ISSUE/628 适配QY C610 GPU，增加编译选项，适配已有算子。添加bge类模型所需的算子， (#629)

* ISSUE/628 适配QY C610 GPU，增加编译选项，适配已有算子。添加bge类模型所需的算子，包括gelu,layer_norm，lp_norm(支持l1，l2 norm)，relu，softmax，tanh。

---------
Co-authored-by: xgqdut2016 <kenan_gewei@163.com>
Co-authored-by: xgqdut2016 <140036308+xgqdut2016@users.noreply.github.com>

src/infiniop/ops/tanh/cuda/kernel.cuh

+#ifndef __TANH_CUDA_H__
+#define __TANH_CUDA_H__
+
+#include <cmath>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+namespace op::tanh::cuda {
+typedef struct TanhOp {
+    static constexpr size_t num_inputs = 1;
+
+    __device__ __forceinline__ float tanh_f32_func(float x) const {
+        return tanhf(x);
+    }
+    template <typename T>
+    __device__ __forceinline__ T operator()(const T &input) const {
+        if constexpr (std::is_same_v<T, half2>) {
+            float2 vf = __half22float2(input);
+            float2 vr = make_float2(tanh_f32_func(vf.x), tanh_f32_func(vf.y));
+            return __float22half2_rn(vr);
+        } else if constexpr (std::is_same_v<T, half>) {
+            float xf = __half2float(input);
+            float yf = tanh_f32_func(xf);
+            return __float2half_rn(yf);
+        } else if constexpr (std::is_same_v<T, cuda_bfloat162>) {
+            float f0 = __bfloat162float(__low2bfloat16(input));
+            float f1 = __bfloat162float(__high2bfloat16(input));
+            float r0 = tanh_f32_func(f0);
+            float r1 = tanh_f32_func(f1);
+            return __floats2bfloat162_rn(r0, r1);
+        } else if constexpr (std::is_same_v<T, cuda_bfloat16>) {
+            float xf = __bfloat162float(input);
+            float rf = tanh_f32_func(xf);
+            return __float2bfloat16_rn(rf);
+        } else if constexpr (std::is_same_v<T, float>) {
+            return tanh_f32_func(input);
+        } else if constexpr (std::is_same_v<T, double>) {
+            return std::tanh(input);
+        } else {
+            return std::tanh(input);
+        }
+    }
+} TanhOp;
+} // namespace op::tanh::cuda
+
+#endif // __TANH_CUDA_H__

src/infiniop/ops/tanh/metax/tanh_metax.h

+#ifndef __TANH_METAX_API_H__
+#define __TANH_METAX_API_H__
+
+#include "../../../elementwise/metax/elementwise_metax_api.h"
+
+ELEMENTWISE_DESCRIPTOR(tanh, metax)
+
+#endif // __TANH_METAX_API_H__

src/infiniop/ops/tanh/metax/tanh_metax.maca

+#include "tanh_metax.h"
+
+#include "../../../elementwise/metax/elementwise_metax.h"
+
+#include "../cuda/kernel.cuh"
+
+namespace op::tanh::metax {
+
+Descriptor::~Descriptor() = default;
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle_,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t out_desc,
+    std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
+
+    auto handle = reinterpret_cast<device::metax::Handle *>(handle_);
+    auto dtype = out_desc->dtype();
+
+    const auto &input_desc = input_desc_vec.at(0);
+    const auto &output_shape = out_desc->shape();
+    const auto &input_shape = input_desc->shape();
+
+    CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_F32, INFINI_DTYPE_F64, INFINI_DTYPE_BF16);
+
+    CHECK_SAME_SHAPE(output_shape, input_shape);
+
+    // create CUDA elementwise descriptor
+    CREATE_ELEMENTWISE_METAX_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec)
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    std::vector<const void *> inputs,
+    void *stream) const {
+
+    if (workspace_size < _workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
+    }
+
+    switch (_dtype) {
+    case INFINI_DTYPE_F16:
+        return _device_info->calculate<256, cuda::TanhOp, half>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_BF16:
+        return _device_info->calculate<256, cuda::TanhOp, cuda_bfloat16>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F32:
+        return _device_info->calculate<256, cuda::TanhOp, float>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F64:
+        return _device_info->calculate<256, cuda::TanhOp, double>(_info, workspace, output, inputs, stream);
+    default:
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::tanh::metax

