sync misc changes (#42534)

paddle/fluid/platform/device/ipu/ipu_device.cc

@@ -39,7 +39,8 @@ const bool GetBoolEnv(const std::string& str) {
 
 int GetNumDevices() {
   bool ipu_model = GetBoolEnv("POPLAR_IPUMODEL");
-  if (ipu_model) {
+  bool compile_only = GetBoolEnv("IPU_COMPILE_ONLY");
+  if (ipu_model || compile_only) {
     return 1;
   }
   int num_devices =
@@ -52,7 +53,8 @@ int GetNumDevices() {
 
 std::vector<int> GetDeviceIds() {
   bool ipu_model = GetBoolEnv("POPLAR_IPUMODEL");
-  if (ipu_model) {
+  bool compile_only = GetBoolEnv("IPU_COMPILE_ONLY");
+  if (ipu_model || compile_only) {
     return {0};
   }
   std::vector<int> device_ids;

paddle/fluid/platform/device/ipu/ipu_executor.cc

@@ -96,6 +96,7 @@ Executor::~Executor() {
 
 void Executor::Prepare(const std::string &proto) {
   VLOG(10) << "enter Executor::Prepare";
+  compile_only_ = GetBoolEnv("IPU_COMPILE_ONLY");
 
   AcquireDevice();
   executor_resources_ = std::make_unique<ExecutorResources>();
@@ -122,9 +123,18 @@ void Executor::Prepare(const std::string &proto) {
   }
   VLOG(10) << "Creating session from Onnx Model...done";
 
-  VLOG(10) << "Preparing session device...";
-  session_->prepareDevice();
-  VLOG(10) << "Preparing session device...done";
+  if (compile_only_) {
+    LOG(INFO)
+        << "Save the offline cache as offline_cache.popart in current path.";
+    VLOG(10) << "Compile only...";
+    session_->compileAndExport("./offline_cache.popart");
+    VLOG(10) << "Compile only...done";
+    return;
+  } else {
+    VLOG(10) << "Preparing session device...";
+    session_->prepareDevice();
+    VLOG(10) << "Preparing session device...done";
+  }
 
   SetWeightsIO();
 
@@ -141,6 +151,11 @@ void Executor::Prepare(const std::string &proto) {
 void Executor::Run(const std::vector<const Tensor *> &inputs,
                    const std::vector<Tensor *> &outputs,
                    const framework::ExecutionContext &ctx) {
+  if (compile_only_) {
+    LOG(INFO) << "If IPU_COMPILE_ONLY=True, skip exe.run";
+    return;
+  }
+
   VLOG(10) << "enter Executor::Run";
   // inputs
   std::map<popart::TensorId, popart::IArray &> popart_inputs;
@@ -222,6 +237,7 @@ void Executor::AcquireDevice() {
   bool use_ipu_model = GetBoolEnv("POPLAR_IPUMODEL");
   bool enable_distribution = ipu_strategy_->enable_distribution;
   if (use_ipu_model) {
+    VLOG(10) << "Create IPU model device...";
     std::map<std::string, std::string> deviceOpts{
         {
             "numIPUs", std::to_string(ipu_strategy_->num_ipus),
@@ -230,7 +246,21 @@ void Executor::AcquireDevice() {
     };
     device_ = popart::DeviceManager::createDeviceManager().createIpuModelDevice(
         deviceOpts);
+    VLOG(10) << "Create IPU model device...done";
+  } else if (compile_only_) {
+    VLOG(10) << "Create offline device...";
+    std::map<std::string, std::string> deviceOpts{
+        {
+            "numIPUs", std::to_string(ipu_strategy_->num_ipus),
+        },
+        {"ipuVersion", "ipu2"},
+    };
+    device_ =
+        popart::DeviceManager::createDeviceManager().createOfflineIPUDevice(
+            deviceOpts);
+    VLOG(10) << "Create offline device...done";
   } else if (enable_distribution) {
+    VLOG(10) << "Create distribution device...";
     auto ipus_per_replica = ipu_strategy_->num_ipus /
                             ipu_strategy_->popart_options.replicatedGraphCount;
     auto device_id = popdist_get_device(ipus_per_replica);
@@ -240,13 +270,16 @@ void Executor::AcquireDevice() {
         device_,
         errors::Unavailable("Can't attach IPU in distribution, ipu_num = %d.",
                             RequestIpus(ipu_strategy_->num_ipus)));
+    VLOG(10) << "Create distribution device...done";
   } else {
+    VLOG(10) << "Create IPU device...";
     device_ =
         popart::DeviceManager::createDeviceManager().acquireAvailableDevice(
             RequestIpus(ipu_strategy_->num_ipus));
     PADDLE_ENFORCE_NOT_NULL(
         device_, errors::Unavailable("Can't attach IPU, ipu_num = %d.",
                                      RequestIpus(ipu_strategy_->num_ipus)));
+    VLOG(10) << "Create IPU device...done";
   }
   VLOG(10) << "leave Executor::AcquireDevice";
 }

