cui-llama.rn 0.2.0

package/cpp/ggml-alloc.h

@@ -0,0 +1,76 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+typedef struct lm_ggml_backend_buffer_type * lm_ggml_backend_buffer_type_t;
+typedef struct lm_ggml_backend_buffer * lm_ggml_backend_buffer_t;
+typedef struct lm_ggml_backend * lm_ggml_backend_t;
+
+// Tensor allocator
+struct lm_ggml_tallocr {
+    lm_ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
+
+LM_GGML_API struct lm_ggml_tallocr lm_ggml_tallocr_new(lm_ggml_backend_buffer_t buffer);
+LM_GGML_API void                lm_ggml_tallocr_alloc(struct lm_ggml_tallocr * talloc, struct lm_ggml_tensor * tensor);
+
+// Graph allocator
+/*
+  Example usage:
+    lm_ggml_gallocr_t galloc = lm_ggml_gallocr_new(lm_ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    lm_ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct lm_ggml_cgraph * graph = build_graph(batch);
+    lm_ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", lm_ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    lm_ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   lm_ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   lm_ggml_set_output(): output tensors are never freed and never overwritten
+
+typedef struct lm_ggml_gallocr * lm_ggml_gallocr_t;
+
+LM_GGML_API lm_ggml_gallocr_t lm_ggml_gallocr_new(lm_ggml_backend_buffer_type_t buft);
+LM_GGML_API lm_ggml_gallocr_t lm_ggml_gallocr_new_n(lm_ggml_backend_buffer_type_t * bufts, int n_bufs);
+LM_GGML_API void           lm_ggml_gallocr_free(lm_ggml_gallocr_t galloc);
+
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: lm_ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+LM_GGML_API bool lm_ggml_gallocr_reserve(lm_ggml_gallocr_t galloc, struct lm_ggml_cgraph * graph);
+LM_GGML_API bool lm_ggml_gallocr_reserve_n(
+    lm_ggml_gallocr_t galloc,
+    struct lm_ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
+
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call lm_ggml_gallocr_reserve_n first to set the node buffers)
+LM_GGML_API bool lm_ggml_gallocr_alloc_graph(lm_ggml_gallocr_t galloc, struct lm_ggml_cgraph * graph);
+
+LM_GGML_API size_t lm_ggml_gallocr_get_buffer_size(lm_ggml_gallocr_t galloc, int buffer_id);
+
+// Utils
+// Create a buffer and allocate all the tensors in a lm_ggml_context
+LM_GGML_API struct lm_ggml_backend_buffer * lm_ggml_backend_alloc_ctx_tensors_from_buft(struct lm_ggml_context * ctx, lm_ggml_backend_buffer_type_t buft);
+LM_GGML_API struct lm_ggml_backend_buffer * lm_ggml_backend_alloc_ctx_tensors(struct lm_ggml_context * ctx, lm_ggml_backend_t backend);
+
+#ifdef  __cplusplus
+}
+#endif

