llama : refactor model loader with backend registry

Author: slaren

ggml/include/ggml-backend.h

@@ -169,8 +169,8 @@ extern "C" {
 
 
     // Functions that may be obtained using ggml_backend_reg_get_proc_address
-    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
-    typedef void (*ggml_backend_set_n_threads_t)(ggml_backend_t, int);
+    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(int main_device, const float * tesor_split);
+    typedef void (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
 
     //
     // Backend registry

ggml/include/ggml-cuda.h

@@ -28,7 +28,7 @@ GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

ggml/src/ggml-amx.cpp

@@ -16,12 +16,6 @@
 #if defined(__AMX_INT8__)
 
 // AMX buffer interface
-static const char * ggml_backend_amx_buffer_get_name(ggml_backend_buffer_t buffer) {
-    return "AMX";
-
-    GGML_UNUSED(buffer);
-}
-
 static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
 }
@@ -72,7 +66,6 @@ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
-    /* .get_name        = */ ggml_backend_amx_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_amx_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required

ggml/src/ggml-backend-impl.h

@@ -22,7 +22,7 @@ extern "C" {
         size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
         // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // (optional) check if tensor data is in host memory (defaults to false)
+        // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
     };
 
@@ -37,7 +37,6 @@ extern "C" {
     //
 
     struct ggml_backend_buffer_i {
-        const char * (*get_name)     (ggml_backend_buffer_t buffer);
         // (optional) free the buffer
         void         (*free_buffer)  (ggml_backend_buffer_t buffer);
         // base address of the buffer

ggml/src/ggml-backend.cpp

@@ -34,6 +34,11 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
 }
 
 ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    if (size == 0) {
+        // return a dummy buffer for zero-sized allocations
+        return ggml_backend_buffer_init(buft, {}, NULL, 0);
+    }
+
     return buft->iface.alloc_buffer(buft, size);
 }
 
@@ -89,7 +94,7 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
 }
 
 const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name(buffer);
+    return ggml_backend_buft_name(ggml_backend_buffer_get_type(buffer));
 }
 
 void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -108,6 +113,11 @@ size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
 }
 
 void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    // get_base is optional if the buffer is zero-sized
+    if (buffer->iface.get_base == NULL && buffer->size == 0) {
+        return NULL;
+    }
+
     void * base = buffer->iface.get_base(buffer);
 
     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -238,43 +248,42 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
-    if (!size) {
-        return;
-    }
-
     buf->iface.set_tensor(buf, tensor, data, offset, size);
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
-    if (!size) {
-        return;
-    }
-
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
 GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(buf != NULL && "tensor buffer not set");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-
-    if (!size) {
+    if (size == 0) {
         return;
     }
 
-    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not supported by backend buffer");
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not implemented by backend buffer");
 
     buf->iface.memset_tensor(buf, tensor, value, offset, size);
 }
@@ -713,12 +722,6 @@ ggml_backend_t ggml_backend_init_best(void) {
 
 // backend CPU
 
-static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) {
-    return "CPU";
-
-    GGML_UNUSED(buffer);
-}
-
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
     uintptr_t data = (uintptr_t)buffer->context;
 
@@ -767,7 +770,6 @@ static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
-    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required
@@ -780,7 +782,6 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
 };
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
-    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
     /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required
@@ -799,19 +800,14 @@ static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_ty
 }
 
 static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    auto alloc_size = size;
-    if (alloc_size == 0) {
-        alloc_size = 1;
-    }
-
-    void * data = ggml_aligned_malloc(alloc_size);
+    void * data = ggml_aligned_malloc(size);
 
     if (data == NULL) {
-        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, alloc_size);
+        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, size);
         return NULL;
     }
 
-    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, alloc_size);
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
 }
 
 static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -1315,12 +1311,6 @@ struct ggml_backend_multi_buffer_context {
     size_t n_buffers;
 };
 
-static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
-    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
-
-    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
-}
-
 static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
@@ -1339,7 +1329,6 @@ static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_
 }
 
 static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
-    /* .get_name        = */ ggml_backend_multi_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
     /* .get_base        = */ NULL,
     /* .init_tensor     = */ NULL,
@@ -1368,7 +1357,7 @@ ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer
 }
 
 bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+    return buffer->iface.free_buffer == ggml_backend_multi_buffer_free_buffer;
 }
 
 void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {

ggml/src/ggml-cann.cpp

@@ -489,23 +489,6 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
-/**
- * @brief Retrieve the name associated with a CANN buffer.
- *
- * This function returns the name of a CANN buffer, which is stored in the
- * context of the buffer.
- *
- * @param buffer The CANN buffer whose name is to be retrieved.
- * @return A pointer to a C-string containing the name of the buffer.
- */
-
-static const char* ggml_backend_cann_buffer_get_name(
-    ggml_backend_buffer_t buffer) {
-    return "CANN";
-
-    GGML_UNUSED(buffer);
-}
-
 /**
  * @brief Check if a buffer is a CANN buffer.
  *
@@ -515,9 +498,10 @@ static const char* ggml_backend_cann_buffer_get_name(
  * @param buffer The buffer to check.
  * @return true if the buffer is a CANN buffer, false otherwise.
  */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
 static bool ggml_backend_buffer_is_cann(
     ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
+    return ggml_backend_buft_is_cann(buffer->buft);
 }
 
 /**
@@ -965,7 +949,6 @@ static void ggml_backend_cann_buffer_clear(
  * on a CANN buffer within the backend.
  */
 static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
-    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
@@ -999,9 +982,10 @@ struct ggml_backend_cann_buffer_type_context {
  */
 static const char* ggml_backend_cann_buffer_type_name(
     ggml_backend_buffer_type_t buft) {
-    return "CANN";
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
 
-    GGML_UNUSED(buft);
+    return buft_ctx->name.c_str();
 }
 
 /**