build turbomind (#35)

* build turbomind

* change namespace fastertransformer to turbomind

* change logger name

src/turbomind/utils/cuda_utils.cc

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/cuda_fp8_utils.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/cuda_fp8_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 /* **************************** debug tools ********************************* */
 
@@ -57,7 +57,7 @@ template<typename T>
 void print_abs_mean(const T* buf, uint size, cudaStream_t stream, std::string name)
 {
     if (buf == nullptr) {
-        FT_LOG_WARNING("It is an nullptr, skip!");
+        TM_LOG_WARNING("It is an nullptr, skip!");
         return;
     }
     cudaDeviceSynchronize();
@@ -110,7 +110,7 @@ template<typename T>
 void print_to_screen(const T* result, const int size)
 {
     if (result == nullptr) {
-        FT_LOG_WARNING("It is an nullptr, skip! \n");
+        TM_LOG_WARNING("It is an nullptr, skip! \n");
         return;
     }
     T* tmp = reinterpret_cast<T*>(malloc(sizeof(T) * size));
@@ -366,7 +366,7 @@ FtCudaDataType getModelFileType(std::string ini_file, std::string section_name)
     FtCudaDataType model_file_type;
     INIReader      reader = INIReader(ini_file);
     if (reader.ParseError() < 0) {
-        FT_LOG_WARNING("Can't load %s. Use FP32 as default", ini_file.c_str());
+        TM_LOG_WARNING("Can't load %s. Use FP32 as default", ini_file.c_str());
         model_file_type = FtCudaDataType::FP32;
     }
     else {
@@ -381,7 +381,7 @@ FtCudaDataType getModelFileType(std::string ini_file, std::string section_name)
             model_file_type = FtCudaDataType::BF16;
         }
         else {
-            FT_LOG_WARNING("Invalid type %s. Use FP32 as default", weight_data_type_str.c_str());
+            TM_LOG_WARNING("Invalid type %s. Use FP32 as default", weight_data_type_str.c_str());
             model_file_type = FtCudaDataType::FP32;
         }
     }
@@ -389,4 +389,4 @@ FtCudaDataType getModelFileType(std::string ini_file, std::string section_name)
 }
 
 /* ************************** end of common utils ************************** */
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/cuda_utils.h

@@ -17,8 +17,8 @@
 #pragma once
 
 #include "3rdparty/INIReader.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/logger.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/logger.h"
 
 #include <cublasLt.h>
 #include <cublas_v2.h>
@@ -31,7 +31,7 @@
 #include <cusparseLt.h>
 #endif
 
-namespace fastertransformer {
+namespace turbomind {
 
 #define MAX_CONFIG_NUM 20
 #define COL32_ 32
@@ -137,7 +137,7 @@ inline void syncAndCheck(const char* const file, int const line)
                 throw std::runtime_error(std::string("[FT][ERROR] CUDA runtime error: ") + (_cudaGetErrorEnum(result))
                                          + " " + file + ":" + std::to_string(line) + " \n");
             }
-            FT_LOG_DEBUG(fmtstr("run syncAndCheck at %s:%d", file, line));
+            TM_LOG_DEBUG(fmtstr("run syncAndCheck at %s:%d", file, line));
         }
     }
 
@@ -212,7 +212,7 @@ inline void myAssert(bool result, const char* const file, int const line, std::s
     do {                                                                                                               \
         bool is_valid_val = (val);                                                                                     \
         if (!is_valid_val) {                                                                                           \
-            fastertransformer::myAssert(is_valid_val, __FILE__, __LINE__, (info));                                     \
+            turbomind::myAssert(is_valid_val, __FILE__, __LINE__, (info));                                     \
         }                                                                                                              \
     } while (0)
 
@@ -442,14 +442,14 @@ void compareTwoTensor(
     }
 
     if (print_size > 0) {
-        FT_LOG_INFO("  id |   pred  |   ref   |abs diff | rel diff (%) |");
+        TM_LOG_INFO("  id |   pred  |   ref   |abs diff | rel diff (%) |");
     }
     float mean_abs_diff = 0.0f;
     float mean_rel_diff = 0.0f;
     int   count         = 0;
     for (int i = 0; i < size; i++) {
         if (i < print_size) {
-            FT_LOG_INFO("%4d | % 6.4f | % 6.4f | % 6.4f | % 7.4f |",
+            TM_LOG_INFO("%4d | % 6.4f | % 6.4f | % 6.4f | % 7.4f |",
                         i,
                         (float)h_pred[i],
                         (float)h_ref[i],
@@ -474,7 +474,7 @@ void compareTwoTensor(
     }
     mean_abs_diff = mean_abs_diff / (float)count;
     mean_rel_diff = mean_rel_diff / (float)count;
-    FT_LOG_INFO("mean_abs_diff: % 6.4f, mean_rel_diff: % 6.4f (%%)", mean_abs_diff, mean_rel_diff);
+    TM_LOG_INFO("mean_abs_diff: % 6.4f, mean_rel_diff: % 6.4f (%%)", mean_abs_diff, mean_rel_diff);
 
     if (fd != nullptr) {
         fprintf(fd, "mean_abs_diff: % 6.4f, mean_rel_diff: % 6.4f (%%)", mean_abs_diff, mean_rel_diff);
@@ -485,4 +485,4 @@ void compareTwoTensor(
 }
 
 /* ************************** end of common utils ************************** */
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/custom_ar_comm.cc

@@ -16,7 +16,7 @@
 
 #include "custom_ar_comm.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 CustomAllReduceComm<T>::CustomAllReduceComm(size_t rank_size, size_t rank): rank_size_(rank_size), rank_(rank)
@@ -137,7 +137,7 @@ void initCustomAllReduceComm(std::vector<std::shared_ptr<AbstractCustomComm>>* c
     if (rank_size != RANKS_PER_NODE) {
 #ifdef BUILD_MULTI_GPU
         if (rank_size > 1) {
-            FT_LOG_WARNING("Custom All Reduce only supports 8 Ranks currently. Using NCCL as Comm.");
+            TM_LOG_WARNING("Custom All Reduce only supports 8 Ranks currently. Using NCCL as Comm.");
         }
 #else
         FT_CHECK_WITH_INFO(rank_size == 1,
@@ -158,7 +158,7 @@ void initCustomAllReduceComm(std::vector<std::shared_ptr<AbstractCustomComm>>* c
     }
     custom_all_reduce_comms->at(0)->allocateAndExchangePeerAccessPointer(custom_all_reduce_comms);
 #else
-    FT_LOG_WARNING("Custom All Reduce is not supported before CUDA 11.2. Using NCCL as Comm.");
+    TM_LOG_WARNING("Custom All Reduce is not supported before CUDA 11.2. Using NCCL as Comm.");
     for (size_t i = 0; i < rank_size; i++) {
         custom_all_reduce_comms->push_back(nullptr);
     }
@@ -186,4 +186,4 @@ initCustomAllReduceComm<uint32_t>(std::vector<std::shared_ptr<AbstractCustomComm
                                   int                                               enable_custom_all_reduce,
                                   size_t                                            rank_size);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/custom_ar_comm.h

@@ -21,12 +21,12 @@
 #include <stdio.h>
 #include <stdlib.h>
 
-#include "src/fastertransformer/kernels/custom_ar_kernels.h"
-#include "src/fastertransformer/utils/Tensor.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/logger.h"
+#include "src/turbomind/kernels/custom_ar_kernels.h"
+#include "src/turbomind/utils/Tensor.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/logger.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 class AbstractCustomComm {
 public:
@@ -84,4 +84,4 @@ struct CustomARCommTypeConverter<__nv_bfloat16> {
 };
 #endif
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/gemm.h"
+#include "src/turbomind/utils/gemm.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 /* ***************************** GEMM Impl ******************************** */
 
@@ -222,7 +222,7 @@ void Gemm::gemm(const GemmOp   transa,
                 const float    alpha,
                 const float    beta)
 {
-    FT_LOG_TRACE("Gemm::gemm [m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]", m, n, k, lda, ldb, ldc);
+    TM_LOG_TRACE("Gemm::gemm [m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]", m, n, k, lda, ldb, ldc);
 
     // Implementation copied from cublasMMWrapper::Gemm
     // Switch A and B since both cublas and cublasLt assume a column major layout,
@@ -423,7 +423,7 @@ void Gemm::batchedGemm(const GemmOp       transa,
                        const float        alpha,
                        const float        beta)
 {
-    FT_LOG_TRACE(
+    TM_LOG_TRACE(
         "Gemm::batchedGemm [b=%ld m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]", batch_size, m, n, k, lda, ldb, ldc);
 
     // Switch A and B.
@@ -624,7 +624,7 @@ void Gemm::stridedBatchedGemm(GemmOp        transa,
                               const float   alpha,
                               const float   beta)
 {
-    FT_LOG_TRACE("Gemm::stridedBatchedGemm [b=%ld, m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]",
+    TM_LOG_TRACE("Gemm::stridedBatchedGemm [b=%ld, m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]",
                  batch_size,
                  m,
                  n,
@@ -873,7 +873,7 @@ void SpGemm::gemm(const GemmOp   transa,
                   const float    alpha,
                   const float    beta)
 {
-    FT_LOG_TRACE("SpGemm::gemm [m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]", m, n, k, lda, ldb, ldc);
+    TM_LOG_TRACE("SpGemm::gemm [m=%ld, n=%ld, k=%ld, lda=%ld, ldb=%ld, ldc=%ld]", m, n, k, lda, ldb, ldc);
     checkDataTypeValidity(Atype);
     checkDataTypeValidity(Btype);
     checkDataTypeValidity(Ctype);
@@ -994,7 +994,7 @@ void SpGemm::gemm(const GemmOp   transa,
 
 std::shared_ptr<Gemm> createGemm(IAllocator* allocator, cudaStream_t stream, bool sparse, bool quantized)
 {
-    FT_LOG_TRACE(
+    TM_LOG_TRACE(
         "Create Gemm instance [sparse=%s, quantized=%s]", sparse ? "true" : "false", quantized ? "true" : "false");
     std::shared_ptr<Gemm> gemm;
     if (!sparse) {
@@ -1105,7 +1105,7 @@ cusparseComputeType getCusparseComputeType(DataType ctype)
 
 void pruneMatrixB(void* data, const cudaStream_t& stream, const size_t k, const size_t n, const GemmOp trans)
 {
-    FT_LOG_TRACE("Prune matrix B [k=%ld, n=%ld, op=%s]", k, n, getGemmOpString(trans).c_str());
+    TM_LOG_TRACE("Prune matrix B [k=%ld, n=%ld, op=%s]", k, n, getGemmOpString(trans).c_str());
 
     // Due to A/B switching, the matrix B will be used as a matrix A.
     const cusparseOrder_t order     = CUSPARSE_ORDER_COL;
@@ -1141,7 +1141,7 @@ size_t compressMatrixB(void**              output,
                        const size_t        n,
                        const GemmOp        trans)
 {
-    FT_LOG_TRACE("compressMatrix [k=%ld, n=%ld, dtype=FP16]", k, n);
+    TM_LOG_TRACE("compressMatrix [k=%ld, n=%ld, dtype=FP16]", k, n);
 
     // swap A/B due to column/row major layout mismatch.
     cusparseOrder_t order = CUSPARSE_ORDER_COL;
@@ -1181,4 +1181,4 @@ size_t compressMatrixB(void**              output,
 
