昇腾和寒武纪相关代码退场 npu相关代码退场3 (#53699)

* rm npu

* rm use_npu

* rm npuid

* rm use_npu

* rm npuid

* delete npupinned

* roll back sth.

* roll back sth.

* delete npupinned

* roll back sth.

* roll back sth.

* rm npu

* rollback something

* rollback npu identity

* rollback npu identity

paddle/fluid/framework/library_type.h

@@ -68,8 +68,6 @@ inline LibraryType StringToLibraryType(const char* ctype) {
     return LibraryType::kPlain;
   } else if (s == std::string("IPU")) {
     return LibraryType::kPlain;
-  } else if (s == std::string("NPU")) {
-    return LibraryType::kPlain;
   } else if (s == std::string("CUDA")) {
     return LibraryType::kPlain;
   } else {

paddle/fluid/operators/beam_search_decode_op.h

@@ -34,13 +34,10 @@ struct BeamSearchDecodeFunctor {
         id_tensor_(id_tensor),
         score_tensor_(score_tensor) {
     tensor_on_gpu_ = false;
-    tensor_on_npu_ = false;
     // First make a copy of GPU data on CPU
     if (platform::is_gpu_place(step_ids_origin_[0].place())) {
       if (platform::is_gpu_place(step_ids_origin_[0].place())) {
         tensor_on_gpu_ = true;
-      } else {
-        tensor_on_npu_ = true;
       }
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();
@@ -63,8 +60,6 @@ struct BeamSearchDecodeFunctor {
     if (platform::is_gpu_place(step_scores_origin_[0].place())) {
       if (platform::is_gpu_place(step_scores_origin_[0].place())) {
         tensor_on_gpu_ = true;
-      } else {
-        tensor_on_npu_ = true;
       }
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();
@@ -96,7 +91,7 @@ struct BeamSearchDecodeFunctor {
     } else {
       BeamSearchDecoder<T> beam_search_decoder(beam_size_, end_id_);
       // Check if the tensor is on GPU. If so, use the CPU copy instead
-      if (tensor_on_gpu_ || tensor_on_npu_) {
+      if (tensor_on_gpu_) {
         beam_search_decoder.Backtrace(
             step_ids_, step_scores_, id_tensor_, score_tensor_);
       } else {
@@ -107,7 +102,6 @@ struct BeamSearchDecodeFunctor {
   }
 
   bool tensor_on_gpu_;
-  bool tensor_on_npu_;
   size_t beam_size_;
   int end_id_;
   // TODO(Superjomn) Here might result serious performance issue in the

paddle/fluid/operators/memcpy_op.cc

@@ -111,7 +111,6 @@ class MemcpyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
                  "1: dst is on CUDAPlace. "
                  "2: dst is on CUDAPinnedPlace. "
                  "3: dst is on XPUPlace. "
-                 "4: dst is on NPUPinnerPlace. "
                  "5: dst is on CustomDevicePlace");
     AddComment(R"DOC(
     Memcpy Operator.

paddle/fluid/operators/memcpy_op.h

@@ -39,8 +39,6 @@ class MemcpyFunctor {
     CUDA = 1,
     CUDA_PINNED = 2,
     XPU = 3,
-    NPU = 4,
-    NPU_PINNED = 5,
     CUSTOM_DEVICE = 6,
   };
 

paddle/fluid/operators/reader/buffered_reader.cc

@@ -98,7 +98,6 @@ BufferedReader::BufferedReader(
 
   cpu_buffer_.resize(buffer_size);
   cuda_buffer_.resize(buffer_size);
-  npu_buffer_.resize(buffer_size);
   xpu_buffer_.resize(buffer_size);
   custom_device_buffer_.resize(buffer_size);
   ReadTillBufferFullAsync();

paddle/fluid/operators/reader/buffered_reader.h

@@ -77,7 +77,6 @@ class BufferedReader : public framework::DecoratedReader {
   // buffers and prevent alloc every time.
   std::vector<TensorVec> cpu_buffer_;
   std::vector<TensorVec> cuda_buffer_;
-  std::vector<TensorVec> npu_buffer_;
   std::vector<TensorVec> xpu_buffer_;
   std::vector<TensorVec> custom_device_buffer_;
   size_t prev_pos_{-1UL};

paddle/fluid/platform/CMakeLists.txt

@@ -123,7 +123,6 @@ cc_library(
        framework_proto
        ${IPU_CTX_DEPS}
        ${GPU_CTX_DEPS}
-       ${NPU_CTX_DEPS}
        ${MKLDNN_CTX_DEPS}
        ${dgc_deps}
        dlpack

paddle/fluid/platform/device_context.h

@@ -93,7 +93,6 @@ namespace platform {
 enum DeviceType {
   CPU = 0,
   CUDA = 1,
-  NPU = 2,
   XPU = 3,
   IPU = 4,
   CUSTOM_DEVICE = 6,

paddle/phi/common/backend.h

@@ -52,7 +52,6 @@ enum class Backend : uint8_t {
 
   // various acceleration devices' backends
   XPU,  // XPU currently does not exist at the same time as CUDA
-  NPU,  // NPU currently does not exist at the same time as CUDA
   IPU,
 
   // paddle kernel primitives backend

paddle/phi/common/place.h

@@ -31,7 +31,6 @@ enum class AllocationType : int8_t {
   GPU = 2,
   GPUPINNED = 3,
   XPU = 4,
-  NPU = 5,
   IPU = 7,
   CUSTOM = 9,
 };

python/paddle/distributed/auto_parallel/cluster.py

@@ -450,9 +450,8 @@ class Cluster:
         """Generate cluster by default config."""
         gpu_models = ["V100", "A100", "H100", "A2", "A10", "A16", "A30", "A40"]
         xpu_models = ["XPU"]
-        npu_models = ["NPU"]
         dcu_models = ["DCU"]
-        all_gpu_models = gpu_models + xpu_models + npu_models + dcu_models
+        all_gpu_models = gpu_models + xpu_models + dcu_models
         self._num_devices_per_machine = device_count
 
         def _convert_to_type(gpu_model):
@@ -461,8 +460,6 @@ class Cluster:
                 type = "GPU"
             elif gpu_model in xpu_models:
                 type = "XPU"
-            elif gpu_model in npu_models:
-                type = "NPU"
             elif gpu_model in dcu_models:
                 type = "DCU"
             else:

python/paddle/distributed/auto_parallel/cluster_v2.py

@@ -25,7 +25,6 @@ class DeviceType(IntEnum):
     CPU = 1
     GPU = 2
     XPU = 3
-    NPU = 4
     DCU = 5
     NIC = 6