Update src/streamlit_app.py

src/streamlit_app.py

@@ -204,7 +204,7 @@ def render_attribution_html(result, log_color: bool = False, cmap_name: str = "B
             grad_magnitude[target_idx_in_grad] = 1e-8
         bar_idx = target_idx_in_grad
 
-    grad_np = grad_magnitude.cpu().numpy()
+    grad_np = grad_magnitude.float().cpu().numpy()
     log_norm = Log_Norm(vmin=grad_np.min(), vmax=grad_np.max())
     norm = Norm(vmin=grad_np.min(), vmax=grad_np.max())
 
@@ -289,9 +289,9 @@ def render_attribution_barplot(result, log_color: bool = False, cmap_name: str =
     front_pad = 2  # assumed
 
     if len(grads.shape) == 2:
-        grad_magnitude = grads.norm(dim=-1).squeeze().detach().clone().cpu().numpy()
+        grad_magnitude = grads.norm(dim=-1).squeeze().detach().clone().float().cpu().numpy()
     else:
-        grad_magnitude = grads.detach().clone().cpu().numpy()
+        grad_magnitude = grads.detach().clone().float().cpu().numpy()
 
     hardset_target_grad = True
     target_bar_index = None
@@ -506,7 +506,7 @@ def main():
             probs = torch.softmax(logit_vector, dim=-1)
             top_probs, top_indices = torch.topk(probs, k)
             top_tokens = [tokenizer.decode([idx]) for idx in top_indices]
-            for i, (tok, prob) in enumerate(zip(top_tokens, top_probs.cpu().numpy()), 1):
+            for i, (tok, prob) in enumerate(zip(top_tokens, top_probs.float().cpu().numpy()), 1):
                 st.write(f"{i}. P(**{repr(tok)}**)={prob:.3f}")