src/infiniop/ops/tanh/nvidia/tanh_nvidia.cu

+#include "../../../elementwise/nvidia/elementwise_nvidia.cuh"
+
+#include "../cuda/kernel.cuh"
+#include "tanh_nvidia.cuh"
+
+namespace op::tanh::nvidia {
+
+Descriptor::~Descriptor() = default;
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle_,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t out_desc,
+    std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
+
+    auto handle = reinterpret_cast<device::nvidia::Handle *>(handle_);
+    auto dtype = out_desc->dtype();
+
+    const auto &input_desc = input_desc_vec.at(0);
+    const auto &output_shape = out_desc->shape();
+    const auto &input_shape = input_desc->shape();
+
+    CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_F32, INFINI_DTYPE_F64, INFINI_DTYPE_BF16);
+
+    CHECK_SAME_SHAPE(output_shape, input_shape);
+
+    // create CUDA elementwise descriptor
+    CREATE_ELEMENTWISE_CUDA_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec)
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    std::vector<const void *> inputs,
+    void *stream) const {
+
+    if (workspace_size < _workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
+    }
+
+    switch (_dtype) {
+    case INFINI_DTYPE_F16:
+        return _device_info->calculate<256, cuda::TanhOp, half>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_BF16:
+        return _device_info->calculate<256, cuda::TanhOp, cuda_bfloat16>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F32:
+        return _device_info->calculate<256, cuda::TanhOp, float>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F64:
+        return _device_info->calculate<256, cuda::TanhOp, double>(_info, workspace, output, inputs, stream);
+    default:
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::tanh::nvidia

src/infiniop/ops/tanh/nvidia/tanh_nvidia.cuh

+#ifndef __TANH_CUDA_API_H__
+#define __TANH_CUDA_API_H__
+
+#include "../../../elementwise/nvidia/elementwise_nvidia_api.cuh"
+
+ELEMENTWISE_DESCRIPTOR(tanh, nvidia)
+
+#endif // __TANH_CUDA_API_H__

src/infiniop/ops/tanh/operator.cc

+#include "../../operator.h"
+#include "../../handle.h"
+#include "infiniop/ops/tanh.h"
+
+#ifdef ENABLE_CPU_API
+#include "cpu/tanh_cpu.h"
+#endif
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
+#include "nvidia/tanh_nvidia.cuh"
+#endif
+#ifdef ENABLE_METAX_API
+#include "metax/tanh_metax.h"
+#endif
+
+__C infiniStatus_t infiniopCreateTanhDescriptor(
+    infiniopHandle_t handle,
+    infiniopTanhDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t output_desc,
+    infiniopTensorDescriptor_t input_desc) {
+
+#define CREATE(CASE, NAMESPACE)                                             \
+    case CASE:                                                              \
+        return op::tanh::NAMESPACE::Descriptor::create(                     \
+            handle,                                                         \
+            reinterpret_cast<op::tanh::NAMESPACE::Descriptor **>(desc_ptr), \
+            output_desc,                                                    \
+            {input_desc})
+
+    switch (handle->device) {
+
+#ifdef ENABLE_CPU_API
+        CREATE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        CREATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        CREATE(INFINI_DEVICE_METAX, metax);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef CREATE
+}
+
+__C infiniStatus_t infiniopGetTanhWorkspaceSize(infiniopTanhDescriptor_t desc, size_t *size) {
+
+#define GET(CASE, NAMESPACE)                                                                \
+    case CASE:                                                                              \
+        *size = reinterpret_cast<op::tanh::NAMESPACE::Descriptor *>(desc)->workspaceSize(); \
+        return INFINI_STATUS_SUCCESS;
+
+    switch (desc->device_type) {
+#ifdef ENABLE_CPU_API
+        GET(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        GET(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        GET(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        GET(INFINI_DEVICE_METAX, metax);
+#endif
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+#undef GET
+
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+__C infiniStatus_t infiniopTanh(
+    infiniopTanhDescriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    const void *input,
+    void *stream) {
+
+#define CALCULATE(CASE, NAMESPACE)                                             \
+    case CASE:                                                                 \
+        return reinterpret_cast<const op::tanh::NAMESPACE::Descriptor *>(desc) \
+            ->calculate(workspace, workspace_size, output, {input}, stream)
+
+    switch (desc->device_type) {
+
+#ifdef ENABLE_CPU_API
+        CALCULATE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        CALCULATE(INFINI_DEVICE_METAX, metax);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef CALCULATE
+}
+
+__C infiniStatus_t
+infiniopDestroyTanhDescriptor(infiniopTanhDescriptor_t desc) {
+
+#define DELETE(CASE, NAMESPACE)                                                 \
+    case CASE:                                                                  \
+        delete reinterpret_cast<const op::tanh::NAMESPACE::Descriptor *>(desc); \
+        return INFINI_STATUS_SUCCESS
+
+    switch (desc->device_type) {
+
+#ifdef ENABLE_CPU_API
+        DELETE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        DELETE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        DELETE(INFINI_DEVICE_METAX, metax);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef DELETE
+}

