simplify the to_cpu op

candle-core/src/metal_backend/device.rs

@@ -336,6 +336,8 @@ impl MetalDevice {
 
  /// Copies data from one buffer to another.
  /// This method is blocking as it relies on GPU-CPU synchronization.
+ /// NOTE: this method does not end the current command buffer, it is the responsibility of the caller
+ /// to do so via the [`close_compute_buffer`] function.
  pub fn copy_buffer(
  &self,
  source_buffer: &BufferRef,

candle-core/src/metal_backend/mod.rs

@@ -1594,21 +1594,11 @@ impl MetalStorage {
  let size = (self.count * self.dtype.size_in_bytes()) as NSUInteger;
  let buffer = self.device.new_buffer_managed(size)?;
 
- // Handle the creation and initialization of a new command buffer
- {
- // Finalize current compute operations
- self.device.end_compute_encoding()?;
-
- let command_buffer = self.device.command_buffer()?;
+ // Enqueue the copy operation
+ self.device.copy_buffer(&self.buffer, 0, &buffer, 0, size)?;
 
- // Setup the blit encoder to perform the copy operation
- let blit = command_buffer.new_blit_command_encoder();
- blit.copy_from_buffer(&self.buffer, 0, &buffer, 0, size);
- blit.end_encoding();
-
- // Commit the command buffer to the queue
- self.device.close_compute_buffer()?;
- }
+ // Commit the command buffer to the queue to ensure the copy operation is completed
+ self.device.close_compute_buffer()?;
 
  let result = read_to_vec(&buffer, self.count);
  Ok(result)