Cuda acceleration for quantized model. (#1754)

* Boilerplate for the quantized cuda support.

* More basic cuda support.

* More cuda quantization (quantize on cpu for now).

* Add the dequantization bit.

* Start adding some dedicated cuda kernels from llama.cpp.

* Move the kernel code.

* Start interfacing with the kernel.

* Tweak the kernel launch params.

* Bugfix for quantized metal.

* Fix some clippy lints.

* Tweak the launch parameters.

* Tweak cuda basics to perform a quantized matmul.

* Perform the dequantization on the cpu + use cublas for matmul.

* Add the dequantization kernel.

* Test the qmatmul.

* More kernels.

* Matmul-vec kernel.

* Add a couple kernels.

* More dequantization kernels.

candle-core/src/quantized/dummy_cuda.rs

@@ -0,0 +1,50 @@
+#![allow(unused)]
+use super::GgmlDType;
+use crate::{CudaDevice, CudaStorage, Error, Result};
+
+pub struct QCudaStorage {
+    dtype: GgmlDType,
+    device: CudaDevice,
+}
+
+impl QCudaStorage {
+    pub fn zeros(_: &CudaDevice, _: usize, _: GgmlDType) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn dtype(&self) -> GgmlDType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &CudaDevice {
+        &self.device
+    }
+
+    pub fn dequantize(&self, _elem_count: usize) -> Result<CudaStorage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn quantize(&mut self, _src: &CudaStorage) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        0
+    }
+
+    pub fn fwd(
+        &self,
+        _self_shape: &crate::Shape,
+        _storage: &CudaStorage,
+        _layout: &crate::Layout,
+    ) -> Result<(CudaStorage, crate::Shape)> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+}
+
+pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
+    _device: &CudaDevice,
+    _data: &[T],
+) -> Result<super::QStorage> {
+    Err(Error::NotCompiledWithCudaSupport)
+}