@fugood/llama.node 0.3.2 → 0.3.3

package/src/llama.cpp/ggml/src/ggml-sycl/vecdotq.hpp

@@ -968,8 +968,8 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs[0], signs[0], std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs[1], signs[1], std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
         q8 += 8;
         aux32 >>= 7;
     }
@@ -1009,8 +1009,8 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs0, signs0, std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs1, signs1, std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
         q8 += 8;
     }
     const float d =

package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp

@@ -0,0 +1,138 @@
+#include <sycl/sycl.hpp>
+#include "wkv6.hpp"
+
+constexpr int WKV_BLOCK_SIZE = 64;  // Matching CUDA_WKV_BLOCK_SIZE
+
+// Helper function for the main kernel
+static void rwkv_wkv_f32_kernel(
+    const int B, const int T, const int C, const int H,
+    const float* k, const float* v, const float* r,
+    const float* tf, const float* td, const float* s,
+    float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) {
+
+    const int tid = item_ct1.get_local_id(2);
+    const int bid = item_ct1.get_group(2);
+
+    const int head_size = WKV_BLOCK_SIZE;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    // Set up shared memory pointers
+    float* _k = shared_mem;
+    float* _r = _k + head_size;
+    float* _tf = _r + head_size;
+    float* _td = _tf + head_size;
+
+    // Local state array
+    float state[WKV_BLOCK_SIZE];
+
+    // Load initial state
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
+    }
+
+    // Sync threads before shared memory operations
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+    // Load time-mixing parameters
+    _tf[tid] = tf[head_i * head_size + tid];
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+    // Main sequence processing loop
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid;
+         t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid;
+         t += C) {
+
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+
+        // Load current timestep data to shared memory
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+
+        const float _v = v[t];
+        float y = 0;
+
+        // Process in chunks of 4 for better vectorization
+        sycl::float4 k4, r4, tf4, td4, s4, kv4;
+        #pragma unroll
+        for (int j = 0; j < head_size; j += 4) {
+            // Load data in vec4 chunks
+            k4 = sycl::float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            r4 = sycl::float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            tf4 = sycl::float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            td4 = sycl::float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            // Compute key-value product
+            sycl::float4 kv4 = k4 * _v;
+
+            // Accumulate weighted sum
+            y += sycl::dot(r4, tf4 * kv4 + s4);
+
+            // Update state
+            s4 = s4 * td4 + kv4;
+
+            // Store updated state
+            state[j] = s4.x();
+            state[j+1] = s4.y();
+            state[j+2] = s4.z();
+            state[j+3] = s4.w();
+        }
+
+        dst[t] = y;
+    }
+
+    // Save final state
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
+    }
+}
+
+void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
+    const ggml_tensor* src1, ggml_tensor* dst) {
+
+    const float* k_d = (const float*)dst->src[0]->data;
+    const float* v_d = (const float*)dst->src[1]->data;
+    const float* r_d = (const float*)dst->src[2]->data;
+    const float* tf_d = (const float*)dst->src[3]->data;
+    const float* td_d = (const float*)dst->src[4]->data;
+    const float* s_d = (const float*)dst->src[5]->data;
+    float* dst_d = (float*)dst->data;
+
+    const int64_t B = dst->src[5]->ne[1];
+    const int64_t T = dst->src[0]->ne[3];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[2];
+
+    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == WKV_BLOCK_SIZE); // The current sycl kernel is designed for RWKV6, HEAD_SIZE == 64
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    // Calculate execution configuration
+    const size_t shared_mem_size = WKV_BLOCK_SIZE * 4 * sizeof(float); // For k, r, tf, td
+    sycl::range<3> block_dims(1, 1, C / H);
+    sycl::range<3> grid_dims(1, 1, B * H);
+
+    // Submit kernel
+    stream->submit([&](sycl::handler& cgh) {
+        sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                rwkv_wkv_f32_kernel(
+                    B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
+                    item_ct1, shared_mem_acc.get_pointer()
+                );
+            });
+    });
+}

package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_WKV6_HPP
+#define GGML_SYCL_WKV6_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst);
+
+
+#endif // GGML_SYCL_WKV6_HPP

package/src/llama.cpp/ggml/src/ggml-threading.cpp

@@ -0,0 +1,12 @@
+#include "ggml-threading.h"
+#include <mutex>
+
+std::mutex ggml_critical_section_mutex;
+
+void ggml_critical_section_start() {
+    ggml_critical_section_mutex.lock();
+}
+
+void ggml_critical_section_end(void) {
+    ggml_critical_section_mutex.unlock();
+}

package/src/llama.cpp/ggml/src/ggml-threading.h

@@ -0,0 +1,12 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ggml_critical_section_start(void);
+void ggml_critical_section_end(void);
+
+#ifdef __cplusplus
+}
+#endif

package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt

@@ -0,0 +1,78 @@
+find_package(Vulkan COMPONENTS glslc REQUIRED)
+
+if (Vulkan_FOUND)
+    message(STATUS "Vulkan found")
+
+    add_library(ggml-vulkan
+                ggml-vulkan.cpp
+                ../../include/ggml-vulkan.h
+                )
+
+    target_link_libraries(ggml-vulkan PRIVATE ggml-base Vulkan::Vulkan)
+    target_include_directories(ggml-vulkan PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+
+    # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+    # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+    if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+        add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+    endif()
+
+    if (GGML_VULKAN_CHECK_RESULTS)
+        add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+    endif()
+
+    if (GGML_VULKAN_DEBUG)
+        add_compile_definitions(GGML_VULKAN_DEBUG)
+    endif()
+
+    if (GGML_VULKAN_MEMORY_DEBUG)
+        add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
+    endif()
+
+    if (GGML_VULKAN_SHADER_DEBUG_INFO)
+        add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+    endif()
+
+    if (GGML_VULKAN_PERF)
+        add_compile_definitions(GGML_VULKAN_PERF)
+    endif()
+
+    if (GGML_VULKAN_VALIDATE)
+        add_compile_definitions(GGML_VULKAN_VALIDATE)
+    endif()
+
+    if (GGML_VULKAN_RUN_TESTS)
+        add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+    endif()
+
+    add_subdirectory(vulkan-shaders)
+
+    set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
+    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
+    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
+    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
+    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+
+    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
+
+    add_custom_command(
+        OUTPUT ${_ggml_vk_header}
+                ${_ggml_vk_source}
+
+        COMMAND ${_ggml_vk_genshaders_cmd}
+            --glslc      ${Vulkan_GLSLC_EXECUTABLE}
+            --input-dir  ${_ggml_vk_input_dir}
+            --output-dir ${_ggml_vk_output_dir}
+            --target-hpp ${_ggml_vk_header}
+            --target-cpp ${_ggml_vk_source}
+            --no-clean
+
+        DEPENDS ${_ggml_vk_shader_deps}
+        COMMENT "Generate vulkan shaders"
+    )
+
+    target_sources(ggml-vulkan PRIVATE ${_ggml_vk_source} ${_ggml_vk_header})
+
+else()
+    message(WARNING "Vulkan not found")
+endif()