llama_cpp 0.9.5 → 0.10.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 4fd4e1a5e4d7e2442ab43255996da3ce92f898f9876f1bda343e2433c5050dd7
-  data.tar.gz: dece2da6c9befa15e6990d18fb58e2bf13d8da6c62033969b6b5104f82df736d
+  metadata.gz: 9bd5dbea8695fdb41ea6d97e372c2cea452ee8ed070e26bd558a720d6c24fe27
+  data.tar.gz: 8d7bfd02445df81644eebb3a6db49bb1ddf241a344fef76a0b949f1c12d0639d
 SHA512:
-  metadata.gz: 51a383690b6e90e9493e1f318e916dfd94a909f4e554afd8ea822d047f05e96be3e2f371e83f0da5a37a9837d9ae5ecc6992bb9d9c0fd60a9de521bcd148e8f7
-  data.tar.gz: 15bbe94edb232d1979f2907c6c3ab7325a1089f9dcdd5d4262d7f0955fd6183e6b01cfee16593165f6e9901991e765ea30740bc1a83cca8fad60df4417551e3b
+  metadata.gz: 116423f4581f605ee379bcd690299e152087c03665be6171bc137205f0824be5d5e0ce6d3c0b548fc6d28193c601679c3cc27a101b6f5b58968c69388a70cbfc
+  data.tar.gz: fd162b1c4e26732573d32ba7439fc44dd2ff09a9024c64ff5adce997e154a934d9556f0ff3cfd6b83dd7dcbfb8366cc6ce394b8324f39ef349d4cb5834ea43f1

data/CHANGELOG.md

@@ -1,3 +1,19 @@
+## [[0.10.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.10.0...v0.10.1)] - 2023-12-16
+
+- Bump bundled llama.cpp from b1620 to b1641.
+- Add attribute reader for `params` to `Model`.
+- Add `Batch` class, this class was not published because the author forgot to write `rb_define_class`.
+
+## [[0.10.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.9.5...v0.10.0)] - 2023-12-09
+
+- Bump bundled llama.cpp from b1593 to b1620.
+  - Add `ModelKVOverride` class.
+  - Add `offload_kqv`, `type_k`, and `type_v` to ContextParams.
+  - Add kv overwrite type constants.
+
+**Breaking Changes**
+- Remove `f16_kv` from ContextParams.
+
 ## [[0.9.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.9.4...v0.9.5)] - 2023-12-02
 
 - Bump bundled llama.cpp from b1555 to b1593.

data/ext/llama_cpp/llama_cpp.cpp

@@ -3,6 +3,7 @@
 VALUE rb_mLLaMACpp;
 VALUE rb_cLLaMABatch;
 VALUE rb_cLLaMAModel;
+VALUE rb_cLLaMAModelKVOverride;
 VALUE rb_cLLaMAModelParams;
 VALUE rb_cLLaMATimings;
 VALUE rb_cLLaMAContext;
@@ -612,6 +613,78 @@ const rb_data_type_t RbLLaMATimings::llama_timings_type = {
   RUBY_TYPED_FREE_IMMEDIATELY
 };
 
