Fix CUDA build and CI issues

- cudarc 0.16: device_ptr() now requires &CudaStream, store stream in DeviceBuffer
- Manual Debug impl for GpuTensor (DeviceBuffer doesn't derive Debug)
- Use macos-latest for x86_64 Darwin cross-compile (macos-13 deprecated)
- Add rust-cache to CI workflow
- Add sccache to release builds

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: lightsofapollo

.github/workflows/ci.yml

@@ -19,6 +19,9 @@ jobs:
       - uses: dtolnay/rust-toolchain@stable
         with:
           components: clippy, rustfmt
+      - uses: Swatinem/rust-cache@v2
+        with:
+          workspaces: crates
       - name: Check formatting
         run: cargo fmt --workspace --check
         working-directory: crates

.github/workflows/release.yml

@@ -35,7 +35,7 @@ jobs:
             target: x86_64-unknown-linux-gnu
           - os: ubuntu-24.04-arm
             target: aarch64-unknown-linux-gnu
-          - os: macos-13
+          - os: macos-latest
             target: x86_64-apple-darwin
           - os: macos-latest
             target: aarch64-apple-darwin
@@ -47,6 +47,7 @@ jobs:
         with:
           target: ${{ matrix.target }}
           args: --release --out dist
+          sccache: true
           manylinux: auto
           before-script-linux: |
             yum install -y libcurl-devel openssl-devel || (apt-get update && apt-get install -y libcurl4-openssl-dev libssl-dev pkg-config)

crates/sst-gpu/src/copy.rs

@@ -15,12 +15,13 @@ use crate::GpuError;
 /// (e.g. to hand to PyTorch via `from_blob`).
 pub struct DeviceBuffer {
     inner: CudaSlice<u8>,
+    stream: Arc<CudaStream>,
 }
 
 impl DeviceBuffer {
     /// Raw device pointer as `u64`, suitable for FFI / PyTorch interop.
     pub fn device_ptr(&self) -> u64 {
-        let (ptr, _sync) = self.inner.device_ptr();
+        let (ptr, _sync) = self.inner.device_ptr(&self.stream);
         ptr as u64
     }
 
@@ -93,7 +94,7 @@ impl GpuCopier {
                     reason: e.to_string(),
                 }
             })?;
-            return Ok(DeviceBuffer { inner: empty });
+            return Ok(DeviceBuffer { inner: empty, stream: Arc::clone(&self.stream) });
         }
 
         // Build a host slice from the raw pointer (no ownership taken).
@@ -111,7 +112,7 @@ impl GpuCopier {
             })?;
 
         tracing::trace!(bytes = len, "queued H2D copy");
-        Ok(DeviceBuffer { inner: device_buf })
+        Ok(DeviceBuffer { inner: device_buf, stream: Arc::clone(&self.stream) })
     }
 
     /// Block until all work on this copier's stream has completed.

crates/sst-gpu/src/pipeline.rs

@@ -15,7 +15,6 @@ use crate::pinned::PinnedBuffer;
 use crate::GpuError;
 
 /// A tensor whose data now lives in GPU device memory.
-#[derive(Debug)]
 pub struct GpuTensor {
     /// Tensor name from the safetensors file.
     pub name: String,
@@ -31,6 +30,18 @@ pub struct GpuTensor {
     _device_buf: crate::copy::DeviceBuffer,
 }
 
+impl std::fmt::Debug for GpuTensor {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("GpuTensor")
+            .field("name", &self.name)
+            .field("device_ptr", &self.device_ptr)
+            .field("len", &self.len)
+            .field("dtype", &self.dtype)
+            .field("shape", &self.shape)
+            .finish()
+    }
+}
+
 /// Orchestrates the full streaming pipeline:
 ///
 /// `Consumer` -> stage in pinned host memory -> async H2D copy -> [`GpuTensor`]