Improve metal buffer usage (#1807)

* Improve metal buffer usage

* Clone cpu storage when loading to reduce wait_until_complete calls
* Use powers of two for buffer sizes so reuse is more likely.
* Select best available buffer by size.
* Add count to MetalStorage -> can use buffer with different size

Co-authored-by: Chris Fleetwood <christopher.fleetwood@huggingface.co>

* Simplify new buffer creation without blit copy. Revert &[] -> Vec

* Add documentation on newBufferWithBytes safety / synchronization

* Drop unused buffers after command buffer is done syncing.

---------

Co-authored-by: Chris Fleetwood <christopher.fleetwood@huggingface.co>

1 files changed, 7 insertions, 2 deletions

diff --git a/candle-core/src/quantized/metal.rs b/candle-core/src/quantized/metal.rs
index af1cf369..7be0f74e 100644
--- a/candle-core/src/quantized/metal.rs
+++ b/candle-core/src/quantized/metal.rs
@@ -106,7 +106,12 @@ impl QMetalStorage {
         }
 
         let buffer = self.device.new_buffer_with_data(&out)?;
-        Ok(MetalStorage::new(buffer, self.device.clone(), DType::F32))
+        Ok(MetalStorage::new(
+            buffer,
+            self.device.clone(),
+            elem_count,
+            DType::F32,
+        ))
     }
 
     pub fn quantize(&mut self, src: &MetalStorage) -> Result<()> {
@@ -170,7 +175,7 @@ impl QMetalStorage {
             &dst,
         )
         .map_err(MetalError::from)?;
-        let dst_storage = crate::MetalStorage::new(dst, device, DType::F32);
+        let dst_storage = crate::MetalStorage::new(dst, device, dst_shape.elem_count(), DType::F32);
         Ok((dst_storage, dst_shape))
     }
 }