chore: see if it fixes

crates/ratchet-core/src/ops/matmul.rs

@@ -352,12 +352,12 @@ def matmul(a, b):
         let ground = ground_truth(&a, &b)?;
 
         let device = Device::request_device(DeviceRequest::GPU)?;
-        let a_gpu = a.to(device.clone())?;
-        let b_gpu = b.to(device.clone())?;
+        let a_gpu = a.to(&device)?;
+        let b_gpu = b.to(&device)?;
         let c_gpu = a_gpu.matmul(&b_gpu)?;
         c_gpu.resolve()?;
 
-        let d_gpu = c_gpu.to(Device::CPU)?;
+        let d_gpu = c_gpu.to(&Device::CPU)?;
         ground.all_close(&d_gpu, 1e-4, 1e-4)?;
         Ok(())
     }
@@ -370,11 +370,11 @@ def matmul(a, b):
         let quantizer = Quantizer::new(Quantization::SInt8);
         let bq = quantizer.sint8_quantize(b);
         let device = Device::request_device(DeviceRequest::GPU)?;
-        let a_gpu = a.to(device.clone())?;
-        let b_gpu = bq.to(device.clone())?;
+        let a_gpu = a.to(&device)?;
+        let b_gpu = bq.to(&device)?;
         let c_gpu = a_gpu.matmul(&b_gpu)?;
         c_gpu.resolve()?;
-        let ours = c_gpu.to(Device::CPU)?;
+        let ours = c_gpu.to(&Device::CPU)?;
 
         println!("RATCHET WQ8\n{:?}\n", ours);
         println!("PYTORCH FP32:\n{:?}", ground);

crates/ratchet-core/src/ops/softmax.rs

@@ -129,10 +129,10 @@ def softmax(a):
         let a = Tensor::randn::<f32>(shape![64, 64], Device::CPU);
         let ground = ground_truth(&a)?;
 
-        let a_gpu = a.to(gpu_device.clone())?;
+        let a_gpu = a.to(&gpu_device)?;
         let b = a_gpu.softmax(1)?;
         b.resolve()?;
-        let ours = b.to(Device::CPU)?;
+        let ours = b.to(&Device::CPU)?;
         println!("GROUND: \n{:?}", ground);
         println!("OURS: \n{:?}", ours);
         ground.all_close(&ours, 1e-6, 1e-6)?;

crates/ratchet-core/src/tensor.rs

@@ -414,10 +414,10 @@ impl Tensor {
     /// If the tensor is already on the specified device, it will be returned as-is,
     /// and the underlying storage will not be copied.
     /// If the tensor is on a different device, it will be copied to the specified device.
-    pub fn to(&self, device: Device) -> Result<Tensor, TensorError> {
-        match (self.device(), &device) {
+    pub fn to(&self, device: &Device) -> Result<Tensor, TensorError> {
+        match (self.device(), device) {
             (Device::GPU(_), Device::CPU) => self.to_cpu(),
-            (Device::CPU, Device::GPU(_)) => self.to_gpu(&device),
+            (Device::CPU, Device::GPU(_)) => self.to_gpu(device),
             _ => Ok(self.clone()),
         }
     }