+class RbLLaMAModelKVOverride {
+public:
+  static VALUE llama_model_kv_override_alloc(VALUE self) {
+    llama_model_kv_override* ptr = (llama_model_kv_override*)ruby_xmalloc(sizeof(llama_model_kv_override));
+    new (ptr) llama_model_kv_override();
+    return TypedData_Wrap_Struct(self, &llama_model_kv_override_type, ptr);
+  }
+
+  static void llama_model_kv_override_free(void* ptr) {
+    ((llama_model_kv_override*)ptr)->~llama_model_kv_override();
+    ruby_xfree(ptr);
+  }
+
+  static size_t llama_model_kv_override_size(const void* ptr) {
+    return sizeof(*((llama_model_kv_override*)ptr));
+  }
+
+  static llama_model_kv_override* get_llama_model_kv_override(VALUE self) {
+    llama_model_kv_override* ptr;
+    TypedData_Get_Struct(self, llama_model_kv_override, &llama_model_kv_override_type, ptr);
+    return ptr;
+  }
+
+  static void define_class(VALUE outer) {
+    rb_cLLaMAModelKVOverride = rb_define_class_under(outer, "ModelKVOverride", rb_cObject);
+    rb_define_alloc_func(rb_cLLaMAModelKVOverride, llama_model_kv_override_alloc);
+    rb_define_method(rb_cLLaMAModelKVOverride, "key", RUBY_METHOD_FUNC(_llama_model_kv_override_get_key), 0);
+    rb_define_method(rb_cLLaMAModelKVOverride, "tag", RUBY_METHOD_FUNC(_llama_model_kv_override_get_tag), 0);
+    rb_define_method(rb_cLLaMAModelKVOverride, "int_value", RUBY_METHOD_FUNC(_llama_model_kv_override_get_int_value), 0);
+    rb_define_method(rb_cLLaMAModelKVOverride, "float_value", RUBY_METHOD_FUNC(_llama_model_kv_override_get_float_value), 0);
+    rb_define_method(rb_cLLaMAModelKVOverride, "bool_value", RUBY_METHOD_FUNC(_llama_model_kv_override_get_bool_value), 0);
+  }
+
+  static const rb_data_type_t llama_model_kv_override_type;
+
+private:
+  static VALUE _llama_model_kv_override_get_key(VALUE self) {
+    llama_model_kv_override* ptr = get_llama_model_kv_override(self);
+    return rb_utf8_str_new_cstr(ptr->key);
+  }
+
+  static VALUE _llama_model_kv_override_get_tag(VALUE self) {
+    llama_model_kv_override* ptr = get_llama_model_kv_override(self);
+    return INT2NUM(ptr->tag);
+  }
+
+  static VALUE _llama_model_kv_override_get_int_value(VALUE self) {
+    llama_model_kv_override* ptr = get_llama_model_kv_override(self);
+    return INT2NUM(ptr->int_value);
+  }
+
+  static VALUE _llama_model_kv_override_get_float_value(VALUE self) {
+    llama_model_kv_override* ptr = get_llama_model_kv_override(self);
+    return DBL2NUM(ptr->float_value);
+  }
+
+  static VALUE _llama_model_kv_override_get_bool_value(VALUE self) {
+    llama_model_kv_override* ptr = get_llama_model_kv_override(self);
+    return ptr->bool_value ? Qtrue : Qfalse;
+  }
+};
+
+const rb_data_type_t RbLLaMAModelKVOverride::llama_model_kv_override_type = {
+  "RbLLaMAModelKVOverride",
+  { NULL,
+    RbLLaMAModelKVOverride::llama_model_kv_override_free,
+    RbLLaMAModelKVOverride::llama_model_kv_override_size },
+  NULL,
+  NULL,
+  RUBY_TYPED_FREE_IMMEDIATELY
+};
+
 class LLaMAModelParamsWrapper {
 public:
   struct llama_model_params params;
@@ -812,14 +885,18 @@ public:
     rb_define_method(rb_cLLaMAContextParams, "yarn_beta_slow", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_beta_slow), 0);
     rb_define_method(rb_cLLaMAContextParams, "yarn_orig_ctx=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_orig_ctx), 1);
     rb_define_method(rb_cLLaMAContextParams, "yarn_orig_ctx", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_orig_ctx), 0);
+    rb_define_method(rb_cLLaMAContextParams, "type_k=", RUBY_METHOD_FUNC(_llama_context_params_set_type_k), 1);
+    rb_define_method(rb_cLLaMAContextParams, "type_k", RUBY_METHOD_FUNC(_llama_context_params_get_type_k), 0);
+    rb_define_method(rb_cLLaMAContextParams, "type_v=", RUBY_METHOD_FUNC(_llama_context_params_set_type_v), 1);
+    rb_define_method(rb_cLLaMAContextParams, "type_v", RUBY_METHOD_FUNC(_llama_context_params_get_type_v), 0);
     rb_define_method(rb_cLLaMAContextParams, "mul_mat_q=", RUBY_METHOD_FUNC(_llama_context_params_set_mul_mat_q), 1);
     rb_define_method(rb_cLLaMAContextParams, "mul_mat_q", RUBY_METHOD_FUNC(_llama_context_params_get_mul_mat_q), 0);
-    rb_define_method(rb_cLLaMAContextParams, "f16_kv=", RUBY_METHOD_FUNC(_llama_context_params_set_f16_kv), 1);
-    rb_define_method(rb_cLLaMAContextParams, "f16_kv", RUBY_METHOD_FUNC(_llama_context_params_get_f16_kv), 0);
     rb_define_method(rb_cLLaMAContextParams, "logits_all=", RUBY_METHOD_FUNC(_llama_context_params_set_logits_all), 1);
     rb_define_method(rb_cLLaMAContextParams, "logits_all", RUBY_METHOD_FUNC(_llama_context_params_get_logits_all), 0);
     rb_define_method(rb_cLLaMAContextParams, "embedding=", RUBY_METHOD_FUNC(_llama_context_params_set_embedding), 1);
     rb_define_method(rb_cLLaMAContextParams, "embedding", RUBY_METHOD_FUNC(_llama_context_params_get_embedding), 0);
+    rb_define_method(rb_cLLaMAContextParams, "offload_kqv=", RUBY_METHOD_FUNC(_llama_context_params_set_offload_kqv), 1);
+    rb_define_method(rb_cLLaMAContextParams, "offload_kqv", RUBY_METHOD_FUNC(_llama_context_params_get_offload_kqv), 0);
   }
 
