Async tensor copying. (#1900)

candle-core/src/backend.rs

@@ -129,6 +129,8 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
 
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;

candle-core/src/cpu_backend.rs

@@ -2814,6 +2814,10 @@ impl BackendDevice for CpuDevice {
         Ok(s.clone())
     }
 
+    fn storage_from_cpu_storage_owned(&self, s: CpuStorage) -> Result<Self::Storage> {
+        Ok(s)
+    }
+
     fn new(_: usize) -> Result<Self> {
         Ok(Self)
     }

candle-core/src/cuda_backend.rs

@@ -420,6 +420,43 @@ impl BackendDevice for CudaDevice {
             device: self.clone(),
         })
     }
+
+    fn storage_from_cpu_storage_owned(&self, storage: CpuStorage) -> Result<CudaStorage> {
+        let slice = match storage {
+            CpuStorage::U8(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorage::U32(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorage::I64(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorage::BF16(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorage::F16(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorage::F32(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorage::F64(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
 }
 
 #[derive(Debug)]

candle-core/src/device.rs

@@ -294,12 +294,12 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
             Device::Cuda(device) => {
                 let storage = array.to_cpu_storage();
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
             Device::Metal(device) => {
                 let storage = array.to_cpu_storage();
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Metal(storage))
             }
         }
@@ -310,12 +310,12 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(S::to_cpu_storage_owned(data))),
             Device::Cuda(device) => {
                 let storage = S::to_cpu_storage_owned(data);
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
             Device::Metal(device) => {
                 let storage = S::to_cpu_storage_owned(data);
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Metal(storage))
             }
         }

candle-core/src/dummy_cuda_backend.rs

@@ -214,6 +214,10 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }

candle-core/src/dummy_metal_backend.rs

@@ -226,6 +226,10 @@ impl crate::backend::BackendDevice for MetalDevice {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }

candle-core/src/metal_backend.rs

@@ -1867,6 +1867,10 @@ impl BackendDevice for MetalDevice {
         ))
     }
 
+    fn storage_from_cpu_storage_owned(&self, storage: CpuStorage) -> Result<Self::Storage> {
+        self.storage_from_cpu_storage(&storage)
+    }
+
     fn rand_uniform(
         &self,
         shape: &Shape,