paddle/fluid/platform/device/ipu/ipu_executor.h

@@ -91,6 +91,7 @@ class Executor {
   const Scope *scope_ = nullptr;
   const IpuStrategy *ipu_strategy_ = nullptr;
   CompilerResources *compiler_resources_ = nullptr;
+  bool compile_only_ = false;
 
   // Deviceinfo for popart session
   std::shared_ptr<popart::DeviceInfo> device_;

paddle/fluid/platform/device/ipu/popart_canonicalization/other_ops.cc

@@ -57,14 +57,21 @@ Node *checkpointoutput_handler(Graph *graph, Node *node) {
 Node *custom_nll_loss_handler(Graph *graph, Node *node) {
   auto *op = node->Op();
   auto reduction = BOOST_GET_CONST(int, op->GetAttr("reduction"));
-  auto ignoreIndex = BOOST_GET_CONST(int, op->GetAttr("ignoreIndex"));
+  auto ignoreIndex = BOOST_GET_CONST(std::string, op->GetAttr("ignoreIndex"));
   auto inputIsLogProbability =
       BOOST_GET_CONST(bool, op->GetAttr("inputIsLogProbability"));
-  return CreateBaseOp(graph, node, "popart_nllloss_v2", node->inputs,
-                      node->outputs,
-                      {{"reduction", reduction},
-                       {"ignoreIndex", ignoreIndex},
-                       {"inputIsLogProbability", inputIsLogProbability}});
+  if (ignoreIndex == "None") {
+    return CreateBaseOp(graph, node, "popart_nllloss_v2", node->inputs,
+                        node->outputs,
+                        {{"reduction", reduction},
+                         {"inputIsLogProbability", inputIsLogProbability}});
+  } else {
+    return CreateBaseOp(graph, node, "popart_nllloss_v2", node->inputs,
+                        node->outputs,
+                        {{"reduction", reduction},
+                         {"ignoreIndex", std::atoi(ignoreIndex.c_str())},
+                         {"inputIsLogProbability", inputIsLogProbability}});
+  }
 }
 
 Node *identity_handler(Graph *graph, Node *node) {

python/paddle/fluid/tests/unittests/ipu/test_cast_op_ipu.py

@@ -85,23 +85,6 @@ class TestEnableFp16(TestBase):
         self.attrs['dtype'] = 'float32'
 
 
-class TestDisableTransferCast(TestEnableFp16):
-    def set_data_feed(self):
-        data = np.random.uniform(size=[1, 3, 3, 3])
-        self.feed_fp32 = {'x': data.astype(np.float32)}
-        self.feed_fp16 = {'x': data.astype(np.float16)}
-
-    def set_op_attrs(self):
-        self.attrs = {}
-        self.attrs['dtype'] = 'float32'
-
-    def run_model(self, exec_mode):
-        ipu_strategy = paddle.static.IpuStrategy()
-        ipu_strategy.set_graph_config(is_training=self.is_training)
-        ipu_strategy.set_options({"transfer_cast_op": False})
-        self.run_op_test(exec_mode)
-
-
 class TestCase2(TestBase):
     def set_data_feed(self):
         self.feed_fp32 = {