 private:
@@ -991,28 +1068,40 @@ private:
     return UINT2NUM(ptr->params.yarn_orig_ctx);
   }
 
-  // mul_mat_q
-  static VALUE _llama_context_params_set_mul_mat_q(VALUE self, VALUE mul_mat_q) {
+  // type_k
+  static VALUE _llama_context_params_set_type_k(VALUE self, VALUE type_k) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    ptr->params.mul_mat_q = RTEST(mul_mat_q) ? true : false;
-    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
+    ptr->params.type_k = static_cast<enum ggml_type>(NUM2INT(type_k));
+    return INT2NUM(ptr->params.type_k);
   }
 
-  static VALUE _llama_context_params_get_mul_mat_q(VALUE self) {
+  static VALUE _llama_context_params_get_type_k(VALUE self) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
+    return INT2NUM(ptr->params.type_k);
   }
 
-  // f16_kv
-  static VALUE _llama_context_params_set_f16_kv(VALUE self, VALUE f16_kv) {
+  // type_v
+  static VALUE _llama_context_params_set_type_v(VALUE self, VALUE type_v) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    ptr->params.f16_kv = RTEST(f16_kv) ? true : false;
-    return ptr->params.f16_kv ? Qtrue : Qfalse;
+    ptr->params.type_v = static_cast<enum ggml_type>(NUM2INT(type_v));
+    return INT2NUM(ptr->params.type_v);
   }
 
-  static VALUE _llama_context_params_get_f16_kv(VALUE self) {
+  static VALUE _llama_context_params_get_type_v(VALUE self) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return ptr->params.f16_kv ? Qtrue : Qfalse;
+    return INT2NUM(ptr->params.type_v);
+  }
+
+  // mul_mat_q
+  static VALUE _llama_context_params_set_mul_mat_q(VALUE self, VALUE mul_mat_q) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.mul_mat_q = RTEST(mul_mat_q) ? true : false;
+    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_context_params_get_mul_mat_q(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
   }
 
   // logits_all
@@ -1038,6 +1127,18 @@ private:
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
     return ptr->params.embedding ? Qtrue : Qfalse;
   }
+
+  // offload_kqv
+  static VALUE _llama_context_params_set_offload_kqv(VALUE self, VALUE offload_kqv) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.offload_kqv = RTEST(offload_kqv) ? true : false;
+    return ptr->params.offload_kqv ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_context_params_get_offload_kqv(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return ptr->params.offload_kqv ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAContextParams::llama_context_params_type = {
@@ -1232,6 +1333,7 @@ public:
   static void define_class(VALUE outer) {
     rb_cLLaMAModel = rb_define_class_under(outer, "Model", rb_cObject);
     rb_define_alloc_func(rb_cLLaMAModel, llama_model_alloc);
+    rb_define_attr(rb_cLLaMAModel, "params", 1, 0);
     rb_define_method(rb_cLLaMAModel, "initialize", RUBY_METHOD_FUNC(_llama_model_initialize), -1);
     rb_define_method(rb_cLLaMAModel, "empty?", RUBY_METHOD_FUNC(_llama_model_empty), 0);
     rb_define_method(rb_cLLaMAModel, "free", RUBY_METHOD_FUNC(_llama_model_free), 0);
@@ -2352,7 +2454,7 @@ private:
     const float penalty_present = NUM2DBL(kw_values[2]);
 
     llama_sample_repetition_penalties(ctx_ptr->ctx, &(cnd_ptr->array), last_n_tokens_data.data(), last_tokens_size,
-        penalty_repeat, penalty_freq, penalty_present);
+                                      penalty_repeat, penalty_freq, penalty_present);
 
     return Qnil;
   }
@@ -2970,9 +3072,11 @@ static VALUE rb_llama_max_devices(VALUE self) {
 extern "C" void Init_llama_cpp(void) {
   rb_mLLaMACpp = rb_define_module("LLaMACpp");
 
