Adding the actual backend

Cargo.toml

@@ -13,6 +13,7 @@ members = [
 exclude = [
    "candle-flash-attn",
    "candle-kernels",
+   "candle-metal-kernels",
    "candle-onnx",
 ]
 resolver = "2"
@@ -60,7 +61,7 @@ tracing-subscriber = "0.3.7"
 wav = "1.0.0"
 yoke = { version = "0.7.2", features = ["derive"] }
 zip = { version = "0.6.6", default-features = false }
-metal = { git = "https://github.com/ivarflakstad/metal-rs.git", features = ["mps"] }
+metal = { version = "0.27.1", features = ["mps"], package="candle-metal" }
 
 [profile.release-with-debug]
 inherits = "release"

candle-core/Cargo.toml

@@ -13,6 +13,7 @@ readme = "README.md"
 accelerate-src = { workspace = true, optional = true }
 byteorder = { workspace = true }
 candle-kernels = { path = "../candle-kernels", version = "0.3.1", optional = true }
+candle-metal-kernels = { path = "../candle-metal-kernels", version = "0.3.1", optional = true }
 metal = { workspace = true, optional = true}
 cudarc = { workspace = true, optional = true }
 gemm = { workspace = true }
@@ -40,4 +41,4 @@ cuda = ["cudarc", "dep:candle-kernels"]
 cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
-metal = ["dep:metal"]
+metal = ["dep:metal", "dep:candle-metal-kernels"]

candle-core/src/device.rs

@@ -8,7 +8,7 @@ use crate::{CpuStorage, DType, Result, Shape, Storage, WithDType};
 pub enum DeviceLocation {
     Cpu,
     Cuda { gpu_id: usize },
-    Metal,
+    Metal { gpu_id: usize },
 }
 
 #[derive(Debug, Clone)]
@@ -146,6 +146,7 @@ impl Device {
         match (self, rhs) {
             (Self::Cpu, Self::Cpu) => true,
             (Self::Cuda(lhs), Self::Cuda(rhs)) => lhs.same_device(rhs),
+            (Self::Metal(lhs), Self::Metal(rhs)) => lhs.same_device(rhs),
             _ => false,
         }
     }

candle-core/src/display.rs

@@ -14,7 +14,9 @@ impl Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
-            _ => todo!(),
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(f, "Tensor[")?;
@@ -477,7 +479,9 @@ impl std::fmt::Display for Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
-            crate::DeviceLocation::Metal => todo!(),
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(

candle-core/src/lib.rs

@@ -53,6 +53,8 @@ mod dummy_metal_backend;
 pub mod error;
 mod indexer;
 pub mod layout;
+#[cfg(feature = "metal")]
+pub mod metal_backend;
 #[cfg(feature = "mkl")]
 mod mkl;
 pub mod npy;