package/cpp/ggml-backend-impl.h

@@ -0,0 +1,153 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    // buffer type
+    typedef void * lm_ggml_backend_buffer_type_context_t;
+
+    struct lm_ggml_backend_buffer_type_i {
+        const char *          (*LM_GGML_CALL get_name)        (lm_ggml_backend_buffer_type_t buft);
+        // allocate a buffer of this type
+        lm_ggml_backend_buffer_t (*LM_GGML_CALL alloc_buffer)    (lm_ggml_backend_buffer_type_t buft, size_t size);
+        // tensor alignment
+        size_t                (*LM_GGML_CALL get_alignment)   (lm_ggml_backend_buffer_type_t buft);
+        // max buffer size that can be allocated
+        size_t                (*LM_GGML_CALL get_max_size)    (lm_ggml_backend_buffer_type_t buft);
+        // data size needed to allocate the tensor, including padding
+        size_t                (*LM_GGML_CALL get_alloc_size)  (lm_ggml_backend_buffer_type_t buft, const struct lm_ggml_tensor * tensor);
+        // check if tensor data is in host memory
+        bool                  (*LM_GGML_CALL is_host)         (lm_ggml_backend_buffer_type_t buft);
+    };
+
+    struct lm_ggml_backend_buffer_type {
+        struct lm_ggml_backend_buffer_type_i  iface;
+        lm_ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
+    typedef void * lm_ggml_backend_buffer_context_t;
+
+    struct lm_ggml_backend_buffer_i {
+        const char * (*LM_GGML_CALL get_name)   (lm_ggml_backend_buffer_t buffer);
+        void         (*LM_GGML_CALL free_buffer)(lm_ggml_backend_buffer_t buffer);
+        void *       (*LM_GGML_CALL get_base)   (lm_ggml_backend_buffer_t buffer);
+        void         (*LM_GGML_CALL init_tensor)(lm_ggml_backend_buffer_t buffer, struct lm_ggml_tensor * tensor);
+        void         (*LM_GGML_CALL set_tensor) (lm_ggml_backend_buffer_t buffer,       struct lm_ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*LM_GGML_CALL get_tensor) (lm_ggml_backend_buffer_t buffer, const struct lm_ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*LM_GGML_CALL cpy_tensor) (lm_ggml_backend_buffer_t buffer, const struct lm_ggml_tensor * src, struct lm_ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*LM_GGML_CALL clear)      (lm_ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*LM_GGML_CALL reset)      (lm_ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+    };
+
+    struct lm_ggml_backend_buffer {
+        struct lm_ggml_backend_buffer_i  iface;
+        lm_ggml_backend_buffer_type_t    buft;
+        lm_ggml_backend_buffer_context_t context;
+        size_t size;
+        enum lm_ggml_backend_buffer_usage usage;
+    };
+
+    LM_GGML_CALL lm_ggml_backend_buffer_t lm_ggml_backend_buffer_init(
+                   lm_ggml_backend_buffer_type_t      buft,
+            struct lm_ggml_backend_buffer_i           iface,
+                   lm_ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    // do not use directly, use lm_ggml_backend_tensor_copy instead
+    bool lm_ggml_backend_buffer_copy_tensor(const struct lm_ggml_tensor * src, struct lm_ggml_tensor * dst);
+
+    // buffer that contains a collection of buffers
+    LM_GGML_CALL lm_ggml_backend_buffer_t lm_ggml_backend_multi_buffer_alloc_buffer(lm_ggml_backend_buffer_t * buffers, size_t n_buffers);
+    LM_GGML_CALL bool                  lm_ggml_backend_buffer_is_multi_buffer(lm_ggml_backend_buffer_t buffer);
+    LM_GGML_CALL void                  lm_ggml_backend_multi_buffer_set_usage(lm_ggml_backend_buffer_t buffer, enum lm_ggml_backend_buffer_usage usage);
+
+    //
+    // Backend
+    //
+
+    typedef void * lm_ggml_backend_context_t;
+
+    struct lm_ggml_backend_i {
+        const char * (*LM_GGML_CALL get_name)(lm_ggml_backend_t backend);
+
+        void (*LM_GGML_CALL free)(lm_ggml_backend_t backend);
+
+        // buffer allocation
+        lm_ggml_backend_buffer_type_t (*LM_GGML_CALL get_default_buffer_type)(lm_ggml_backend_t backend);
+
+        // (optional) asynchronous tensor data access
+        void (*LM_GGML_CALL set_tensor_async)(lm_ggml_backend_t backend,       struct lm_ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*LM_GGML_CALL get_tensor_async)(lm_ggml_backend_t backend, const struct lm_ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*LM_GGML_CALL cpy_tensor_async)(lm_ggml_backend_t backend_src, lm_ggml_backend_t backend_dst, const struct lm_ggml_tensor * src, struct lm_ggml_tensor * dst);
+
+        // (optional) complete all pending operations
+        void (*LM_GGML_CALL synchronize)(lm_ggml_backend_t backend);
+
+        // compute graph with a plan (not used currently)
+        // create a new plan for a graph
+        lm_ggml_backend_graph_plan_t (*LM_GGML_CALL graph_plan_create) (lm_ggml_backend_t backend, const struct lm_ggml_cgraph * cgraph);
+        void                      (*LM_GGML_CALL graph_plan_free)   (lm_ggml_backend_t backend, lm_ggml_backend_graph_plan_t plan);
+        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
+        void                      (*LM_GGML_CALL graph_plan_update) (lm_ggml_backend_t backend, lm_ggml_backend_graph_plan_t plan, const struct lm_ggml_cgraph * cgraph);
+        // compute the graph with the plan
+        enum lm_ggml_status          (*LM_GGML_CALL graph_plan_compute)(lm_ggml_backend_t backend, lm_ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan (async)
+        enum lm_ggml_status (*LM_GGML_CALL graph_compute)     (lm_ggml_backend_t backend, struct lm_ggml_cgraph * cgraph);
+
+        // check if the backend can compute an operation
+        bool (*LM_GGML_CALL supports_op)(lm_ggml_backend_t backend, const struct lm_ggml_tensor * op);
+
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*LM_GGML_CALL supports_buft)(lm_ggml_backend_t backend, lm_ggml_backend_buffer_type_t buft);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*LM_GGML_CALL offload_op)(lm_ggml_backend_t backend, const struct lm_ggml_tensor * op);
+
+        // (optional) event synchronization
+        // create a new event that can record events on this backend instance
+        lm_ggml_backend_event_t (*LM_GGML_CALL event_new)         (lm_ggml_backend_t backend);
+        void                 (*LM_GGML_CALL event_free)        (lm_ggml_backend_event_t event);
+        // record an event on the backend instance that created it
+        void                 (*LM_GGML_CALL event_record)      (lm_ggml_backend_event_t event);
+        // wait for an event on on a different backend instance
+        void                 (*LM_GGML_CALL event_wait)        (lm_ggml_backend_t backend, lm_ggml_backend_event_t event);
+        // block until an event is recorded
+        void                 (*LM_GGML_CALL event_synchronize) (lm_ggml_backend_event_t event);
+    };
+
+    struct lm_ggml_backend {
+        lm_ggml_guid_t guid;
+
+        struct lm_ggml_backend_i iface;
+        lm_ggml_backend_context_t context;
+    };
+
+    struct lm_ggml_backend_event {
+        lm_ggml_backend_t backend;
+        void * context;
+    };
+
+    //
+    // Backend registry
+    //
+
+    typedef lm_ggml_backend_t (*LM_GGML_CALL lm_ggml_backend_init_fn)(const char * params, void * user_data);
+
+    LM_GGML_CALL void lm_ggml_backend_register(const char * name, lm_ggml_backend_init_fn init_fn, lm_ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
+#ifdef  __cplusplus
+}
+#endif