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>

README.md

@@ -100,6 +100,7 @@ python scripts/install.py [XMAKE_CONFIG_FLAGS]
 | `--use-mc=[y\|n]`        | 是否沐曦 GPU 接口实现使用maca SDK | n
 | `--moore-gpu=[y\|n]`     | 是否编译摩尔线程 GPU 接口实现     | n
 | `--iluvatar-gpu=[y\|n]`  | 是否编译沐曦 GPU 接口实现         | n
+| `--qy-gpu=[y\|n]`        | 是否编译QY GPU 接口实现           | n
 | `--hygon-dcu=[y\|n]`     | 是否编译海光 DCU 接口实现         | n
 | `--kunlun-xpu=[y\|n]`    | 是否编译昆仑 XPU 接口实现         | n
 | `--ninetoothed=[y\|n]`   | 是否编译九齿实现                 | n
@@ -141,6 +142,9 @@ python scripts/install.py [XMAKE_CONFIG_FLAGS]
      # linux系统：--cuda=$CUDA_HOME
      xmake f --nv-gpu=true --cuda=$CUDA_HOME -cv
 
+     # QY
+     xmake f --qy-gpu=true --cuda=$CUDA_HOME -cv
+
      # 寒武纪
      xmake f --cambricon-mlu=true -cv
 

include/infinicore.h

@@ -46,6 +46,7 @@ typedef enum {
     INFINI_DEVICE_ILUVATAR = 6,
     INFINI_DEVICE_KUNLUN = 7,
     INFINI_DEVICE_HYGON = 8,
+    INFINI_DEVICE_QY = 9,
     INFINI_DEVICE_TYPE_COUNT
 } infiniDevice_t;
 

include/infinicore/device.hpp

@@ -21,6 +21,7 @@ public:
         ILUVATAR = INFINI_DEVICE_ILUVATAR,
         KUNLUN = INFINI_DEVICE_KUNLUN,
         HYGON = INFINI_DEVICE_HYGON,
+        QY = INFINI_DEVICE_QY,
         COUNT = INFINI_DEVICE_TYPE_COUNT,
     };
 

include/infiniop.h

@@ -6,25 +6,30 @@
 #include "infiniop/ops/attention.h"
 #include "infiniop/ops/causal_softmax.h"
 #include "infiniop/ops/clip.h"
-#include "infiniop/ops/logsoftmax.h"
 #include "infiniop/ops/conv.h"
 #include "infiniop/ops/dequantize_awq.h"
+#include "infiniop/ops/gelu.h"
 #include "infiniop/ops/gemm.h"
+#include "infiniop/ops/layer_norm.h"
+#include "infiniop/ops/logsoftmax.h"
+#include "infiniop/ops/lp_norm.h"
 #include "infiniop/ops/mul.h"
+#include "infiniop/ops/ones.h"
 #include "infiniop/ops/random_sample.h"
 #include "infiniop/ops/rearrange.h"
 #include "infiniop/ops/relu.h"
 #include "infiniop/ops/rms_norm.h"
 #include "infiniop/ops/rope.h"
+#include "infiniop/ops/sigmoid.h"
 #include "infiniop/ops/silu.h"
+#include "infiniop/ops/softmax.h"
 #include "infiniop/ops/softplus.h"
 #include "infiniop/ops/sub.h"
 #include "infiniop/ops/swiglu.h"
+#include "infiniop/ops/tanh.h"
 #include "infiniop/ops/topkrouter.h"
-#include "infiniop/ops/zeros.h"
-#include "infiniop/ops/ones.h"
 #include "infiniop/ops/topksoftmax.h"
-#include "infiniop/ops/sigmoid.h"
+#include "infiniop/ops/zeros.h"
 #include "infiniop/tensor_descriptor.h"
 
 #endif // __INFINIOP_API_H__

include/infiniop/ops/gelu.h

+#ifndef __INFINIOP_GELU_API_H__
+#define __INFINIOP_GELU_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopGeluDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateGeluDescriptor(infiniopHandle_t handle,
+                                                         infiniopGeluDescriptor_t *desc_ptr,
+                                                         infiniopTensorDescriptor_t output,
+                                                         infiniopTensorDescriptor_t intput);
+
+__C __export infiniStatus_t infiniopGetGeluWorkspaceSize(infiniopGeluDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopGelu(infiniopGeluDescriptor_t desc,
+                                         void *workspace,
+                                         size_t workspace_size,
+                                         void *output,
+                                         const void *intput,
+                                         void *stream);
+
+__C __export infiniStatus_t infiniopDestroyGeluDescriptor(infiniopGeluDescriptor_t desc);
+
+#endif

include/infiniop/ops/layer_norm.h

+#ifndef __INFINIOP_LAYER_NORM_API_H__
+#define __INFINIOP_LAYER_NORM_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopLayerNormDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateLayerNormDescriptor(
+    infiniopHandle_t handle,
+    infiniopLayerNormDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t output_desc,
+    infiniopTensorDescriptor_t input_standardization_desc,
+    infiniopTensorDescriptor_t input_std_deviation_desc,
+    infiniopTensorDescriptor_t input_desc,
+    infiniopTensorDescriptor_t weight_desc,
+    infiniopTensorDescriptor_t bias_desc,
+    float eps);
+
+__C __export infiniStatus_t infiniopGetLayerNormWorkspaceSize(infiniopLayerNormDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopLayerNorm(infiniopLayerNormDescriptor_t desc,
+                                              void *workspace,
+                                              size_t workspace_size,
+                                              void *output,
+                                              void *input_standardization,
+                                              void *input_std_deviation,
+                                              const void *input,
+                                              const void *weight,
+                                              const void *bias,
+                                              void *stream);
+
+__C __export infiniStatus_t infiniopDestroyLayerNormDescriptor(infiniopLayerNormDescriptor_t desc);
+
+#endif

include/infiniop/ops/lp_norm.h

+#ifndef __INFINIOP_LP_NORM_API_H__
+#define __INFINIOP_LP_NORM_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopLPNormDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateLPNormDescriptor(
+    infiniopHandle_t handle,
+    infiniopLPNormDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t output_desc,
+    infiniopTensorDescriptor_t input_desc,
+    int axis,
+    int p,
+    float eps);
+
+__C __export infiniStatus_t infiniopGetLPNormWorkspaceSize(infiniopLPNormDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopLPNorm(infiniopLPNormDescriptor_t desc,
+                                           void *workspace,
+                                           size_t workspace_size,
+                                           void *output,
+                                           const void *input,
+                                           void *stream);
+
+__C __export infiniStatus_t infiniopDestroyLPNormDescriptor(infiniopLPNormDescriptor_t desc);
+
+#endif

include/infiniop/ops/relu.h

@@ -10,6 +10,8 @@ __C __export infiniStatus_t infiniopCreateReluDescriptor(infiniopHandle_t handle
                                                          infiniopTensorDescriptor_t y,
                                                          infiniopTensorDescriptor_t x);
 
+__C __export infiniStatus_t infiniopGetReluWorkspaceSize(infiniopReluDescriptor_t desc, size_t *size);
+
 __C __export infiniStatus_t infiniopRelu(infiniopReluDescriptor_t desc,
                                          void *workspace,
                                          size_t workspace_size,

include/infiniop/ops/softmax.h

+#ifndef __INFINIOP_SOFTMAX_API_H__
+#define __INFINIOP_SOFTMAX_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopSoftmaxDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateSoftmaxDescriptor(
+    infiniopHandle_t handle,
+    infiniopSoftmaxDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t y_desc,
+    infiniopTensorDescriptor_t x_desc,
+    int axis);
+
+__C __export infiniStatus_t infiniopGetSoftmaxWorkspaceSize(infiniopSoftmaxDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopSoftmax(
+    infiniopSoftmaxDescriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *y,
+    const void *x,
+    void *stream);
+
+__C __export infiniStatus_t infiniopDestroySoftmaxDescriptor(infiniopSoftmaxDescriptor_t desc);
+
+#endif

include/infiniop/ops/tanh.h

+#ifndef __INFINIOP_TANH_API_H__
+#define __INFINIOP_TANH_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopTanhDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateTanhDescriptor(infiniopHandle_t handle,
+                                                         infiniopTanhDescriptor_t *desc_ptr,
+                                                         infiniopTensorDescriptor_t output,
+                                                         infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetTanhWorkspaceSize(infiniopTanhDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopTanh(infiniopTanhDescriptor_t desc,
+                                         void *workspace,
+                                         size_t workspace_size,
+                                         void *output,
+                                         const void *input,
+                                         void *stream);
+
+__C __export infiniStatus_t infiniopDestroyTanhDescriptor(infiniopTanhDescriptor_t desc);
+
+#endif

python/infinicore/device.py

@@ -108,4 +108,5 @@ _TORCH_DEVICE_MAP = {
     _infinicore.Device.Type.ILUVATAR: "cuda",
     _infinicore.Device.Type.KUNLUN: "cuda",
     _infinicore.Device.Type.HYGON: "cuda",
+    _infinicore.Device.Type.QY: "cuda",
 }

scripts/python_test.py

@@ -16,21 +16,29 @@ def run_tests(args):
         "attention.py",
         "causal_softmax.py",
         "clip.py",
+        "conv.py",
+        #"dequantize_awq.py",
+        "gelu.py",
         "gemm.py",
+        #"layer_norm.py",
         "logsoftmax.py",
+        #"lp_norm.py",
         "mul.py",
+        "ones.py",
         "random_sample.py",
         "rearrange.py",
+        "relu.py",
         "rms_norm.py",
         "rope.py",
+        "sigmoid.py",
+        #"softmax.py",
+        "softplus.py",
         "sub.py",
         "swiglu.py",
-        "softplus.py",
-        "ones.py",
-        "zeros.py",
-        "sigmoid.py",
+        "tanh.py",
         "topkrouter.py",
         "topksoftmax.py",
+        "zeros.py",
     ]:
         result = subprocess.run(
             f"python {test} {args} --debug", text=True, encoding="utf-8", shell=True

src/infiniccl-test/main.cpp

@@ -12,7 +12,7 @@ void printUsage() {
     std::cout << "infiniccl-test --<device>" << std::endl
               << std::endl;
     std::cout << "  --<device>" << std::endl;
-    std::cout << "    Specify the device type --(nvidia|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon)." << std::endl
+    std::cout << "    Specify the device type --(nvidia|cambricon|ascend|metax|moore|iluvatar|qy|kunlun|hygon)." << std::endl
               << std::endl;
     std::cout << "The program will run tests on all visible devices of the specified device type."
               << " Use Environmental Variables such as CUDA_VSIBLE_DEVICES to limit visible device IDs.";
@@ -43,6 +43,7 @@ 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/infiniccl/cuda/infiniccl_cuda.h

@@ -4,7 +4,7 @@
 #include "../infiniccl_impl.h"
 
 // Windows does not support CUDA
-#if (defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API)) && defined(ENABLE_CCL) && !defined(_WIN32)
+#if (defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API) || defined(ENABLE_HYGON_API)) && defined(ENABLE_CCL) && !defined(_WIN32)
 INFINICCL_DEVICE_API_IMPL(cuda)
 #else
 INFINICCL_DEVICE_API_NOOP(cuda)

src/infiniccl/infiniccl.cc

@@ -20,6 +20,7 @@ __C infiniStatus_t infinicclCommInitAll(
     switch (device_type) {
         COMM_INIT_ALL(INFINI_DEVICE_NVIDIA, cuda);
         COMM_INIT_ALL(INFINI_DEVICE_ILUVATAR, cuda);
+        COMM_INIT_ALL(INFINI_DEVICE_QY, cuda);
         COMM_INIT_ALL(INFINI_DEVICE_HYGON, cuda);
         COMM_INIT_ALL(INFINI_DEVICE_ASCEND, ascend);
         COMM_INIT_ALL(INFINI_DEVICE_CAMBRICON, cambricon);
@@ -45,6 +46,7 @@ __C infiniStatus_t infinicclCommDestroy(infinicclComm_t comm) {
     switch (comm->device_type) {
         COMM_DESTROY(INFINI_DEVICE_NVIDIA, cuda);
         COMM_DESTROY(INFINI_DEVICE_ILUVATAR, cuda);
+        COMM_DESTROY(INFINI_DEVICE_QY, cuda);
         COMM_DESTROY(INFINI_DEVICE_HYGON, cuda);
         COMM_DESTROY(INFINI_DEVICE_ASCEND, ascend);
         COMM_DESTROY(INFINI_DEVICE_CAMBRICON, cambricon);
@@ -77,6 +79,7 @@ __C infiniStatus_t infinicclAllReduce(
     switch (comm->device_type) {
         ALL_REDUCE(INFINI_DEVICE_NVIDIA, cuda);
         ALL_REDUCE(INFINI_DEVICE_ILUVATAR, cuda);
+        ALL_REDUCE(INFINI_DEVICE_QY, cuda);
         ALL_REDUCE(INFINI_DEVICE_HYGON, cuda);
         ALL_REDUCE(INFINI_DEVICE_ASCEND, ascend);
         ALL_REDUCE(INFINI_DEVICE_CAMBRICON, cambricon);

src/infinicore-test/README.md

@@ -63,6 +63,7 @@ xmake build infinicore-test
 ./infinicore-test --metax
 ./infinicore-test --moore
 ./infinicore-test --iluvatar
+./infinicore-test --qy
 ./infinicore-test --kunlun
 ./infinicore-test --hygon
 ```

src/infinicore-test/main.cc

@@ -39,6 +39,7 @@ void printUsage() {
               << "  metax" << std::endl
               << "  moore" << std::endl
               << "  iluvatar" << std::endl
+              << "  qy" << std::endl
               << "  kunlun" << std::endl
               << "  hygon" << std::endl
               << std::endl
@@ -77,6 +78,8 @@ ParsedArgs parseArgs(int argc, char *argv[]) {
             args.device_type = INFINI_DEVICE_MOORE;
         } else if (arg == "--iluvatar") {
             args.device_type = INFINI_DEVICE_ILUVATAR;
+        } else if (arg == "--qy") {
+            args.device_type = INFINI_DEVICE_QY;
         } else if (arg == "--kunlun") {
             args.device_type = INFINI_DEVICE_KUNLUN;
         } else if (arg == "--hygon") {

src/infinicore/device.cc

@@ -35,6 +35,8 @@ std::string Device::toString(const Type &type) {
         return "MOORE";
     case Type::ILUVATAR:
         return "ILUVATAR";
+    case Type::QY:
+        return "QY";
     case Type::KUNLUN:
         return "KUNLUN";
     case Type::HYGON:

src/infinicore/pybind11/device.hpp

@@ -19,6 +19,7 @@ inline void bind(py::module &m) {
         .value("METAX", Device::Type::METAX)
         .value("MOORE", Device::Type::MOORE)
         .value("ILUVATAR", Device::Type::ILUVATAR)
+        .value("QY", Device::Type::QY)
         .value("KUNLUN", Device::Type::KUNLUN)
         .value("HYGON", Device::Type::HYGON)
         .value("COUNT", Device::Type::COUNT);

src/infiniop-test/src/main.cpp

@@ -22,7 +22,7 @@ void printUsage() {
     std::cout << "    Path to the test gguf file" << std::endl
               << std::endl;
     std::cout << "  --<device>[:id]" << std::endl;
-    std::cout << "    (Optional) Specify the device type --(cpu|nvidia|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon) and device ID (optional). CPU by default." << std::endl
+    std::cout << "    (Optional) Specify the device type --(cpu|nvidia|cambricon|ascend|metax|moore|iluvatar|qy|kunlun|hygon) and device ID (optional). CPU by default." << std::endl
               << std::endl;
     std::cout << "  --warmup <warmups>" << std::endl;
     std::cout << "    (Optional) Number of warmups to perform before timing. Default to 0." << std::endl
@@ -77,6 +77,7 @@ ParsedArgs parseArgs(int argc, char *argv[]) {
             PARSE_DEVICE("--metax", INFINI_DEVICE_METAX)
             PARSE_DEVICE("--moore", INFINI_DEVICE_MOORE)
             PARSE_DEVICE("--iluvatar", INFINI_DEVICE_ILUVATAR)
+            PARSE_DEVICE("--qy", INFINI_DEVICE_QY)
             PARSE_DEVICE("--kunlun", INFINI_DEVICE_KUNLUN)
             PARSE_DEVICE("--hygon", INFINI_DEVICE_HYGON)
             else if (arg == "--warmup" && i + 1 < argc) {