Skip to content

[AIMIGRAPHX-] MIGraphX ROCm 7.14 feature port.#53

Draft
TedThemistokleous wants to merge 2 commits into
mainfrom
migx_ep_714_feature_port
Draft

[AIMIGRAPHX-] MIGraphX ROCm 7.14 feature port.#53
TedThemistokleous wants to merge 2 commits into
mainfrom
migx_ep_714_feature_port

Conversation

@TedThemistokleous

@TedThemistokleous TedThemistokleous commented Jul 21, 2026

Copy link
Copy Markdown
Collaborator

Porting outstanding items and features from the changes in ROCm 7.14_internal_testing from rocm/Onnxruntime

  • Remove reundant syncs in compute and fix stream syncs
  • skip input hashing for identical inputs - helps larger models with many inputs

The MIGraphX EP synchronized the compute stream after every Compute call
(staging, eager, and EPContext paths), forcing a full CPU/GPU serialization
per node and destroying inference overlap - worst for graphs partitioned
into multiple MIGraphX/other nodes.

ORT owns the per-run compute stream (created in CreateSyncStreamForDevice,
handed to the kernel via GetGPUComputeStream()) and already flushes it at
run end through DeviceStreamCollection::CleanUp -> SyncStream::Flush,
honoring the run's sync_stream setting. Intra-run cross-stream ordering is
handled by ORT notifications. The EP therefore does not need to sync in the
hot path.

Also fix OnRunEnd, which synchronized the EP's default stream (stream_) -
a stream that never carried compute work, since all kernels run on ORT's
compute stream. That sync was both incorrect (wrong stream) and a redundant
serialization point, so it is dropped and the now-unused stream_ member is
removed.

Outputs remain correct: all H2D/compute/D2D work is enqueued on the compute
stream in order and ORT flushes it before returning results to the caller.
@TedThemistokleous
TedThemistokleous marked this pull request as draft July 21, 2026 21:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants