whispercpp 1.3.5 → 1.3.7

data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp

@@ -15,6 +15,7 @@
 
 #include "dpct/helper.hpp"
 #include "ggml.h"
+#include "type.hpp"
 #include "quants.hpp"
 
 typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1,
@@ -31,6 +32,18 @@ static __dpct_inline__ int get_int_b1(const void * x, const int & i32) {
     return x32;
 }
 
+static __dpct_inline__ int get_int_b2(const void * x, const int & i32) {
+    const uint16_t * x16 = (const uint16_t *) x; // assume at least 2 byte alignment
+
+    int x32  = x16[2*i32 + 0] <<  0;
+    x32     |= x16[2*i32 + 1] << 16;
+
+    return x32;
+}
+
+static __dpct_inline__ int get_int_b4(const void * x, const int & i32) {
+    return ((const int *) x)[i32]; // assume at least 4 byte alignment
+}
 
 static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) {
   const uint16_t* x16 =
@@ -72,6 +85,32 @@ static __dpct_inline__ int get_int_from_uint8_aligned(
       (const int*)(x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
 }
 
+static __dpct_inline__ int byte_sub_4(const int a, const int b) {
+  const uint32_t ua = static_cast<uint32_t>(a);
+  const uint32_t ub = static_cast<uint32_t>(b);
+  return static_cast<int>(((ua | 0x80808080u) - ub) ^ 0x80808080u);
+}
+
+static __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq_scalar(
+    const int vl, const int vh, const int u0, const int u1, const int8_t sc0,
+    const int8_t sc1, const float d, const float d80, const float d81) {
+    static_assert(QR6_K == 2, "q6_K MMVQ scalar fast path assumes QR6_K == 2");
+
+    const int vil0 = (vl >> 0) & 0x0F0F0F0F;
+    const int vih0 = ((vh >> 0) << 4) & 0x30303030;
+    const int vi0 = byte_sub_4(vil0 | vih0, 0x20202020);
+
+    const int vil1 = (vl >> 4) & 0x0F0F0F0F;
+    const int vih1 = ((vh >> 4) << 4) & 0x30303030;
+    const int vi1 = byte_sub_4(vil1 | vih1, 0x20202020);
+
+    const float sumf =
+        d80 * (dpct::dp4a(vi0, u0, 0) * sc0) +
+        d81 * (dpct::dp4a(vi1, u1, 0) * sc1);
+
+    return d * sumf;
+}
+
 static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4,
                                                   const uint8_t *values,
                                                   int &val1, int &val2) {
@@ -266,24 +305,8 @@ vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh,
                             const int *__restrict__ u,
                             const int8_t *__restrict__ scales, const float &d,
                             const float *__restrict__ d8) {
-
-    float sumf = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR6_K; ++i) {
-        const int sc = scales[4*i];
-
-        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
-
-        const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
-
-        const int vi = dpct::vectorized_binary<sycl::char4>(
-            (vil | vih), 0x20202020, dpct::sub_sat()); // vi = (vil | vih) - 32
-
-        sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product
-    }
-
-    return d*sumf;
+    return vec_dot_q6_K_q8_1_impl_mmvq_scalar(
+        vl, vh, u[0], u[1], scales[0], scales[4], d, d8[0], d8[1]);
 }
 
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
@@ -338,6 +361,74 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     };
 };
 
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q8_0> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q8_0;
+
+    using q8_0_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q8_0>;
+    using q8_0_traits = typename q8_0_block::traits;
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t * base = static_cast<const uint8_t *>(vbq);
+        const int8_t *  qs   = reinterpret_cast<const int8_t *>(base + ibx_offset.first);
+        const ggml_half  d   = *reinterpret_cast<const ggml_half *>(base + d_offset.first);
+
+        int v[q8_0_traits::vdr_mmvq];
+        int u[q8_0_traits::vdr_mmvq];
+
+#pragma unroll
+        for (size_t i = 0; i < q8_0_traits::vdr_mmvq; ++i) {
+            v[i] = get_int_from_int8(qs, iqs + i);
+            u[i] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i);
+        }
+
+        int sumi = 0;
+#pragma unroll
+        for (size_t i = 0; i < q8_0_traits::vdr_mmvq; ++i) {
+            sumi = dpct::dp4a(v[i], u[i], sumi);
+        }
+
+        const sycl::half2 ds_values = *q8_1_ds;
+        return static_cast<float>(d) * static_cast<float>(ds_values[0]) * sumi;
+    }
+};
+
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q3_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q3_K;
+
+    using q3_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q3_K>;
+    using q3_k_traits = typename q3_k_block::traits;
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t *  base   = static_cast<const uint8_t *>(vbq);
+        const uint8_t *  qs     = base + ibx_offset.first;
+        const uint8_t *  hmask  = base + ibx_offset.second;
+        const uint8_t *  scales = base + d_offset.first;
+        const ggml_half  d      = *reinterpret_cast<const ggml_half *>(base + d_offset.second);
+
+        const int bq8_offset   = QR3_K * (iqs / (QI3_K / 2));
+        const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1 / 2);
+
+        const int vl = get_int_from_uint8(qs, iqs);
+        const int vh = ~get_int_from_uint8(hmask, iqs % (QI3_K / 2)) >> bq8_offset;
+
+        int   u[QR3_K];
+        float d8[QR3_K];
+
+#pragma unroll
+        for (int i = 0; i < QR3_K; ++i) {
+            const int8_t * quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            u[i]                          = get_int_from_int8_aligned(quant_base_ptr, iqs % QI8_1);
+            d8[i]                         = (*(q8_1_ds + bq8_offset + i))[0];
+        }
+
+        return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, scales, scale_offset, static_cast<float>(d), d8);
+    }
+};
+
 static inline float vec_dot_q4_K_q8_1_common(const int * __restrict__ q4, const uint16_t * __restrict__ scales,
                                              const ggml_half2 & dm, const block_q8_1 * __restrict__ bq8_1,
                                              const int &        iqs) {
@@ -428,32 +519,76 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     }
 };
 
-template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
-    static constexpr ggml_type gtype = GGML_TYPE_Q6_K;
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q5_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q5_K;
 
-    using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
-    using q6_k_traits = typename q6_k_block::traits;
+    using q5_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q5_K>;
+    using q5_k_traits = typename q5_k_block::traits;
 
-    __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int vl, const int vh, const int * __restrict__ u,
-                                                      const int8_t * __restrict__ scales, const float d,
-                                                      const float * __restrict__ d8) {
-        float sumf = 0.0f;
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
+        const uint8_t *    qs             = base + ibx_offset.first;   // low 4 bits
+        const uint8_t *    qh_base        = base + ibx_offset.second;  // high bit
+        const uint8_t *    scs            = base + d_offset.first;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
 
-#pragma unroll
-        for (int i = 0; i < QR6_K; ++i) {
-            const int sc = scales[4 * i];
+        const int        bq8_offset = QR5_K * ((iqs / 2) / (QI8_1 / 2));
+        const int *      ql_ptr     = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
+        const int *      qh_ptr     = (const int *) (qh_base + 4 * ((iqs / 2) % 4));
+        const uint16_t * scales     = (const uint16_t *) scs;
 
-            const int vil = (vl >> (4 * i)) & 0x0F0F0F0F;
+        int   vl[2];
+        int   vh[2];
+        int   u[2 * QR5_K];
+        float d8[QR5_K];
 
-            const int vih = ((vh >> (4 * i)) << 4) & 0x30303030;
+        vl[0] = ql_ptr[0];
+        vl[1] = ql_ptr[4];
 
-            const int vi = dpct::vectorized_binary<sycl::char4>((vil | vih), 0x20202020,
-                                                                dpct::sub_sat());  // vi = (vil | vih) - 32
+        vh[0] = qh_ptr[0] >> bq8_offset;
+        vh[1] = qh_ptr[4] >> bq8_offset;
 
-            sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc);                        // SIMD dot product
+        uint16_t  aux[2];
+        const int j = (QR5_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+        if (j < 2) {
+            aux[0] = scales[j + 0] & 0x3f3f;
+            aux[1] = scales[j + 2] & 0x3f3f;
+        } else {
+            aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+            aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
         }
 
-        return d * sumf;
+        const uint8_t * sc = (const uint8_t *) aux;
+        const uint8_t * m  = sc + 2;
+
+        for (int i = 0; i < QR5_K; ++i) {
+            const int8_t* quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            sycl::half2 ds_values = *(q8_1_ds + bq8_offset + i);
+
+            d8[i]                   = ds_values[0];
+
+            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
+            u[2 * i + 0]   = q8[0];
+            u[2 * i + 1]   = q8[4];
+        }
+
+        return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, *dms, d8);
+    }
+};
+
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q6_K;
+
+    using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
+    using q6_k_traits = typename q6_k_block::traits;
+
+    __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int vl, const int vh, const int * __restrict__ u,
+                                                      const int8_t * __restrict__ scales, const float d,
+                                                      const float * __restrict__ d8) {
+        return vec_dot_q6_K_q8_1_impl_mmvq_scalar(
+            vl, vh, u[0], u[1], scales[0], scales[4], d, d8[0], d8[1]);
     }
 
     __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
@@ -474,16 +609,15 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
 
         const int8_t * scs = scales + scale_offset;
 
-        int   u[QR6_K];
-        float d8[QR6_K];
+        const int u0 = get_int_from_int8_aligned(
+            q8_1_quant_ptr + bq8_offset * QK8_1, iqs % QI8_1);
+        const int u1 = get_int_from_int8_aligned(
+            q8_1_quant_ptr + (bq8_offset + 2) * QK8_1, iqs % QI8_1);
+        const float d80 = (*(q8_1_ds + bq8_offset + 0))[0];
+        const float d81 = (*(q8_1_ds + bq8_offset + 2))[0];
 
-#pragma unroll
-        for (int i = 0; i < QR6_K; ++i) {
-            u[i] = get_int_from_int8_aligned(q8_1_quant_ptr + (bq8_offset + 2 * i) * QK8_1, iqs % QI8_1);
-            const sycl::half2 ds_values = *(q8_1_ds + bq8_offset + 2 * i);
-            d8[i]                       = ds_values[0];
-        }
-        return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scs, *d, d8);
+        return vec_dot_q6_K_q8_1_impl_mmvq_scalar(
+            vl, vh, u0, u1, scs[0], scs[4], *d, d80, d81);
     }
 };
 #define VDR_Q4_0_Q8_1_MMVQ 2
@@ -650,6 +784,19 @@ static __dpct_inline__ float vec_dot_q8_0_q8_1_impl(const int *v, const int *u,
     return d8_0*d8_1 * sumi;
 }
 
+template <typename T, int vdr>
+static __dpct_inline__ T vec_dot_q8_0_q8_1_impl(const int * v, const int * u, const T & d8_0, const T & d8_1) {
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = ggml_sycl_dp4a(v[i], u[i], sumi);
+    }
+
+    return d8_0*d8_1 * ((T) sumi);
+}
+
 template <int vdr>
 static __dpct_inline__ float vec_dot_q8_1_q8_1_impl(const int *v, const int *u,
                                                     const sycl::half2 &dm8,
@@ -742,6 +889,35 @@ static __dpct_inline__ float vec_dot_mxfp4_q8_1(const void * __restrict__ vbq,
     return d * sumi;
 }
 
+#define VDR_NVFP4_Q8_1_MMVQ 4
+#define VDR_NVFP4_Q8_1_MMQ  8
+
+static __dpct_inline__ float vec_dot_nvfp4_q8_1(const void * __restrict__ vbq,
+                                                const block_q8_1 * __restrict__ bq8_1,
+                                                const int32_t & iqs) {
+    const block_nvfp4 * bq4 = (const block_nvfp4 *) vbq;
+    float sum = 0.0f;
+#pragma unroll
+    for (int i = 0; i < VDR_NVFP4_Q8_1_MMVQ/2; i++) {
+        const int32_t iqs0 = iqs + 2*i;
+        const int32_t iqs1 = iqs0 + 1;
+        const int32_t is = iqs0 >> 1;
+        const sycl::int2   v0   = get_int_from_table_16(get_int_b4(bq4->qs, iqs0), kvalues_mxfp4);
+        const sycl::int2   v1   = get_int_from_table_16(get_int_b4(bq4->qs, iqs1), kvalues_mxfp4);
+        const block_q8_1 * bq8 = bq8_1 + (is >> 1);
+        const int32_t i8 = ((is & 1) << 2);
+
+        int sumi = ggml_sycl_dp4a(v0.x(), get_int_b4(bq8->qs, i8 + 0), 0);
+        sumi     = ggml_sycl_dp4a(v0.y(), get_int_b4(bq8->qs, i8 + 2), sumi);
+        sumi     = ggml_sycl_dp4a(v1.x(), get_int_b4(bq8->qs, i8 + 1), sumi);
+        sumi     = ggml_sycl_dp4a(v1.y(), get_int_b4(bq8->qs, i8 + 3), sumi);
+
+        const float d = ggml_sycl_ue4m3_to_fp32(bq4->d[is]) * (bq8->ds)[0];
+        sum += d * float(sumi);
+    }
+
+    return sum;
+}
 
 static __dpct_inline__ float
 vec_dot_q5_0_q8_1(const void *__restrict__ vbq,
@@ -1020,16 +1196,15 @@ vec_dot_q6_K_q8_1(const void *__restrict__ vbq,
 
     const int8_t * scales = bq6_K->scales + scale_offset;
 
-    int    u[QR6_K];
-    float d8[QR6_K];
-
-#pragma unroll
-    for (int i = 0; i < QR6_K; ++i) {
-        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + 2 * i].ds[0];
-    }
+    const int u0 = get_int_from_int8_aligned(
+        bq8_1[bq8_offset + 0].qs, iqs % QI8_1);
+    const int u1 = get_int_from_int8_aligned(
+        bq8_1[bq8_offset + 2].qs, iqs % QI8_1);
+    const float d80 = bq8_1[bq8_offset + 0].ds[0];
+    const float d81 = bq8_1[bq8_offset + 2].ds[0];
 
-    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
+    return vec_dot_q6_K_q8_1_impl_mmvq_scalar(
+        vl, vh, u0, u1, scales[0], scales[4], bq6_K->d, d80, d81);
 }
 
 

data/ext/sources/ggml/src/ggml-sycl/wkv.cpp

@@ -1,7 +1,7 @@
 #include <sycl/sycl.hpp>
 #include "wkv.hpp"
 
-constexpr int WKV_BLOCK_SIZE = 64;  // Matching CUDA_WKV_BLOCK_SIZE
+constexpr int WKV_BLOCK_SIZE = 64;
 
 // Helper function for the main kernel
 template <int block_size>

data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt

@@ -0,0 +1,70 @@
+cmake_minimum_required(VERSION 3.19)
+cmake_policy(SET CMP0114 NEW)
+
+include(ExternalProject)
+
+message(STATUS "Including the VirtGPU/Virglrenderer API Remoting")
+
+# Download venus_hw.h from virglrenderer repository
+ExternalProject_Add(
+    venus_hw_header
+    URL https://gitlab.freedesktop.org/virgl/virglrenderer/-/raw/virglrenderer-1.2.0/src/venus_hw.h
+    DOWNLOAD_NO_EXTRACT YES
+    DOWNLOAD_DIR ${CMAKE_CURRENT_SOURCE_DIR}/include
+    DOWNLOAD_NAME venus_hw.h
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+    INSTALL_COMMAND ""
+    LOG_DOWNLOAD ON
+)
+
+if (NOT GGML_VIRTGPU_BACKEND STREQUAL "ONLY")
+    message(STATUS "Enable the VirtGPU/Virglrenderer API Remoting frontend library")
+
+    find_package(PkgConfig REQUIRED)
+    pkg_check_modules(DRM REQUIRED libdrm)
+    if (NOT GGML_BACKEND_DL)
+      # cannot simply use USE_VIRTGPU, as in the 'else()' case the
+      # frontend isn't compiled
+      target_compile_definitions(ggml PUBLIC "GGML_USE_VIRTGPU_FRONTEND")
+    endif()
+
+    ggml_add_backend_library(ggml-virtgpu
+                             ggml-backend-buffer.cpp
+                             ggml-backend.cpp
+                             ggml-backend-device.cpp
+                             ggml-backend-reg.cpp
+                             ggml-backend-buffer-type.cpp
+                             virtgpu-apir.h
+                             virtgpu-forward.gen.h
+                             virtgpu.cpp
+                             virtgpu-shm.cpp
+                             virtgpu-utils.cpp
+                             virtgpu-forward-device.cpp
+                             virtgpu-forward-buffer-type.cpp
+                             virtgpu-forward-buffer.cpp
+                             virtgpu-forward-backend.cpp
+                             virtgpu-forward-impl.h
+                             apir_cs_ggml-rpc-front.cpp
+                             ../../include/ggml-virtgpu.h)
+
+    target_include_directories(ggml-virtgpu PUBLIC /usr/include/libdrm/)
+
+    target_link_libraries(ggml-virtgpu PUBLIC ${DRM_LIBRARIES})
+    target_include_directories(ggml-virtgpu PUBLIC ${DRM_INCLUDE_DIRS})
+    target_compile_options(ggml-virtgpu PUBLIC ${DRM_CFLAGS_OTHER})
+
+    target_include_directories(ggml-virtgpu PUBLIC ./include)
+    target_include_directories(ggml-virtgpu PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
+
+    # Ensure venus_hw.h is downloaded before building ggml-virtgpu
+    add_dependencies(ggml-virtgpu venus_hw_header)
+
+    target_compile_options(ggml-virtgpu PRIVATE -std=c++20)
+else()
+    message(STATUS "Not building the VirtGPU/Virglrenderer API Remoting frontend library")
+endif()
+
+if (NOT GGML_VIRTGPU_BACKEND STREQUAL "OFF")
+    add_subdirectory("backend")
+endif()

data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp

@@ -0,0 +1,87 @@
+#include "backend/shared/apir_cs_rpc.h"
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "ggml-remoting.h"
+
+#include <cinttypes>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) {
+    apir_rpc_tensor result;
+    result.id   = reinterpret_cast<uint64_t>(tensor);
+    result.type = tensor->type;
+    if (tensor->buffer) {
+        ggml_backend_buffer_t buffer = tensor->buffer;
+
+        result.buffer = BUFFER_TO_HOST_HANDLE(buffer);
+    } else {
+        result.buffer = 0;
+    }
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result.ne[i] = tensor->ne[i];
+        result.nb[i] = tensor->nb[i];
+    }
+    result.op = tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result.op_params[i] = tensor->op_params[i];
+    }
+    result.flags = tensor->flags;
+    for (uint32_t i = 0; i < GGML_MAX_SRC; i++) {
+        result.src[i] = reinterpret_cast<uint64_t>(tensor->src[i]);
+    }
+    result.view_src  = reinterpret_cast<uint64_t>(tensor->view_src);
+    result.view_offs = tensor->view_offs;
+    result.data      = reinterpret_cast<uint64_t>(tensor->data);
+    if (tensor->data) {
+        if (!tensor->buffer) {
+            GGML_ABORT("%s: tensor has data but not buffer", __func__);
+        }
+        // tensor->data is serialized as an offset to the buffer base address
+        result.data -= reinterpret_cast<uint64_t>(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base);
+    }
+    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
+    return result;
+}
+
+void apir_add_tensor(ggml_tensor *                       tensor,
+                     std::vector<apir_rpc_tensor> &      tensors,
+                     std::unordered_set<ggml_tensor *> & visited) {
+    if (tensor == nullptr) {
+        return;
+    }
+    if (visited.find(tensor) != visited.end()) {
+        return;
+    }
+    visited.insert(tensor);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        apir_add_tensor(tensor->src[i], tensors, visited);
+    }
+    apir_add_tensor(tensor->view_src, tensors, visited);
+    tensors.push_back(apir_serialize_tensor(tensor));
+}
+
+void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
+    uint32_t                          n_nodes = cgraph->n_nodes;
+    std::vector<apir_rpc_tensor>      tensors;
+    std::unordered_set<ggml_tensor *> visited;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        apir_add_tensor(cgraph->nodes[i], tensors, visited);
+    }
+    // serialization format:
+    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(apir_rpc_tensor)) |
+    uint32_t n_tensors = tensors.size();
+    int      output_size =
+        sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(apir_rpc_tensor);
+    output.resize(output_size, 0);
+    memcpy(output.data(), &n_nodes, sizeof(n_nodes));
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t));
+    }
+    uint32_t * out_ntensors = (uint32_t *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
+    *out_ntensors           = n_tensors;
+    apir_rpc_tensor * out_tensors =
+        (apir_rpc_tensor *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t));
+    memcpy(out_tensors, tensors.data(), n_tensors * sizeof(apir_rpc_tensor));
+}

data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt

@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.19)
+cmake_policy(SET CMP0114 NEW)
+
+message(STATUS "Enable the VirtGPU/Virglrenderer backend library")
+
+ggml_add_backend_library(ggml-virtgpu-backend
+                         backend.cpp
+                         backend-dispatched.cpp
+                         backend-dispatched-backend.cpp
+                         backend-dispatched-device.cpp
+                         backend-dispatched-buffer.cpp
+                         backend-dispatched-buffer-type.cpp
+                         shared/api_remoting.h
+                         shared/apir_backend.h
+                         shared/apir_cs.h
+                         apir_cs_ggml-rpc-back.cpp)
+
+target_compile_options(ggml-virtgpu-backend PRIVATE -std=c++20)
+
+# Add include directory for ggml-backend-impl.h and other core headers
+target_include_directories(ggml-virtgpu-backend PRIVATE ../..)

