[naft-san (Nominy)]

Description

With enableGraphCapture, run() binds inputs and outputs only on the first run of a session, but the WebGPU EP unregisters the GPU buffers of external tensors at the end of every run. Any later run that executes kernels (an uncaptured gpu_graph_id: '-1' run, or capturing another graph id) then binds released buffers and fails with Required member is undefined in createBindGroup.

This keeps the binding under graph capture, since ORT-allocated outputs must stay bound for replay, but re-binds inputs and pre-allocated outputs on every run. Because the binding is kept, a graph-capture session must request the same outputs, each pre-allocated or not, on every run. A run that changes this used to return stale or empty outputs; it now throws before binding anything, and the session keeps working.

Adds a unittest for the WebGPU EP build (JSEP rejects re-registering a different external buffer under graph capture). CI does not run unittests for that build; run it with npm test --webgpu-ep -- unittest -b=webgpu. It fails without the fix.

ORT-allocated outputs whose shape changes between graph ids are still unsupported: they returned a stale output before, and now raise an error.

Motivation and Context

Running a captured decode step after an uncaptured prefill (gpu_graph_id: '-1', then '1') fails on the second run.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

Clearing outputs breaks graph replay with GPU-buffer inputs and no preallocated outputs, causing subsequent runs to fail in _OrtGetTensorData rather than return captured outputs.

1 open issue

• Preserve allocated outputs when rebinding captured graphs · New

What changed in this PR

This PR updates ONNX Runtime Web’s WebGPU binding lifecycle to address stale GPU-buffer references across runs when graph capture is enabled.

Changes:

• Clears bound outputs and resets the IO-bound state after successful runs, including graph-capture runs, so subsequent runs rebind their tensors.
• Adds a comment explaining why bindings must not outlive per-run GPU-buffer registrations.

File	Description
js/​web/​lib/​wasm/​wasm-core-impl.ts	Removes the graph-capture exception from binding cleanup and documents the buffer-lifetime rationale.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Output-binding state persists incorrectly across runs, causing graph replay to retrieve unregistered output handles and changed output requests to fail with output-count mismatches.

4 open issues

• Remove stale output bindings between uncaptured runs · New
• Restore per-graph output bindings when switching allocation modes · New
• Skip IO binding tests when WASM proxy mode is enabled · New
• Release returned tensors to avoid leaking native handles · New

1 resolved since last review

• Preserve allocated outputs when rebinding captured graphs

_{Review effort: Max}

[Copilot]

🟢 Approval recommended

No active review issues block the graph-capture I/O binding fix.

No open issues 🎉

4 resolved since last review

• Restore per-graph output bindings when switching allocation modes
• Remove stale output bindings between uncaptured runs
• Release returned tensors to avoid leaking native handles
• Skip IO binding tests when WASM proxy mode is enabled

_{Review effort: Balanced}