diff --git a/candle-core/src/quantized/ggml_file.rs b/candle-core/src/quantized/ggml_file.rs
index e4ad35cd..1dd3d9c0 100644
--- a/candle-core/src/quantized/ggml_file.rs
+++ b/candle-core/src/quantized/ggml_file.rs
@@ -135,7 +135,13 @@ pub fn qtensor_from_ggml(
     dims: Vec<usize>,
 ) -> Result<super::QTensor> {
     let tensor_elems = dims.iter().product::<usize>();
-    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.blck_size();
+    let blck_size = ggml_dtype.blck_size();
+    if tensor_elems % blck_size != 0 {
+        crate::bail!(
+            "the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
+        )
+    }
+    let size_in_bytes = tensor_elems / blck_size * ggml_dtype.type_size();
 
     match ggml_dtype {
         GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims),
diff --git a/candle-core/src/quantized/gguf_file.rs b/candle-core/src/quantized/gguf_file.rs
index b3993034..3a5f2030 100644
--- a/candle-core/src/quantized/gguf_file.rs
+++ b/candle-core/src/quantized/gguf_file.rs
@@ -59,8 +59,13 @@ impl TensorInfo {
         tensor_data_offset: u64,
     ) -> Result<QTensor> {
         let tensor_elems = self.shape.elem_count();
-        let size_in_bytes =
-            tensor_elems * self.ggml_dtype.type_size() / self.ggml_dtype.blck_size();
+        let blck_size = self.ggml_dtype.blck_size();
+        if tensor_elems % blck_size != 0 {
+            crate::bail!(
+            "the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
+        )
+        }
+        let size_in_bytes = tensor_elems / blck_size * self.ggml_dtype.type_size();
         let mut raw_data = vec![0u8; size_in_bytes];
         reader.seek(std::io::SeekFrom::Start(tensor_data_offset + self.offset))?;
         reader.read_exact(&mut raw_data)?;