Struct GraphCaptureGuard 

pub struct GraphCaptureGuard { /* private fields */ }

RAII guard that runs CUDA graph capture in a scoped block.

Call GraphCaptureGuard::begin (or [begin_with_mode] / [begin_with_pool]) to start capture; calling [finish] returns the instantiated graph. If the guard is dropped without calling finish (for example because a kernel returned an error mid-capture), its Drop impl best-effort-ends capture and discards the resulting graph so the stream returns to a usable state. CL-454.

This mirrors PyTorch’s with torch.cuda.graph(g): ... context manager semantics in Rust’s RAII idiom.

Example

use ferrotorch_gpu::graph::GraphCaptureGuard;

let mut guard = GraphCaptureGuard::begin(device.stream())?;
run_kernels()?; // any kernel launched on device.stream() is recorded
let graph = guard.finish()?;
graph.upload()?;
for _ in 0..1000 { graph.launch()?; }

Implementations

impl GraphCaptureGuard

pub fn begin(stream: &Arc<CudaStream>) -> GpuResult<Self>

Begin graph capture on stream in the default CaptureMode::ThreadLocal mode. CL-454.

pub fn begin_with_mode( stream: &Arc<CudaStream>, mode: CaptureMode, ) -> GpuResult<Self>

Begin graph capture with an explicit CaptureMode. CL-454.

pub fn begin_with_pool( stream: &Arc<CudaStream>, pool: Arc<CapturePool>, ) -> GpuResult<Self>

Begin graph capture bound to a CapturePool. The pool is attached to the resulting graph by [finish]. CL-454.

pub fn finish(self) -> GpuResult<CapturedGraph>

Finish capture and return the instantiated CapturedGraph.

Consumes the guard so Drop becomes a no-op. If a pool was attached at construction, the resulting graph is produced via end_capture_with_pool and holds the pool Arc for the lifetime of the graph.

pub fn status(&self) -> GpuResult<CaptureStatus>

Report whether the stream this guard is bound to is still actively capturing. An unexpected Invalidated or None usually means a forbidden API call (alloc, sync, host copy) happened under capture.

Trait Implementations

impl Drop for GraphCaptureGuard

Available on crate feature cuda only.

fn drop(&mut self)

Executes the destructor for this type.