Implementation:Ggml org Ggml Sycl backend api

Metadata

Field	Value
Page Type	Implementation (API Doc)
Knowledge Sources	GGML
Domains	ML_Infrastructure, Tensor_Computing, GPU_Computing
Last Updated	2025-05-15 12:00 GMT

Overview

Public C header declaring the Intel SYCL backend interface for running tensor operations on Intel GPUs and other SYCL-compatible accelerators.

Description

ggml-sycl.h declares the SYCL backend's public API (49 lines). Licensed under MIT by Intel Corporation. It provides:

1. Constants: GGML_SYCL_NAME = "SYCL" and GGML_SYCL_MAX_DEVICES = 48 (the highest device limit among GGML backends).
2. Backend initialization: ggml_backend_sycl_init(device) creates a backend for a specific SYCL device by index.
3. Type checking: ggml_backend_is_sycl(backend) verifies if a backend is SYCL-based.
4. Buffer types:
   • ggml_backend_sycl_buffer_type(device) -- device-specific buffer type
   • ggml_backend_sycl_split_buffer_type(tensor_split) -- split buffer type that distributes matrices by rows across multiple devices for tensor parallelism
   • ggml_backend_sycl_host_buffer_type() -- pinned host buffer for faster CPU-GPU transfers
5. Device enumeration and queries:
   • ggml_backend_sycl_print_sycl_devices -- prints all available SYCL devices
   • ggml_backend_sycl_get_gpu_list -- retrieves GPU device IDs
   • ggml_backend_sycl_get_device_description -- gets human-readable device description
   • ggml_backend_sycl_get_device_count -- returns number of available SYCL devices
   • ggml_backend_sycl_get_device_memory -- queries free and total device memory
6. Registration: ggml_backend_sycl_reg() returns the registration handle.

Note: SYCL does not support host memory registration (commented-out declarations preserved for reference).

Usage

Include this header to use the SYCL backend for inference on Intel GPUs via the oneAPI programming model. The backend supports multi-device tensor parallelism via split buffer types.

Code Reference

Source Location

GGML repo, file: include/ggml-sycl.h (49 lines).

Signatures

#define GGML_SYCL_NAME "SYCL"
#define GGML_SYCL_MAX_DEVICES 48

GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int * id_list, int max_len);
GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API int  ggml_backend_sycl_get_device_count(void);
GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t * free, size_t * total);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);

Import

#include "ggml-sycl.h"

I/O Contract

Inputs

Parameter	Type	Required	Description
device	int	Yes	SYCL device index (0-based) for backend initialization, buffer type, and device queries.
backend	ggml_backend_t	Yes	Backend handle for type checking.
tensor_split	const float *	Yes	Array of split ratios across devices (for split_buffer_type).
id_list	int *	Yes	Output array for GPU device IDs (for get_gpu_list).
max_len	int	Yes	Maximum number of entries to write to id_list.

Outputs

Output	Type	Description
Backend handle	ggml_backend_t	Initialized SYCL backend for the specified device.
Type check	bool	true if the backend is SYCL-based.
Buffer type	ggml_backend_buffer_type_t	Device, split, or host buffer type handle.
Device count	int	Number of available SYCL devices.
Device memory	via output params	Free and total memory on the specified device.

Usage Examples

#include "ggml-sycl.h"

// List available SYCL devices
ggml_backend_sycl_print_sycl_devices();

int n_devices = ggml_backend_sycl_get_device_count();

// Initialize backend on device 0
ggml_backend_t backend = ggml_backend_sycl_init(0);

// Query device memory
size_t free_mem, total_mem;
ggml_backend_sycl_get_device_memory(0, &free_mem, &total_mem);

// For multi-GPU: create split buffer type
float tensor_split[2] = { 0.5f, 0.5f };  // 50/50 split across 2 devices
ggml_backend_buffer_type_t split_buft = ggml_backend_sycl_split_buffer_type(tensor_split);

Related Pages

• Principle:Ggml_org_Ggml_Backend_Interface
• Principle:Ggml_org_Ggml_Backend_Initialization
• Principle:Ggml_org_Ggml_Backend_Scheduling