 /* ************************* End of GEMM utils **************************** */
 
-}  // end of namespace fastertransformer
+}  // end of namespace turbomind

src/turbomind/utils/gemm.h

@@ -27,13 +27,13 @@
 
 // TODO: Need to remove the dependency of the layer module.
 //   e.g. refactor Weight class to some base module.
-#include "src/fastertransformer/layers/DenseWeight.h"
-#include "src/fastertransformer/utils/Tensor.h"
-#include "src/fastertransformer/utils/allocator.h"
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/logger.h"
-#include "src/fastertransformer/utils/memory_utils.h"
+#include "src/turbomind/layers/DenseWeight.h"
+#include "src/turbomind/utils/Tensor.h"
+#include "src/turbomind/utils/allocator.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/logger.h"
+#include "src/turbomind/utils/memory_utils.h"
 
 #ifndef CUDART_VERSION
 #error CUDART_VERSION Undefined!
@@ -42,7 +42,7 @@
 // cublas default workspace size: 32MB. Let me make this as a Gemm property.
 #define WORKSPACE_SIZE 33554432
 
-namespace fastertransformer {
+namespace turbomind {
 
 // A wrapper of cublas or cusparse matrix operator.
 //  - GEMM_OP_N = CUBLAS_OP_N or CUSPARSE_OP_N
@@ -677,4 +677,4 @@ size_t compressMatrixB(void**              output,
 
 /* ************************* End of GEMM utils **************************** */
 
-}  // end of namespace fastertransformer
+}  // end of namespace turbomind

src/turbomind/utils/gemm_test/decoding_gemm_func.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/gemm_test/decoding_gemm_func.h"
+#include "src/turbomind/utils/gemm_test/decoding_gemm_func.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_decoding_gemm_config(int   batch_size,
@@ -401,4 +401,4 @@ size_t calDecodingGemmTestBufSizeInByte(int            batch_size,
     return buf_size_in_byte;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/decoding_gemm_func.h

@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/gemm_test/gemm_func.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/gemm_test/gemm_func.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -31,7 +31,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_decoding_gemm_config(int   batch_size,
@@ -55,4 +55,4 @@ size_t calDecodingGemmTestBufSizeInByte(int            batch_size,
                                         int            vocab_size,
                                         CublasDataType data_type);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/encoder_gemm_func.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/gemm_test/encoder_gemm_func.h"
+#include "src/turbomind/utils/gemm_test/encoder_gemm_func.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_encoder_gemm_config(
@@ -560,4 +560,4 @@ template void generate_encoder_gemm_config<__nv_bfloat16>(
     int batch_size, int seq_len, int head_num, int size_per_head, void* buffer, bool isAppend, int tensor_para_size);
 #endif
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/encoder_gemm_func.h

@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/gemm_test/gemm_func.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/gemm_test/gemm_func.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -31,7 +31,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_encoder_gemm_config(int   batch_size,
@@ -42,4 +42,4 @@ void generate_encoder_gemm_config(int   batch_size,
                                   bool  isAppend         = true,
                                   int   tensor_para_size = 1);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/encoder_igemm_func.cc

@@ -20,7 +20,7 @@
 #error CUDART_VERSION Undefined!
 #endif
 
-namespace fastertransformer {
+namespace turbomind {
 
 int batch_size_;
 int seq_len_;
@@ -1329,4 +1329,4 @@ int generate_encoder_igemm_config(
     return 0;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/encoder_igemm_func.h

@@ -16,8 +16,8 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_utils.h"
 #include <algorithm>
 #include <cublasLt.h>
 #include <cuda_runtime.h>
@@ -29,7 +29,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 /* CAUTION : must match cublasLtMatmulTile_t */
 const char* const matmulTileName[] = {"UNDEF",  "8x8",     "8x16",    "16x8",    "8x32",    "16x16",  "32x8",
@@ -77,4 +77,4 @@ int LtBatchIgemmCustomFind(cublasLtHandle_t ltHandle,
 
 void matInit(int rows, int cols, int8_t* p, int ld);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/gemm_func.cc

@@ -22,7 +22,7 @@
 #error CUDART_VERSION Undefined!
 #endif
 
-namespace fastertransformer {
+namespace turbomind {
 
 // Utility function to print customMatmulPerf_t structure
 int printPerfStructure(int                       batch_size,
@@ -986,4 +986,4 @@ size_t calGemmTestBufSizeInByteXlnet(
     return max_size * size_per_ele;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/gemm_func.h

@@ -17,8 +17,8 @@
 #pragma once
 
 #include "encoder_igemm_func.h"  // TODO(bhsueh) Remove this include
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_utils.h"
 #include <cstdio>
 #include <cstdlib>
 #include <ctime>
@@ -34,7 +34,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 // Scale Type Converter
 // is_fp16_compute_type is only valid when T = half
@@ -98,4 +98,4 @@ int printPerfStructure(int                       batch_size,
                        int                       hasPrint,
                        int                       batch_count = 1);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/gpt_gemm_func.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/gemm_test/gpt_gemm_func.h"
+#include "src/turbomind/utils/gemm_test/gpt_gemm_func.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 bool isSparseGemmAvailable(size_t m, size_t n, size_t k)
 {
@@ -804,4 +804,4 @@ size_t calGptGemmTestBufSizeInByte(int            batch_size,
     return buf_size_in_byte;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/gpt_gemm_func.h

@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/gemm_test/gemm_func.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/gemm_test/gemm_func.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -36,7 +36,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_gpt_gemm_config(int   batch_size,
@@ -60,4 +60,4 @@ size_t calGptGemmTestBufSizeInByte(int            batch_size,
                                    int            tensor_para_size,
                                    CublasDataType data_type);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/swin_gemm_func.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/gemm_test/swin_gemm_func.h"
+#include "src/turbomind/utils/gemm_test/swin_gemm_func.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_swin_gemm_config(
@@ -398,4 +398,4 @@ template void generate_swin_gemm_config<__nv_bfloat16>(
     int batch_size, int seq_len, int head_num, int size_per_head, void* buffer, bool isAppend);
 #endif
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/swin_gemm_func.h

@@ -16,10 +16,10 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/gemm_test/gemm_func.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/gemm_test/gemm_func.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -31,10 +31,10 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void generate_swin_gemm_config(
     int batch_size, int seq_len, int head_num, int size_per_head, void* buffer, bool isAppend = true);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/swin_igemm_func.cc

@@ -16,7 +16,7 @@
 
 #include "swin_igemm_func.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 static const char* showStatus(cublasStatus_t error)
 {
@@ -278,4 +278,4 @@ int generate_swin_igemm_config(
     return 0;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/utils/gemm_test/swin_igemm_func.h

@@ -16,9 +16,9 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cublasAlgoMap.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
-#include "src/fastertransformer/utils/gemm_test/encoder_igemm_func.h"
+#include "src/turbomind/utils/cublasAlgoMap.h"
+#include "src/turbomind/utils/cuda_utils.h"
+#include "src/turbomind/utils/gemm_test/encoder_igemm_func.h"
 #include <algorithm>
 #include <cublasLt.h>
 #include <cuda_runtime.h>
@@ -30,7 +30,7 @@
 #include <unistd.h>
 #include <vector>
 
-namespace fastertransformer {
+namespace turbomind {
 
 /* CAUTION : must match cublasLtMatmulTile_t */
 // const char* const matmulTileName[] = {
@@ -42,4 +42,4 @@ namespace fastertransformer {
 int generate_swin_igemm_config(
     int batch_size, int seq_len, int head_num, int size_per_head, void* buffer, bool isAppend = true);
 
-}  // namespace fastertransformer
+}  // namespace turbomind