abetlen · abetlen · Jun 29, 2026 · Jun 29, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- feat: update vendored ggml to 0.15.3 and sync SYCL bindings by @abetlen in #177
 - docs: add JupyterLite browser playground by @abetlen in #173
 
 ## [0.0.44]

diff --git a/ggml/ggml.py b/ggml/ggml.py
@@ -308,8 +308,8 @@ def f_(*args: Any, **kwargs: Any):
 
 GGML_VERSION_MAJOR = 0
 GGML_VERSION_MINOR = 15
-GGML_VERSION_PATCH = 2
-GGML_VERSION = "0.15.2"
+GGML_VERSION_PATCH = 3
+GGML_VERSION = "0.15.3"
 
 GGML_ROPE_TYPE_NORMAL = 0
 GGML_ROPE_TYPE_NEOX = 2
@@ -15251,6 +15251,50 @@ def ggml_backend_sycl_split_buffer_type(
     ...
 
 
+# GGML_API GGML_CALL void * ggml_backend_sycl_comm_init(ggml_backend_t * backends, size_t n_backends);
+@ggml_function(
+    "ggml_backend_sycl_comm_init",
+    [ctypes.POINTER(ggml_backend_t_ctypes), ctypes.c_size_t],
+    ctypes.c_void_p,
+    enabled=GGML_USE_SYCL_BACKEND,
+)
+def ggml_backend_sycl_comm_init(
+    backends: CtypesPointer[ggml_backend_t_ctypes],
+    n_backends: Union[ctypes.c_size_t, int],
+    /,
+) -> Optional[ctypes.c_void_p]:
+    ...
+
+
+# GGML_API GGML_CALL void ggml_backend_sycl_comm_free(void * comm_ctx);
+@ggml_function(
+    "ggml_backend_sycl_comm_free",
+    [ctypes.c_void_p],
+    None,
+    enabled=GGML_USE_SYCL_BACKEND,
+)
+def ggml_backend_sycl_comm_free(
+    comm_ctx: Optional[ctypes.c_void_p],
+    /,
+) -> None:
+    ...
+
+
+# GGML_API GGML_CALL bool ggml_backend_sycl_comm_allreduce_tensor(void * comm_ctx, struct ggml_tensor ** tensors);
+@ggml_function(
+    "ggml_backend_sycl_comm_allreduce_tensor",
+    [ctypes.c_void_p, ctypes.POINTER(ctypes.POINTER(ggml_tensor))],
+    ctypes.c_bool,
+    enabled=GGML_USE_SYCL_BACKEND,
+)
+def ggml_backend_sycl_comm_allreduce_tensor(
+    comm_ctx: Optional[ctypes.c_void_p],
+    tensors: CtypesPointer[ggml_tensor_p],
+    /,
+) -> bool:
+    ...
+
+
 # GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 @ggml_function(
     "ggml_backend_sycl_host_buffer_type",

diff --git a/vendor/ggml b/vendor/ggml
+1 −2		CMakeLists.txt
+8 −0		include/ggml-sycl.h
+1 −1		scripts/sync-llama.last
+5 −6		src/ggml-cpu/amx/mmq.cpp
+50 −23		src/ggml-cpu/ops.cpp
+2 −2		src/ggml-cpu/vec.cpp
+90 −46		src/ggml-cuda/binbcast.cu
+35 −29		src/ggml-cuda/cpy.cu
+1 −1		src/ggml-cuda/ggml-cuda.cu
+55 −12		src/ggml-cuda/out-prod.cu
+1 −0		src/ggml-cuda/vendors/hip.h
+1 −0		src/ggml-cuda/vendors/musa.h
+0 −4		src/ggml-hexagon/CMakeLists.txt
+1,365 −1,280		src/ggml-hexagon/ggml-hexagon.cpp
+156 −50		src/ggml-hexagon/htp-opnode.h
+14 −38		src/ggml-hexagon/htp/CMakeLists.txt
+13 −15		src/ggml-hexagon/htp/cmake-toolchain.cmake
+2 −3		src/ggml-hexagon/htp/flash-attn-ops.c
+80 −0		src/ggml-hexagon/htp/hex-common.h
+1 −5		src/ggml-hexagon/htp/hex-dma.h
+1 −56		src/ggml-hexagon/htp/hex-utils.h
+13 −13		src/ggml-hexagon/htp/hmx-flash-attn-ops.c
+0 −2,080		src/ggml-hexagon/htp/hmx-matmul-ops.c
+1,306 −0		src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h
+0 −6		src/ggml-hexagon/htp/hmx-ops.c
+0 −88		src/ggml-hexagon/htp/hmx-ops.h
+9 −3		src/ggml-hexagon/htp/htp-ctx.h
+15 −8		src/ggml-hexagon/htp/htp-ops.h
+2 −1		src/ggml-hexagon/htp/htp_iface.idl
+13 −18		src/ggml-hexagon/htp/hvx-base.h
+1,024 −0		src/ggml-hexagon/htp/hvx-mm-kernels-flat.h
+1,140 −0		src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h
+81 −23		src/ggml-hexagon/htp/main.c
+2,716 −4,104		src/ggml-hexagon/htp/matmul-ops.c
+508 −0		src/ggml-hexagon/htp/matmul-ops.h
+10 −9		src/ggml-hexagon/htp/ssm-conv.c
+0 −4		src/ggml-hexagon/libggml-htp.inf
+9 −13		src/ggml-opencl/ggml-opencl.cpp
+1 −1		src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl
+5 −2		src/ggml-opencl/kernels/norm.cl
+5 −0		src/ggml-sycl/binbcast.cpp
+11 −5		src/ggml-sycl/conv3d.cpp
+155 −53		src/ggml-sycl/element_wise.cpp
+255 −0		src/ggml-sycl/ggml-sycl.cpp
+2 −2		src/ggml-sycl/softmax.cpp
+5 −0		src/ggml-vulkan/CMakeLists.txt
+372 −69		src/ggml-vulkan/ggml-vulkan.cpp
+0 −17		src/ggml-vulkan/vulkan-shaders/clamp.comp
+1 −1		src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
+431 −0		src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp
+0 −17		src/ggml-vulkan/vulkan-shaders/cos.comp
+1 −0		src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+1 −0		src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+25 −0		src/ggml-vulkan/vulkan-shaders/get_rows_back.comp
+4 −7		src/ggml-vulkan/vulkan-shaders/l2_norm.comp
+0 −22		src/ggml-vulkan/vulkan-shaders/leaky_relu.comp
+31 −23		src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+6 −5		src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
+84 −67		src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
+10 −10		src/ggml-vulkan/vulkan-shaders/norm.comp
+0 −17		src/ggml-vulkan/vulkan-shaders/sin.comp
+0 −17		src/ggml-vulkan/vulkan-shaders/sqrt.comp
+0 −17		src/ggml-vulkan/vulkan-shaders/square.comp
+24 −0		src/ggml-vulkan/vulkan-shaders/unary.comp
+64 −28		src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+10 −3		src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
+28 −36		src/ggml-webgpu/ggml-webgpu.cpp
+2 −2		src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl
+45 −39		src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl
+540 −465		src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl
+63 −69		src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl
+14 −9		src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl
+7 −10		src/ggml.c
+73 −8		tests/test-backend-ops.cpp
+61 −28		tests/test-quantize-fns.cpp