src/infiniop/ops/topkrouter/nvidia/topkrouter_nvidia.cu

-#ifdef ENABLE_NVIDIA_API
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_QY_API)
 
 #include "../../../devices/nvidia/nvidia_common.cuh"
 #include "../../../devices/nvidia/nvidia_kernel_common.cuh"

src/infiniop/ops/topkrouter/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/topkrouter_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_QY_API)
 #include "nvidia/topkrouter_nvidia.cuh"
 #endif
 
@@ -23,6 +23,9 @@ __C infiniStatus_t infiniopCreateTopkrouterDescriptor(infiniopHandle_t handle, i
 #endif
 #ifdef ENABLE_NVIDIA_API
         CREATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
 #endif
     }
 
@@ -43,6 +46,9 @@ __C infiniStatus_t infiniopGetTopkrouterWorkspaceSize(infiniopTopkrouterDescript
 #endif
 #ifdef ENABLE_NVIDIA_API
         GET(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
 #endif
     }
 
@@ -66,6 +72,9 @@ __C infiniStatus_t infiniopTopkrouter(infiniopTopkrouterDescriptor_t desc, void
 #endif
 #ifdef ENABLE_NVIDIA_API
         CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
 #endif
     }
 
@@ -86,6 +95,9 @@ __C infiniStatus_t infiniopDestroyTopkrouterDescriptor(infiniopTopkrouterDescrip
 #endif
 #ifdef ENABLE_NVIDIA_API
         DESTROY(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        DESTROY(INFINI_DEVICE_QY, nvidia);
 #endif
     }
 

src/infiniop/ops/topksoftmax/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/topksoftmax_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_QY_API)
 #include "nvidia/topksoftmax_nvidia.cuh"
 #endif
 #ifdef ENABLE_METAX_API
@@ -28,6 +28,9 @@ __C infiniStatus_t infiniopCreateTopksoftmaxDescriptor(infiniopHandle_t handle,
 #ifdef ENABLE_NVIDIA_API
         CREATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -52,6 +55,9 @@ __C infiniStatus_t infiniopGetTopksoftmaxWorkspaceSize(infiniopTopksoftmaxDescri
 #ifdef ENABLE_NVIDIA_API
         GET(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, metax);
 #endif
@@ -81,6 +87,9 @@ __C infiniStatus_t infiniopTopksoftmax(infiniopTopksoftmaxDescriptor_t desc, voi
 #ifdef ENABLE_NVIDIA_API
         CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -105,6 +114,9 @@ __C infiniStatus_t infiniopDestroyTopksoftmaxDescriptor(infiniopTopksoftmaxDescr
 #ifdef ENABLE_NVIDIA_API
         DESTROY(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DESTROY(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         DESTROY(INFINI_DEVICE_METAX, metax);
 #endif

