diff --git a/README.md b/README.md
index 95e48c27..2b588dde 100644
--- a/README.md
+++ b/README.md
@@ -281,7 +281,7 @@ two smoothing methods are supported:
 Usage: `python cam.py --image-path <path_to_image> --method <method> --output-dir <output_dir_path> `
 
 
-To use with a specific device, like cpu, cuda, cuda:0 or mps:
+To use with a specific device, like cpu, cuda, cuda:0, mps or hpu:
 `python cam.py --image-path <path_to_image> --device cuda  --output-dir <output_dir_path> `
 
 ----------
diff --git a/cam.py b/cam.py
index 9f3e377c..935afeb2 100644
--- a/cam.py
+++ b/cam.py
@@ -77,6 +77,9 @@ def get_args():
         'kpcacam': KPCA_CAM
     }
 
+    if args.device=='hpu':
+        import habana_frameworks.torch.core as htcore
+
     model = models.resnet50(pretrained=True).to(torch.device(args.device)).eval()
 
     # Choose the target layer you want to compute the visualization for.
diff --git a/pytorch_grad_cam/base_cam.py b/pytorch_grad_cam/base_cam.py
index 4b2850a9..44ae5b90 100644
--- a/pytorch_grad_cam/base_cam.py
+++ b/pytorch_grad_cam/base_cam.py
@@ -25,6 +25,13 @@ def __init__(
 
         # Use the same device as the model.
         self.device = next(self.model.parameters()).device
+        if 'hpu' in str(self.device):
+            try:
+                import habana_frameworks.torch.core as htcore
+            except ImportError as error:
+                error.msg = f"Could not import habana_frameworks.torch.core. {error.msg}."
+                raise error
+            self.__htcore = htcore
         self.reshape_transform = reshape_transform
         self.compute_input_gradient = compute_input_gradient
         self.uses_gradients = uses_gradients
@@ -97,6 +104,8 @@ def forward(
             self.model.zero_grad()
             loss = sum([target(output) for target, output in zip(targets, outputs)])
             loss.backward(retain_graph=True)
+            if 'hpu' in str(self.device):
+                self.__htcore.mark_step()
 
         # In most of the saliency attribution papers, the saliency is
         # computed with a single target layer.