Vulkan support via Blade

Cargo.toml

@@ -47,3 +47,6 @@ exclude = [
     "crates/luminal_2",
     "demos/flash_attention",
 ]
+
+#[patch."https://github.com/kvark/blade"]
+#blade-graphics = { path = "../blade/blade-graphics" }

crates/luminal_2/Cargo.toml

@@ -7,6 +7,7 @@ edition = "2024"
 default = []
 cuda = ["dep:cudarc", "dep:luminal_cuda"]
 metal = ["dep:objc2", "dep:objc2-metal", "dep:objc2-foundation"]
+blade = ["dep:blade-graphics"]
 
 [dependencies]
 luminal = { path = "../../" }
@@ -42,3 +43,4 @@ objc2-foundation = { version = "0.3.1", optional = true }
 eframe = "0.28"
 egui = "0.28"
 anyhow = "1.0.99"
+blade-graphics = { git = "https://github.com/kvark/blade", rev = "99162daf690e05d4489b57adeb37a2cc0086d1be", optional = true }

crates/luminal_2/src/lib.rs

@@ -8,6 +8,9 @@ pub mod utils;
 #[cfg(test)]
 mod tests;
 
+#[cfg(feature = "blade")]
+use blade_graphics as gpu;
+
 #[cfg(feature = "cuda")]
 use itertools::Itertools;
 use luminal::prelude::*;
@@ -30,10 +33,17 @@ pub type Buffer = Retained<ProtocolObject<dyn MTLBuffer>>;
 #[cfg(feature = "metal")]
 pub type Function = Retained<ProtocolObject<dyn MTLFunction>>;
 
+#[cfg(feature = "cuda")]
+pub type Device = CudaContext;
+
+#[cfg(feature = "blade")]
+pub type Device = gpu::Context;
+
 #[derive(Clone, PartialEq, Eq)]
 pub enum GPUArch {
     CUDA,
     Metal(HashMap<usize, &'static str>),
+    Blade,
 }
 
 impl GPUArch {
@@ -179,6 +189,12 @@ impl Operator for CompatKernel {
         todo!()
     }
 }
+#[cfg(feature = "blade")]
+impl Operator for CompatKernel {
+    fn process(&mut self, _inputs: Vec<(InputTensor, ShapeTracker)>) -> Vec<Tensor> {
+        todo!()
+    }
+}
 
 pub fn custom_kernel(
     inputs: &[GraphTensor],
@@ -226,6 +242,12 @@ impl Operator for Diff {
         todo!()
     }
 }
+#[cfg(feature = "blade")]
+impl Operator for Diff {
+    fn process(&mut self, _inp: Vec<(InputTensor, ShapeTracker)>) -> Vec<Tensor> {
+        todo!()
+    }
+}
 
 pub trait GT2 {
     fn diff2(self, name: impl ToString) -> Self;
@@ -268,3 +290,9 @@ impl Operator for GraphBreak {
         inp.into_iter().map(|i| i.0.cloned()).collect()
     }
 }
+
+#[cfg(feature = "blade")]
+pub struct Buffer {
+    pub raw: gpu::Buffer,
+    pub size: usize,
+}