data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp

@@ -0,0 +1,115 @@
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "shared/apir_cs_rpc.h"
+
+#include <cinttypes>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+std::unordered_set<ggml_backend_buffer_t> backend_buffers;
+
+void apir_track_backend_buffer(ggml_backend_buffer_t buffer) {
+    backend_buffers.insert(buffer);
+}
+
+bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer) {
+    auto it = backend_buffers.find(buffer);
+    if (it == backend_buffers.end()) {
+        return false;
+    }
+
+    backend_buffers.erase(it);
+    return true;
+}
+
+std::unordered_set<ggml_backend_buffer_t> apir_get_track_backend_buffers() {
+    return backend_buffers;
+}
+
+ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor) {
+    ggml_tensor * result =
+        ggml_new_tensor_4d(ctx, (ggml_type) tensor->type, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = tensor->nb[i];
+    }
+    result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
+    if (result->buffer && backend_buffers.find(result->buffer) == backend_buffers.end()) {
+        printf("WARNING: HOST BUFFER NOT FOUND | %p\n", (void *) result->buffer);
+        result->buffer = nullptr;
+    }
+
+    uint64_t tensor_data = tensor->data;
+    if (result->buffer) {
+        // require that the tensor data does not go beyond the buffer end
+        uint64_t tensor_size  = (uint64_t) ggml_nbytes(result);
+        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+        uint64_t buffer_size  = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+
+        // tensor->data is serialized as an offset to the buffer base address
+        tensor_data += buffer_start;
+
+        GGML_ASSERT(tensor_data + tensor_size >= tensor_data);  // check for overflow
+        GGML_ASSERT(tensor_data >= buffer_start && tensor_data + tensor_size <= buffer_start + buffer_size);
+    }
+
+    result->op = (ggml_op) tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result->op_params[i] = tensor->op_params[i];
+    }
+    result->flags = tensor->flags;
+    result->data  = reinterpret_cast<void *>(tensor_data);
+    ggml_set_name(result, tensor->name);
+    return result;
+}
+
+ggml_tensor * apir_create_node(uint64_t                                                      id,
+                               ggml_context *                                                ctx,
+                               const std::unordered_map<uint64_t, const apir_rpc_tensor *> & tensor_ptrs,
+                               std::unordered_map<uint64_t, ggml_tensor *> &                 tensor_map) {
+    if (id == 0) {
+        return nullptr;
+    }
+    if (tensor_map.find(id) != tensor_map.end()) {
+        return tensor_map[id];
+    }
+    const apir_rpc_tensor * tensor = tensor_ptrs.at(id);
+    ggml_tensor *           result = apir_deserialize_tensor(ctx, tensor);
+    if (result == nullptr) {
+        return nullptr;
+    }
+    tensor_map[id] = result;
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        result->src[i] = apir_create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+    }
+    result->view_src  = apir_create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+    result->view_offs = tensor->view_offs;
+    return result;
+}
+
+ggml_cgraph * apir_deserialize_graph(uint32_t                n_nodes,
+                                     uint32_t                n_tensors,
+                                     const apir_rpc_tensor * tensors,
+                                     const uint64_t *        nodes) {
+    size_t buf_size = ggml_tensor_overhead() * (n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+    ggml_init_params params = {
+        /*.mem_size   =*/buf_size,
+        /*.mem_buffer =*/NULL,
+        /*.no_alloc   =*/true,
+    };
+    ggml_context * ctx   = ggml_init(params);
+    ggml_cgraph *  graph = ggml_new_graph_custom(ctx, n_nodes, false);
+    graph->n_nodes       = n_nodes;
+    std::unordered_map<uint64_t, const apir_rpc_tensor *> tensor_ptrs;
+    for (uint32_t i = 0; i < n_tensors; i++) {
+        tensor_ptrs[tensors[i].id] = &tensors[i];
+    }
+    std::unordered_map<uint64_t, ggml_tensor *> tensor_map;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        int64_t id;
+        memcpy(&id, &nodes[i], sizeof(id));
+        graph->nodes[i] = apir_create_node(id, ctx, tensor_ptrs, tensor_map);
+    }
+
+    return graph;
+}

data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h

@@ -0,0 +1,13 @@
+#include "shared/apir_backend.h"
+
+#define BUFFER_TO_HOST_HANDLE(name) ggml_buffer_to_apir_handle(name)
+
+static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_host_handle_t) buffer;
+}
+
+static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_type_host_handle_t) buft;
+}