WebGPU EP crashes on macOS when running concurrent multi threaded inference

[altunenes]

Describe the issue

Note: Re opening this issue as a follow up to #27592, which was automatically closed by the stale bot despite being an active (and I can't click to re open), reproducible bug. Confirmed still present on latest ORT (v1.28) so wanted to create a new anyway.

The WebGPU Execution Provider crashes on macOS (Apple Silicon) when multiple threads create sessions and run inference concurrently. The crash is non-deterministic and manifests as different errors on each run, indicating a race condition in Dawn's Metal backend.

Observed crashes (non-deterministic, varies per run)
Run 1 — Metal assertion:

  **-[AGXG15GFamilyCommandBuffer blitCommandEncoderCommon:]:832:**
    failed assertion `A command encoder is already encoding to this command buffer'                

Run 2 — Memory corruption:
malloc: *** error for object 0x784797e90: pointer being freed was not allocated

Run 3 — Trace trap:
[1] 56040 trace trap cargo run --release -- version-RFB-640.onnx

Notes:

• Single-threaded session creation and inference works fine - Sequential (non-concurrent) multi-session usage works fine
• The same workload works correctly with the CoreML EP

To reproduce

minimal Repo: https://github.com/altunenes/ort-webgpu-thread-crash

git clone https://github.com/altunenes/ort-webgpu-thread-crash.git
cargo run --release -- version-RFB-640.onnx
(simple face detection model)

The repro creates 2 persistent WebGPU sessions, then spawns 5 threads that each create a session and run inference concurrently while the main thread also runs inference on the persistent sessions.

Urgency

No response

Platform

Mac

OS Version

26.2

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.28

ONNX Runtime API

Other / Unknown

Architecture

X64

Execution Provider

Other / Unknown

Execution Provider Library Version

No response