+  RbLLaMABatch::define_class(rb_mLLaMACpp);
   RbLLaMATokenData::define_class(rb_mLLaMACpp);
   RbLLaMATokenDataArray::define_class(rb_mLLaMACpp);
   RbLLaMAModel::define_class(rb_mLLaMACpp);
+  RbLLaMAModelKVOverride::define_class(rb_mLLaMACpp);
   RbLLaMAModelParams::define_class(rb_mLLaMACpp);
   RbLLaMATimings::define_class(rb_mLLaMACpp);
   RbLLaMAContext::define_class(rb_mLLaMACpp);
@@ -3023,6 +3127,10 @@ extern "C" void Init_llama_cpp(void) {
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_INT", INT2NUM(LLAMA_KV_OVERRIDE_INT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_FLOAT", INT2NUM(LLAMA_KV_OVERRIDE_FLOAT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_BOOL", INT2NUM(LLAMA_KV_OVERRIDE_BOOL));
+
   rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_END", INT2NUM(LLAMA_GRETYPE_END));
   rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_ALT", INT2NUM(LLAMA_GRETYPE_ALT));
   rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_RULE_REF", INT2NUM(LLAMA_GRETYPE_RULE_REF));

data/ext/llama_cpp/src/ggml-alloc.c

@@ -168,10 +168,6 @@ static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor *
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
 
-    if (!alloc->measure) {
-        ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
-    }
-
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
 #endif
@@ -237,7 +233,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
 }
 
 ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
 
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -449,7 +445,6 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
 static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
     ggml_tallocr_t alloc = node_tallocr(galloc, view);
 
-    //printf("init_view: %s from src %s\n", view->name, view->view_src->name);
     GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
     if (update_backend) {
         view->backend = view->view_src->backend;
@@ -459,7 +454,7 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
 
     // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
     // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
+    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
         ggml_backend_buffer_init_tensor(alloc->buffer, view);
@@ -765,3 +760,43 @@ size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
 size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
     return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
 }
+
+// utils
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+
+    size_t nbytes = 0;
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL && t->view_src == NULL) {
+            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+    }
+
+    if (nbytes == 0) {
+        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
+
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(tallocr, t);
+            } else {
+                ggml_backend_view_init(buffer, t);
+            }
+        }
+    }
+
+    ggml_tallocr_free(tallocr);
+
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}

data/ext/llama_cpp/src/ggml-alloc.h

@@ -8,6 +8,7 @@ extern "C" {
 
 struct ggml_backend;
 struct ggml_backend_buffer;
+struct ggml_backend_buffer_type;
 
 //
 // Legacy API
@@ -42,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
 // ggml-backend v2 API
 //
 
-// Seperate tensor and graph allocator objects
+// Separate tensor and graph allocator objects
 // This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
 // The original API is kept as a wrapper around the new API
 
@@ -80,6 +81,12 @@ GGML_API void   ggml_gallocr_alloc_graph_n(
                     struct ggml_hash_set hash_set,
                     ggml_tallocr_t * hash_node_talloc);
 
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+
 #ifdef  __cplusplus
 }
 #endif

data/ext/llama_cpp/src/ggml-backend-impl.h

@@ -12,31 +12,50 @@ extern "C" {
     // Backend buffer
     //
 
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+        void     (*free_buffer)(ggml_backend_buffer_t buffer);
+        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        void *   (*get_base)   (ggml_backend_buffer_t buffer);
+        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
     };
 
     struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
-
         size_t size;
     };
 
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
+    ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t      buft,
             struct ggml_backend_buffer_i           iface,
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+
     //
     // Backend
     //
@@ -49,20 +68,17 @@ extern "C" {
         void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        // (optional) asynchroneous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
 
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        // (optional) asynchroneous tensor copy
+        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        void (*synchronize)     (ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
@@ -82,6 +98,15 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+
+    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
 #ifdef  __cplusplus
 }
 #endif