // Context tensor enumeration and lookup
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
@@ -1358,7 +1363,7 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_alibi(
+    GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
@@ -1367,7 +1372,7 @@ extern "C" {
 
     // clamp
     // in-place, returns view(a)
-    struct ggml_tensor * ggml_clamp(
+    GGML_API struct ggml_tensor * ggml_clamp(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             float                 min,
@@ -2102,7 +2107,7 @@ extern "C" {
         enum ggml_type    vec_dot_type;
     } ggml_type_traits_t;
 
-    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }

data/ext/llama_cpp/src/k_quants.h

@@ -29,7 +29,7 @@
 
 // 2-bit quantization
 // weight is represented as x = a * q + b
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 2.5625 bits per weight
 typedef struct {
     uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
@@ -41,7 +41,7 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "w
 
 // 3-bit quantization
 // weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
 #ifdef GGML_QKK_64
 typedef struct {
@@ -62,7 +62,7 @@ static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 +
 #endif
 
 // 4-bit quantization
-// 16 blocks of 32 elements each
+// 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
 #ifdef GGML_QKK_64
@@ -83,7 +83,7 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/
 #endif
 
 // 5-bit quantization
-// 16 blocks of 32 elements each
+// 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
 #ifdef GGML_QKK_64
@@ -107,7 +107,7 @@ static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/
 
 // 6-bit quantization
 // weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 6.5625 bits per weight
 typedef struct {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits