refactor: 统一硬件相关头文件引用

将分散在各文件中的CUDA/HIP/MUSA硬件相关头文件引用统一到vendors目录下的对应头文件中，提高代码可维护性。移除重复的头文件引用，优化构建配置。

csrc/attention/attention_kernels.cuh

@@ -17,9 +17,7 @@
  * limitations under the License.
  */
 
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
+#include "../vendors/functions.h"
 #include <algorithm>
 
 #include "attention_dtypes.h"
@@ -27,9 +25,7 @@
 #include "../cuda_compat.h"
 
 #ifdef USE_ROCM
-  #include <hip/hip_bf16.h>
   #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
-typedef __hip_bfloat16 __nv_bfloat16;
 #else
   #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

csrc/attention/dtype_bfloat16.cuh

@@ -23,16 +23,8 @@
 #include "attention_generic.cuh"
 #include "dtype_float32.cuh"
 
-#ifndef USE_ROCM
-  #include <cuda_bf16.h>
-  #include <cuda_fp16.h>
-#else
-  #include <hip/hip_bf16.h>
-  #include <hip/hip_fp16.h>
 
-typedef __hip_bfloat162 __nv_bfloat162;
-typedef __hip_bfloat16 __nv_bfloat16;
-#endif
+#include "../vendors/functions.h"
 
 #include <stdint.h>
 

csrc/attention/dtype_float16.cuh

@@ -23,9 +23,7 @@
 #include "attention_generic.cuh"
 #include "dtype_float32.cuh"
 
-#ifdef USE_ROCM
-  #include <hip/hip_fp16.h>
-#endif
+#include "../vendors/functions.h"
 
 #include <stdint.h>
 

csrc/attention/dtype_fp8.cuh

@@ -3,11 +3,7 @@
 #include "attention_generic.cuh"
 
 #include <stdint.h>
-#ifdef ENABLE_FP8
-  #ifndef USE_ROCM
-    #include <cuda_fp8.h>
-  #endif  // USE_ROCM
-#endif    // ENABLE_FP8
+#include "../vendors/functions.h"
 
 namespace vllm {
 

csrc/attention/merge_attn_states.cu

@@ -1,7 +1,5 @@
 #include <optional>
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
+#include "../vendors/functions.h"
 #include <algorithm>
 
 #include "attention_dtypes.h"

csrc/attention/vertical_slash_index.cu

@@ -3,9 +3,7 @@
 
 #include <assert.h>
 
-#include <cuda.h>
-
-#include <torch/all.h>
+#include "../vendors/functions.h"
 
 __device__ int64_t save_blocks(int* block_offset, int64_t range_start,
                                int64_t range_end, int64_t block_size,