modify blas

backends/metax_gpu/CMakeLists.txt

@@ -733,6 +733,7 @@ target_compile_definitions(
   ${TARGET_NAME}
   PUBLIC PADDLE_WITH_CUDA=1
          PADDLE_WITH_CUSTOM_DEVICE=1
+         mcblasContext=cublasContext
          GPUContext=CustomContext
          KPSContext=CustomContext
          STREAM_TYPE=cudaStream_t

backends/metax_gpu/kernels/metax_kernel/metax_context.cc

@@ -56,16 +56,4 @@ void DnnWorkspaceHandle::ReallocWorkspace(size_t required_workspace_bytes) {
   allocation_.reset();
   allocation_ = allocator_->Allocate(required_workspace_bytes);
 }
-
-static std::function<blasLtHandle_t()> blaslt_handle_creator_{nullptr};
-static blasLtHandle_t blaslt_handle_{nullptr};
-static std::once_flag flag_blaslt_;
-
-static void InitBlasLtHandle(blasLtHandle_t* blaslt_handle) {
-#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 11060
-  mcblasLtCreate(blaslt_handle);
-#elif defined(PADDLE_WITH_HIP)
-  phi::dynload::hipblasLtCreate(blaslt_handle);
-#endif
-}
 }  // namespace phi

backends/metax_gpu/kernels/metax_kernel/metax_context.h

@@ -27,9 +27,7 @@
 #include "paddle/phi/core/attribute.h"
 #include "paddle/phi/core/device_context.h"
 
-using blasLtHandle_t = struct mcblasLtContext*;
-
-blasLtHandle_t GetBlasLtHandle();
+cublasLtHandle_t GetBlasLtHandle();
 
 namespace phi {
 class DnnWorkspaceHandle {

backends/metax_gpu/patch/paddle.patch

@@ -488,7 +488,6 @@ index 4eae698648..5c047723ea 100644
  #endif
    return block_dim >= kMaxBlockDim ? kMaxBlockDim : lwarpSize;
  }
-
 diff --git a/paddle/phi/kernels/funcs/math/context_project.h b/paddle/phi/kernels/funcs/math/context_project.h
 index 15e1a4a3c3..e4780538d7 100644
 --- a/paddle/phi/kernels/funcs/math/context_project.h