docs: document graph runtime API

docs/api/runtime.md

@@ -50,6 +50,9 @@ otherwise.
 runtime::Error
 runtime::Stream
 runtime::Event
+runtime::Graph
+runtime::GraphExec
+runtime::StreamCaptureMode
 runtime::MemcpyKind
 runtime::kSuccess
 runtime::kMemcpyHostToHost
@@ -58,6 +61,15 @@ runtime::kMemcpyDeviceToHost
 runtime::kMemcpyDeviceToDevice
 ```
 
+`StreamCaptureMode` supports the same capture scopes used by CUDA-style stream
+capture:
+
+```cpp
+runtime::StreamCaptureMode::kStreamCaptureModeGlobal
+runtime::StreamCaptureMode::kStreamCaptureModeThreadLocal
+runtime::StreamCaptureMode::kStreamCaptureModeRelaxed
+```
+
 ## Device Functions
 
 ```cpp
@@ -107,6 +119,58 @@ runtime::Error EventElapsedTime(float* ms, runtime::Event start,
                                 runtime::Event end);
 ```
 
+## Graph Capture and Replay Functions
+
+When the configured build includes a graph-capable backend, the generated
+runtime API also includes graph capture and replay entry points.
+
+The graph API captures stream work into a reusable `Graph`, instantiates it as a
+launchable `GraphExec`, and replays it on a stream:
+
+```cpp
+runtime::Error StreamBeginCapture(runtime::Stream stream,
+                                  runtime::StreamCaptureMode mode);
+runtime::Error StreamEndCapture(runtime::Stream stream, runtime::Graph* graph);
+runtime::Error GraphDestroy(runtime::Graph graph);
+runtime::Error GraphInstantiate(runtime::GraphExec* graph_exec,
+                                runtime::Graph graph);
+runtime::Error GraphExecDestroy(runtime::GraphExec graph_exec);
+runtime::Error GraphLaunch(runtime::GraphExec graph_exec,
+                           runtime::Stream stream);
+```
+
+Use the same stream for `StreamBeginCapture` and `StreamEndCapture`. Operations
+issued to that stream between the two calls become part of the captured graph.
+After `GraphInstantiate` succeeds, the executable graph can be launched multiple
+times while inputs and outputs remain at the recorded addresses.
+
+```cpp
+namespace rt = infini::rt::runtime;
+
+rt::Stream stream = nullptr;
+rt::Graph graph = nullptr;
+rt::GraphExec graph_exec = nullptr;
+
+rt::StreamCreate(&stream);
+
+rt::StreamBeginCapture(
+    stream, rt::StreamCaptureMode::kStreamCaptureModeRelaxed);
+rt::MemcpyAsync(dst, src, count, rt::kMemcpyDeviceToDevice, stream);
+rt::StreamEndCapture(stream, &graph);
+
+rt::GraphInstantiate(&graph_exec, graph);
+rt::GraphLaunch(graph_exec, stream);
+rt::StreamSynchronize(stream);
+
+rt::GraphExecDestroy(graph_exec);
+rt::GraphDestroy(graph);
+rt::StreamDestroy(stream);
+```
+
+Backends that do not support graph capture return a non-success status for the
+graph entry points when those functions are present in the configured public
+header. See [Backends](../backends.md) for current support.
+
 ## Backend-Specific Runtime Templates
 
 Backend runtime specializations are available as implementation-facing headers,

docs/backends.md

@@ -28,16 +28,16 @@ status on backends that do not support them.
 
 Current test expectations are:
 
-| Backend | Async memcpy | Host memory | Async memory | Memory info | Async memset | Events |
-| --- | --- | --- | --- | --- | --- | --- |
-| CPU | No | Yes | No | Yes | No | Yes |
-| NVIDIA | Yes | Yes | Yes | Yes | Yes | Yes |
-| Iluvatar | Yes | Yes | Yes | Yes | Yes | Yes |
-| MetaX | Yes | Yes | Yes | Yes | Yes | Yes |
-| Moore | Yes | Yes | No | Yes | Yes | Yes |
-| Hygon | Yes | Yes | Yes | Yes | Yes | Yes |
-| Cambricon | Yes | No | No | No | No | No |
-| Ascend | Yes | No | No | No | No | No |
+| Backend | Async memcpy | Host memory | Async memory | Memory info | Async memset | Events | Graph capture |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| CPU | No | Yes | No | Yes | No | Yes | No |
+| NVIDIA | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
+| Iluvatar | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
+| MetaX | Yes | Yes | Yes | Yes | Yes | Yes | No |
+| Moore | Yes | Yes | No | Yes | Yes | Yes | No |
+| Hygon | Yes | Yes | Yes | Yes | Yes | Yes | No |
+| Cambricon | Yes | No | No | No | No | No | No |
+| Ascend | Yes | No | No | No | No | No | Yes |
 
 Treat a non-`kSuccess` status as the portable way to detect unsupported
 operations.

docs/compatibility.md

@@ -18,7 +18,7 @@ The stable user-facing surface is:
 - `infini::rt::TensorView`
 - `infini::rt::set_runtime_device_type`
 - `infini::rt::runtime_device_type`
-- `infini::rt::runtime` dispatching functions and constants documented in
+- `infini::rt::runtime` types, dispatching functions, and constants documented in
   [Runtime API](api/runtime.md)
 
 ## Implementation-Facing Headers