src/infiniop/ops/zeros/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/zeros_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
 #include "nvidia/zeros_nvidia.cuh"
 #endif
 #ifdef ENABLE_METAX_API
@@ -40,6 +40,9 @@ __C infiniStatus_t infiniopCreateZerosDescriptor(
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -70,6 +73,9 @@ __C infiniStatus_t infiniopGetZerosWorkspaceSize(infiniopZerosDescriptor_t desc,
 #ifdef ENABLE_ILUVATAR_API
         GET(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, metax);
 #endif
@@ -108,6 +114,9 @@ __C infiniStatus_t infiniopZeros(
 #ifdef ENABLE_ILUVATAR_API
         CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -140,6 +149,9 @@ infiniopDestroyZerosDescriptor(infiniopZerosDescriptor_t desc) {
 #ifdef ENABLE_ILUVATAR_API
         DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         DELETE(INFINI_DEVICE_METAX, metax);
 #endif

src/infiniop/reduce/cuda/reduce.cuh

 #ifndef __INFINIOP_REDUCE_CUDA_H__
 #define __INFINIOP_REDUCE_CUDA_H__
-
+#include <cub/block/block_reduce.cuh>
 /*
  * Device functions for reduction operations on CUDA.
  *

src/infinirt-test/main.cc

@@ -20,6 +20,7 @@ void printUsage() {
               << "  metax" << std::endl
               << "  moore" << std::endl
               << "  iluvatar" << std::endl
+              << "  qy" << std::endl
               << "  kunlun" << std::endl
               << "  hygon" << std::endl
               << std::endl;
@@ -51,6 +52,7 @@ ParsedArgs parseArgs(int argc, char *argv[]) {
         else PARSE_DEVICE("--metax", INFINI_DEVICE_METAX)
         else PARSE_DEVICE("--moore", INFINI_DEVICE_MOORE)
         else PARSE_DEVICE("--iluvatar", INFINI_DEVICE_ILUVATAR)
+        else PARSE_DEVICE("--qy", INFINI_DEVICE_QY)
         else PARSE_DEVICE("--kunlun", INFINI_DEVICE_KUNLUN)
         else PARSE_DEVICE("--hygon", INFINI_DEVICE_HYGON)
         else {

src/infinirt/cuda/infinirt_cuda.cuh

@@ -3,7 +3,7 @@
 #include "../infinirt_impl.h"
 
 namespace infinirt::cuda {
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API) || defined(ENABLE_HYGON_API)
 INFINIRT_DEVICE_API_IMPL
 #else
 INFINIRT_DEVICE_API_NOOP

src/infinirt/infinirt.cc

@@ -23,7 +23,7 @@ __C infiniStatus_t infinirtGetAllDeviceCount(int *count_array) {
         return INFINI_STATUS_NULL_POINTER;
     }
     for (size_t i = 0; i < INFINI_DEVICE_TYPE_COUNT; i++) {
-        if (i == INFINI_DEVICE_ILUVATAR || i == INFINI_DEVICE_HYGON) {
+        if (i == INFINI_DEVICE_ILUVATAR || i == INFINI_DEVICE_QY || i == INFINI_DEVICE_KUNLUN || i == INFINI_DEVICE_HYGON) {
             count_array[i] = 0;
             continue;
         }
@@ -77,6 +77,9 @@ __C infiniStatus_t infinirtGetDevice(infiniDevice_t *device_ptr, int *device_id_
         case INFINI_DEVICE_ILUVATAR:                                   \
             _status = infinirt::cuda::API PARAMS;                      \
             break;                                                     \
+        case INFINI_DEVICE_QY:                                         \
+            _status = infinirt::cuda::API PARAMS;                      \
+            break;                                                     \
         case INFINI_DEVICE_HYGON:                                      \
             _status = infinirt::cuda::API PARAMS;                      \
             break;                                                     \

test/infinicore/framework/config.py

@@ -23,6 +23,7 @@ def get_supported_hardware_platforms():
         ("--moore", "Moore Threads GPUs (requires torch_musa)"),
         ("--kunlun", "Kunlun XPUs (requires torch_xmlir)"),
         ("--hygon", "Hygon DCUs"),
+        ("--qy", "QY GPUs"),
     ]
 
 
@@ -194,6 +195,15 @@ def get_test_devices(args):
             devices_to_test.append(InfiniDeviceEnum.HYGON)
         except ImportError:
             print("Warning: Hygon DCU support not available")
+            
+    if args.qy:
+        try:
+            # Iluvatar GPU detection
+            import torch
+
+            devices_to_test.append(InfiniDeviceEnum.QY)
+        except ImportError:
+            print("Warning: QY GPU support not available")
 
     # Default to CPU if no devices specified
     if not devices_to_test:

test/infinicore/framework/devices.py

@@ -8,6 +8,7 @@ class InfiniDeviceEnum:
     ILUVATAR = 6
     KUNLUN = 7
     HYGON = 8
+    QY = 9
 
 
 InfiniDeviceNames = {
@@ -18,6 +19,7 @@ InfiniDeviceNames = {
     InfiniDeviceEnum.METAX: "Metax",
     InfiniDeviceEnum.MOORE: "Moore",
     InfiniDeviceEnum.ILUVATAR: "Iluvatar",
+    InfiniDeviceEnum.QY: "Qy",
     InfiniDeviceEnum.KUNLUN: "Kunlun",
     InfiniDeviceEnum.HYGON: "Hygon",
 }
@@ -32,4 +34,5 @@ torch_device_map = {
     InfiniDeviceEnum.ILUVATAR: "cuda",
     InfiniDeviceEnum.KUNLUN: "cuda",
     InfiniDeviceEnum.HYGON: "cuda",
+    InfiniDeviceEnum.QY: "cuda",
 }

test/infiniop/attention.py

@@ -23,7 +23,6 @@ from libinfiniop import (
 )
 
 
-
 def causal_softmax(x):
     type = x.dtype
     mask = torch.tril(torch.ones_like(x), diagonal=-1).flip(dims=[-2, -1])

test/infiniop/conv.py

@@ -96,27 +96,27 @@ NUM_ITERATIONS = 1000
 
 
 def conv(x, w, stride, padding, dilation, y_tensor, bias=None):
-    match len(x.shape) - 2:
-        case 1:
-            y_tensor.copy_(
-                F.conv1d(
-                    x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
-                )
+    ndim = len(x.shape) - 2#不要使用match，会导致CI无法通过
+    if ndim == 1:
+        y_tensor.copy_(
+            F.conv1d(
+                x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
             )
-        case 2:
-            y_tensor.copy_(
-                F.conv2d(
-                    x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
-                )
+        )
+    elif ndim == 2:
+        y_tensor.copy_(
+            F.conv2d(
+                x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
             )
-        case 3:
-            y_tensor.copy_(
-                F.conv3d(
-                    x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
-                )
+        )
+    elif ndim == 3:
+        y_tensor.copy_(
+            F.conv3d(
+                x, w, bias=bias, stride=stride, padding=padding, dilation=dilation
             )
-        case _:
-            print("Error: Pytorch -> Unsupported tensor dimension")
+        )
+    else:
+        print("Error: Pytorch -> Unsupported tensor dimension")
 
