Cudarc static-linking enabled.

Cargo.toml

@@ -43,7 +43,7 @@ candle-onnx = { path = "./candle-onnx", version = "0.8.4" }
 candle-transformers = { path = "./candle-transformers", version = "0.8.4" }
 clap = { version = "4.2.4", features = ["derive"] }
 criterion = { version = "0.5.1", default-features=false }
-cudarc = { version = "0.13.5", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
+cudarc = { version = "0.13.5", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16"], default-features=false }
 fancy-regex = "0.13.0"
 gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
 hf-hub = "0.4.1"

candle-core/Cargo.toml

@@ -43,8 +43,9 @@ criterion = { workspace = true }
 
 [features]
 default = []
-cuda = ["cudarc", "dep:candle-kernels", "dep:ug-cuda"]
+_cuda = ["dep:cudarc", "dep:candle-kernels", "dep:ug-cuda"]
-cudnn = ["cuda", "cudarc/cudnn"]
+# cuda = ["_cuda", "cudarc?/cuda-version-from-build-system", "cudarc?/dynamic-linking"]
+cudnn = ["_cuda", "cudarc?/cudnn"]
 _mkl = ["dep:libc", "dep:intel-mkl-src"]
 mkl = ["_mkl", "intel-mkl-src?/mkl-static-lp64-iomp"] 
 accelerate = ["dep:libc", "dep:accelerate-src"]

candle-core/benches/benchmarks/mod.rs

@@ -20,9 +20,9 @@ impl BenchDevice for Device {
         match self {
             Device::Cpu => Ok(()),
             Device::Cuda(device) => {
-                #[cfg(feature = "cuda")]
+                #[cfg(feature = "_cuda")]
                 return Ok(device.synchronize()?);
-                #[cfg(not(feature = "cuda"))]
+                #[cfg(not(feature = "_cuda"))]
                 panic!("Cuda device without cuda feature enabled: {:?}", device)
             }
             Device::Metal(device) => {
@@ -61,7 +61,7 @@ impl BenchDeviceHandler {
         let mut devices = Vec::new();
         if cfg!(feature = "metal") {
             devices.push(Device::new_metal(0)?);
-        } else if cfg!(feature = "cuda") {
+        } else if cfg!(feature = "_cuda") {
             devices.push(Device::new_cuda(0)?);
         }
         devices.push(Device::Cpu);

candle-core/src/custom_op.rs

@@ -378,7 +378,7 @@ impl Tensor {
 
 pub struct UgIOp1 {
     name: &'static str,
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     func: cudarc::driver::CudaFunction,
     #[cfg(feature = "metal")]
     func: metal::ComputePipelineState,
@@ -392,7 +392,7 @@ impl UgIOp1 {
         kernel: ug::lang::ssa::Kernel,
         device: &crate::Device,
     ) -> Result<Self> {
-        #[cfg(feature = "cuda")]
+        #[cfg(feature = "_cuda")]
         {
             let device = device.as_cuda_device()?;
             let func = device.compile(name, kernel)?;
@@ -404,7 +404,7 @@ impl UgIOp1 {
             let func = device.compile(name, kernel)?;
             Ok(Self { name, func })
         }
-        #[cfg(not(any(feature = "cuda", feature = "metal")))]
+        #[cfg(not(any(feature = "_cuda", feature = "metal")))]
         {
             Ok(Self { name })
         }
@@ -456,7 +456,7 @@ impl InplaceOp1 for UgIOp1 {
         Ok(())
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(&self, sto: &mut CudaStorage, layout: &Layout) -> Result<()> {
         use crate::cuda_backend::WrapErr;
         use cudarc::driver::LaunchAsync;

candle-core/src/lib.rs

@@ -55,7 +55,7 @@ pub mod conv;
 mod convert;
 pub mod cpu;
 pub mod cpu_backend;
-#[cfg(feature = "cuda")]
+#[cfg(feature = "_cuda")]
 pub mod cuda_backend;
 mod custom_op;
 mod device;
@@ -104,10 +104,10 @@ pub use strided_index::{StridedBlocks, StridedIndex};
 pub use tensor::{Tensor, TensorId};
 pub use variable::Var;
 
-#[cfg(feature = "cuda")]
+#[cfg(feature = "_cuda")]
 pub use cuda_backend as cuda;
 
-#[cfg(not(feature = "cuda"))]
+#[cfg(not(feature = "_cuda"))]
 pub use dummy_cuda_backend as cuda;
 
 pub use cuda::{CudaDevice, CudaStorage};

candle-core/src/quantized/mod.rs

@@ -16,9 +16,9 @@ pub mod metal;
 mod metal {
     pub use super::dummy_metal::*;
 }
-#[cfg(feature = "cuda")]
+#[cfg(feature = "_cuda")]
 pub mod cuda;
-#[cfg(not(feature = "cuda"))]
+#[cfg(not(feature = "_cuda"))]
 mod cuda {
     pub use super::dummy_cuda::*;
 }

candle-core/src/sort.rs

@@ -52,7 +52,7 @@ impl ArgSort {
     }
 }
 
-#[cfg(feature = "cuda")]
+#[cfg(feature = "_cuda")]
 mod cuda {
     use super::*;
     use crate::cuda_backend::cudarc::driver::{
@@ -118,7 +118,7 @@ impl crate::CustomOp1 for ArgSort {
         Ok((sort_indexes, layout.shape().into()))
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         storage: &crate::CudaStorage,

candle-core/src/test_utils.rs

@@ -10,7 +10,7 @@ macro_rules! test_device {
             $fn_name(&Device::Cpu)
         }
 
