[chore] enable notify when submitting tensor write task

csrc/pthread_backend.cpp

@@ -89,22 +89,31 @@ void PthreadAsyncIO::sync_h2d() {
 
 void PthreadAsyncIO::write_tensor(int fd, torch::Tensor t, unsigned long long offset, callback_t callback, std::optional<torch::Tensor> pinned) {
     auto stream = c10::cuda::getCurrentCUDAStream();
+    if (!t.is_cuda()) {
+        this->h2d_in_progress.fetch_sub(1);  // already moved to cpu
+        if (this->h2d_in_progress.load() == 0) {  // notify when all h2d are completed and safe to optimizer.step()
+            std::lock_guard<std::mutex> lock(this->mtx);
+            cv.notify_one();
+        }
+    }
     auto fut = this->pool.submit_task(
         [this, fd, t, offset, pinned, stream] {
-            at::cuda::CUDAStreamGuard guard(stream);  // https://pytorch.org/cppdocs/notes/tensor_cuda_stream.html
             torch::Tensor cpu_tensor;
             if (t.is_cuda()) {
+                at::cuda::CUDAStreamGuard guard(stream);  // https://pytorch.org/cppdocs/notes/tensor_cuda_stream.html
                 if (pinned.has_value()) {
                     pinned.value().copy_(t, /*non_blocking*/ false);
                     cpu_tensor = pinned.value();
                 } else {
                     cpu_tensor = t.to(t.options().device(c10::DeviceType::CPU), /*non_blocking*/ false, /*copy*/ false);  // modified from torch::Tensor::cpu()
                 }
-            }
-            this->h2d_in_progress.fetch_sub(1);
-            if (this->h2d_in_progress.load() == 0) {  // notify when all h2d are completed and safe to optimizer.step()
-                std::lock_guard<std::mutex> lock(this->mtx);
-                cv.notify_one();
+                this->h2d_in_progress.fetch_sub(1);
+                if (this->h2d_in_progress.load() == 0) {  // notify when all h2d are completed and safe to optimizer.step()
+                    std::lock_guard<std::mutex> lock(this->mtx);
+                    cv.notify_one();
+                }
+            } else {
+                cpu_tensor = t;
             }
             void *buf = cpu_tensor.data_ptr();
             size_t n_bytes = cpu_tensor.numel() * cpu_tensor.element_size();