Module graph 

CUDA Graphs — record a sequence of operations once, replay cheaply.

Two construction paths:

1. Stream capture (Stream::begin_capture, Stream::end_capture) — run your work normally on a stream; the driver records it into a graph instead of executing it. This is the recommended entry point for most users.
2. Explicit construction (Graph::new) — build a graph node by node. baracuda v0.1 does not yet expose typed node builders; use capture for now.

Either way, instantiate the Graph to a GraphExec and launch it on any stream as many times as you like.

Modules

instantiate_flags

Re-export of cuGraphInstantiateWithFlags flag constants.

Structs

Graph

A CUDA graph — a DAG of CUDA operations.

GraphExec

An instantiated (executable) CUDA graph.

GraphNode

A node inside a Graph. Lightweight handle that borrows the parent graph’s storage — destroying the graph invalidates all of its nodes.

UpdateResult

Outcome of GraphExec::update. Inspect result against the constants in baracuda_cuda_sys::types::CUgraphExecUpdateResult — SUCCESS (0) means the executable graph was patched in place.

Enums

CaptureMode

Stream-capture mode, matching CUstreamCaptureMode.

Functions

device_graph_mem_attribute

Query a CUgraphMem_attribute (used / reserved memory, high-water marks) — see baracuda_cuda_sys::types::CUgraphMem_attribute.

device_graph_mem_trim

Trim the per-device graph-mem reserve back to the minimum footprint.

device_set_graph_mem_attribute

Reset a CUgraphMem_attribute (typically the high-water marks).