diff --git a/ggml/include/ggml-rpc.h b/ggml/include/ggml-rpc.h index ecf0a3a6dc25..00be60f0fbc7 100644 --- a/ggml/include/ggml-rpc.h +++ b/ggml/include/ggml-rpc.h @@ -8,10 +8,10 @@ extern "C" { #define RPC_PROTO_MAJOR_VERSION 8 #define RPC_PROTO_MINOR_VERSION 0 -#define RPC_PROTO_PATCH_VERSION 1 +#define RPC_PROTO_PATCH_VERSION 2 #ifdef __cplusplus -static_assert(GGML_OP_COUNT == 102, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION"); +static_assert(GGML_OP_COUNT == 103, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION"); #endif #define GGML_RPC_MAX_SERVERS 16 diff --git a/ggml/src/ggml-sycl/gemm.hpp b/ggml/src/ggml-sycl/gemm.hpp index c3f6fcaa206a..54a6c7a91f0b 100644 --- a/ggml/src/ggml-sycl/gemm.hpp +++ b/ggml/src/ggml-sycl/gemm.hpp @@ -107,7 +107,10 @@ class DnnlGemmWrapper { args.insert({ DNNL_ARG_WEIGHTS, dnnl::memory(w_md, eng, const_cast(w4)) }); args.insert({ DNNL_ARG_DST, dnnl::memory(pd.dst_desc(), eng, dst) }); args.insert({ DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, dnnl::memory(sc_md, eng, const_cast(scales)) }); - args.insert({ DNNL_ARG_SCRATCHPAD, ctx.get_scratchpad_mem(pd.scratchpad_desc(), eng, q) }); + const auto scratchpad_md = pd.scratchpad_desc(); + ggml_sycl_pool_alloc scratchpad(ctx.pool()); + void * scratchpad_ptr = scratchpad_md.get_size() > 0 ? scratchpad.alloc(scratchpad_md.get_size()) : nullptr; + args.insert({ DNNL_ARG_SCRATCHPAD, dnnl::memory(scratchpad_md, eng, scratchpad_ptr) }); prim.execute(stream, args); } @@ -127,7 +130,10 @@ class DnnlGemmWrapper { args.insert({ DNNL_ARG_SRC, dnnl::memory(act_md, eng, const_cast(act)) }); args.insert({ DNNL_ARG_WEIGHTS, dnnl::memory(w_md, eng, const_cast(w)) }); args.insert({ DNNL_ARG_DST, dnnl::memory(pd.dst_desc(), eng, dst) }); - args.insert({ DNNL_ARG_SCRATCHPAD, ctx.get_scratchpad_mem(pd.scratchpad_desc(), eng, q) }); + const auto scratchpad_md = pd.scratchpad_desc(); + ggml_sycl_pool_alloc scratchpad(ctx.pool()); + void * scratchpad_ptr = scratchpad_md.get_size() > 0 ? scratchpad.alloc(scratchpad_md.get_size()) : nullptr; + args.insert({ DNNL_ARG_SCRATCHPAD, dnnl::memory(scratchpad_md, eng, scratchpad_ptr) }); prim.execute(stream, args); } diff --git a/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h b/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h index 219c477c6170..2506fe9e8064 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h +++ b/ggml/src/ggml-vulkan/ggml-vulkan-push-constants.h @@ -158,6 +158,15 @@ struct vk_op_fwht_push_constants { float scale; }; +struct vk_op_fwht_signed_push_constants { + uint32_t n_rows; + uint32_t src_offset; + uint32_t signs_offset; + uint32_t signs_ne; + uint32_t dst_offset; + float scale; +}; + struct vk_op_dsv4_hc_comb_push_constants { uint32_t n_tokens;