 
 # infer the shape of the output given the inputs for a N-ary convolution

test/infiniop/gelu.py

+import torch
+import ctypes
+from ctypes import c_uint64
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    debug,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+from enum import Enum, auto
+
+# ==============================================================================
+#  Configuration (Internal Use Only)
+# ==============================================================================
+# These are not meant to be imported from other modules
+_TEST_CASES_ = [
+    # shape, input_stride, output_stride
+    ((13, 4), None, None),
+    ((13, 4), (10, 1), (10, 1)),
+    #((13, 4), (0, 1), None),
+    ((13, 4, 4), None, None),
+    ((13, 4, 4), (20, 4, 1), (20, 4, 1)),
+    #((13, 4, 4), (4, 0, 1), None),
+    ((16, 5632), None, None),
+    ((16, 5632), (13312, 1), (13312, 1)),
+    ((4, 4, 5632), None, None),
+    ((4, 4, 5632), (45056, 5632, 1), (45056, 5632, 1)),
+]
+
+
+class Inplace(Enum):
+    OUT_OF_PLACE = auto()
+    INPLACE = auto()
+
+
+# Inplace options applied for each test case in _TEST_CASES_
+_INPLACE = [
+    Inplace.OUT_OF_PLACE,
+    Inplace.INPLACE,
+]
+
+# Form the test cases by appending each element of _INPLACE to each tuple in _TEST_CASES_
+_TEST_CASES = [
+    test_case + (inplace_item,)
+    for test_case in _TEST_CASES_
+    for inplace_item in _INPLACE
+]
+
+# Data types used for testing
+_TENSOR_DTYPES = [InfiniDtype.BF16, InfiniDtype.F16, InfiniDtype.F32, InfiniDtype.F64]
+
+# Tolerance map for different data types
+_TOLERANCE_MAP = {
+    InfiniDtype.BF16: {"atol": 1e-2, "rtol": 1e-2},
+    InfiniDtype.F16: {"atol": 1e-3, "rtol": 1e-3},
+    InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-5},
+    InfiniDtype.F64: {"atol": 1e-6, "rtol": 1e-6},
+}
+
+DEBUG = False
+PROFILE = False
+NUM_PRERUN = 10
+NUM_ITERATIONS = 1000
+
+
+def test(
+    handle,
+    device,
+    shape,
+    input_stride=None,
+    output_stride=None,
+    inplace=Inplace.OUT_OF_PLACE,
+    dtype=torch.float16,
+    sync=None,
+):
+    input = TestTensor(shape, input_stride, dtype, device)
+    if inplace == Inplace.INPLACE:
+        if input_stride != output_stride:
+            return
+        output = input
+    else:
+        output = TestTensor(shape, output_stride, dtype, device, mode="ones")
+
+    if output.is_broadcast():
+        return
+
+    print(
+        f"Testing Gelu on {InfiniDeviceNames[device]} with shape:{shape} input_stride:{input_stride} output_stride:{output_stride}"
+        f"dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
+    )
+
+    new_tensor = torch.nn.functional.gelu(input.torch_tensor())
+    output.update_torch_tensor(new_tensor)
+
+    if sync is not None:
+        sync()
+
+    descriptor = infiniopOperatorDescriptor_t()
+    check_error(
+        LIBINFINIOP.infiniopCreateGeluDescriptor(
+            handle,
+            ctypes.byref(descriptor),
+            output.descriptor,
+            input.descriptor,
+        )
+    )
+
+    # Invalidate the shape and strides in the descriptor to prevent them from being directly used by the kernel
+    for tensor in [input, output]:
+        tensor.destroy_desc()
+
+    workspace_size = c_uint64(0)
+    check_error(
+        LIBINFINIOP.infiniopGetGeluWorkspaceSize(
+            descriptor, ctypes.byref(workspace_size)
+        )
+    )
+    workspace = TestWorkspace(workspace_size.value, output.device)
+
+    def lib_gelu():
+        check_error(
+            LIBINFINIOP.infiniopGelu(
+                descriptor,
+                workspace.data(),
+                workspace.size(),
+                output.data(),
+                input.data(),
+                None,
+            )
+        )
+
+    lib_gelu()
+
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    if DEBUG:
+        debug(output.actual_tensor(), output.torch_tensor(), atol=atol, rtol=rtol)
+    assert torch.allclose(
+        output.actual_tensor(), output.torch_tensor(), atol=atol, rtol=rtol
+    )
+
+    # Profiling workflow
+    if PROFILE:
+        # fmt: off
+        profile_operation("PyTorch", lambda: torch.nn.functional.gelu(input.torch_tensor()), device, NUM_PRERUN, NUM_ITERATIONS)
+        profile_operation("    lib", lambda: lib_gelu(), device, NUM_PRERUN, NUM_ITERATIONS)
+        # fmt: on
+    check_error(LIBINFINIOP.infiniopDestroyGeluDescriptor(descriptor))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    # Configure testing options
+    DEBUG = args.debug
+    PROFILE = args.profile
+    NUM_PRERUN = args.num_prerun
+    NUM_ITERATIONS = args.num_iterations
+
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+
+    print("\033[92mTest passed!\033[0m")