-        #[cfg(feature = "cuda")]
+        #[cfg(feature = "_cuda")]
         #[test]
         fn $test_cuda() -> Result<()> {
             $fn_name(&Device::new_cuda(0)?)

candle-core/src/utils.rs

@@ -21,7 +21,7 @@ pub fn has_mkl() -> bool {
 }
 
 pub fn cuda_is_available() -> bool {
-    cfg!(feature = "cuda")
+    cfg!(feature = "_cuda")
 }
 
 pub fn metal_is_available() -> bool {

candle-core/tests/custom_op_tests.rs

@@ -144,7 +144,7 @@ fn inplace_op1() -> Result<()> {
     Ok(())
 }
 
-#[cfg(any(feature = "cuda", feature = "metal"))]
+#[cfg(any(feature = "_cuda", feature = "metal"))]
 #[allow(clippy::approx_constant)]
 #[test]
 fn ug_op() -> Result<()> {

candle-flash-attn/Cargo.toml

@@ -11,7 +11,7 @@ license = "MIT OR Apache-2.0"
 readme = "README.md"
 
 [dependencies]
-candle = { path = "../candle-core", features = ["cuda"], package = "candle-core", version = "0.8.4" }
+candle = { path = "../candle-core", features = ["_cuda"], package = "candle-core", version = "0.8.4" }
 half = { version = "2.3.1", features = ["num-traits"] }
 
 [build-dependencies]
@@ -21,4 +21,4 @@ anyhow = { version = "1", features = ["backtrace"] }
 
 [dev-dependencies]
 anyhow = { version = "1", features = ["backtrace"] }
-candle-nn = { path = "../candle-nn", features = ["cuda"] }
+candle-nn = { path = "../candle-nn", features = ["_cuda"] }

candle-nn/Cargo.toml

@@ -32,6 +32,7 @@ criterion = { workspace = true }
 [features]
 default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate"]
+_cuda = ["candle/_cuda"]
 cuda = ["candle/cuda"]
 _mkl = ["dep:intel-mkl-src", "candle/_mkl"]
 mkl = ["candle/mkl"]

candle-nn/benches/benchmarks/mod.rs

@@ -15,9 +15,9 @@ impl BenchDevice for Device {
         match self {
             Device::Cpu => Ok(()),
             Device::Cuda(device) => {
-                #[cfg(feature = "cuda")]
+                #[cfg(feature = "_cuda")]
                 return Ok(device.synchronize()?);
-                #[cfg(not(feature = "cuda"))]
+                #[cfg(not(feature = "_cuda"))]
                 panic!("Cuda device without cuda feature enabled: {:?}", device)
             }
             Device::Metal(device) => {
@@ -56,7 +56,7 @@ impl BenchDeviceHandler {
         let mut devices = Vec::new();
         if cfg!(feature = "metal") {
             devices.push(Device::new_metal(0)?);
-        } else if cfg!(feature = "cuda") {
+        } else if cfg!(feature = "_cuda") {
             devices.push(Device::new_cuda(0)?);
         }
         devices.push(Device::Cpu);

candle-nn/src/ops.rs

@@ -82,7 +82,7 @@ impl candle::CustomOp1 for Sigmoid {
         Ok((storage, layout.shape().clone()))
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         storage: &candle::CudaStorage,
@@ -333,7 +333,7 @@ impl candle::CustomOp1 for SoftmaxLastDim {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         storage: &candle::CudaStorage,
@@ -507,7 +507,7 @@ impl candle::CustomOp2 for RmsNorm {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         s1: &candle::CudaStorage,
@@ -740,7 +740,7 @@ impl candle::CustomOp3 for LayerNorm {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         s1: &candle::CudaStorage,

candle-nn/src/rotary_emb.rs

@@ -77,7 +77,7 @@ impl candle::CustomOp3 for RotaryEmbI {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         s1: &candle::CudaStorage,
@@ -322,7 +322,7 @@ impl candle::CustomOp3 for RotaryEmb {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         s1: &candle::CudaStorage,
@@ -576,7 +576,7 @@ impl candle::CustomOp3 for RotaryEmbThd {
         }
     }
 
-    #[cfg(feature = "cuda")]
+    #[cfg(feature = "_cuda")]
     fn cuda_fwd(
         &self,
         s1: &candle::CudaStorage,

candle-transformers/Cargo.toml

@@ -28,6 +28,7 @@ tracing = { workspace = true }
 [features]
 default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate"]
+_cuda = ["candle/_cuda", "candle-nn/_cuda"]
 cuda = ["candle/cuda", "candle-nn/cuda"]
 flash-attn = ["cuda", "dep:candle-flash-attn"]
 mkl = ["dep:intel-mkl-src", "candle/mkl", "candle-nn/mkl"]