AMDGPU EP: register umbrella plugin EP and forward profile options - #2
Open
zz002 wants to merge 12 commits into
Open
AMDGPU EP: register umbrella plugin EP and forward profile options#2zz002 wants to merge 12 commits into
zz002 wants to merge 12 commits into
Conversation
Add the AMDGPU execution provider to ONNX Runtime GenAI. The AMDGPU EP resolves a profile to a concrete backend (MIGraphX or DML) at runtime; OGA drives it as a single DeviceType::AMDGPU. Provider naming: exposed as "amdgpu"; OGA also accepts "AMDGPUExecutionProvider" (the catalog form used by the AMD-shipped Windows ML EP MSIX) so test harnesses that match config strings against WinML-discovered names work without bypass hacks. Both normalize to "AMDGPU". GPU-resident KV cache: a GPU-resident DeviceInterface keeps the KV cache on the device (no per-token CPU-to-GPU roundtrip), with backend-agnostic opaque DeviceBuffer copies that dispatch to the active backend. Static-shape prefill: emit ep.migraphx.static_pad_* and hip_graph_enable session-config entries so the EP pads the prefill token axis and reuses a captured graph. DML ignores the migraphx-namespaced keys. Known limitations: - Beam search not supported (needs past_present_share_buffer=true, which requires num_beams=1)
Route the small decode inputs (input_ids/position_ids/attention_mask) through a host-accessible (CPU-writable, GPU-readable) allocator so the CPU updates them in place with no per-step copy. Resolved via GetSharedAllocator; KV cache and scoring stay on the default device interface, and the path falls back to default inputs if no host-accessible allocator is available. Single-GPU only for now (device_id 0).
Logits is GPU-written and CPU-read (the sampler), the opposite of the pinned decode inputs. On AMDGPU the inputs use a host-accessible allocator whose heap is not CPU-read-coherent, so reading logits from it returns stale data. Route logits to the CPU interface instead via a new p_logits_ member; only the decode inputs stay on the host-accessible allocator.
…ice (review microsoft#2165) Resolve the AMDGPU device id from the filtered EP device instead of hardcoding 0. The EP keys its allocator on this id, so a hardcoded 0 pinned compute to device 0 regardless of the selected device. Correlate the host-accessible pool to the same id so pinned decode inputs live on the device that runs the model. Single-GPU resolves to id 0, unchanged.
The umbrella EP routes backends by model architecture but OGA never sent it, so every model routed as non-LLM. Emit config.model.type as the ep.amdgpuexecutionprovider.model_arch provider option alongside the existing static-padding hints.
Trailing-comment spacing and argument-continuation alignment flagged by the lint-cpp CI check (clang-format 20.1.0). Formatting only, no behavior change.
zz002
force-pushed
the
oga-amdgpu-hip-ep-registration
branch
from
July 23, 2026 07:05
70afe53 to
24ad31f
Compare
Emit the ep.directml.enable_host_accessible provider option so the DirectML backend uses host-accessible decode inputs. Sits alongside the existing static-padding and model_arch config entries.
zz002
force-pushed
the
oga-amdgpu-hip-ep-registration
branch
from
July 27, 2026 02:00
24ad31f to
de1f8c9
Compare
The AMDGPU umbrella is a plugin EP, so its OrtEpDevice is only visible after RegisterExecutionProviderLibrary. The C model_benchmark has no --ep_library flag, so self-register amdgpu-ep.dll on the OrtEnv the way the RyzenAI interface does, resolving the DLL next to the genai/ort module or the executable. The umbrella reads its provider options (e.g. profile) from session config entries prefixed "ep.amdgpuexecutionprovider.", not the AppendExecutionProvider_V2 ep_options channel, so bridge the genai_config provider options into that prefixed form. This is what makes profile=hip actually reach the umbrella and select the hipgpu backend. Override ShapeInitSessionProviderOptions so the trivial device-init session that EnsureDeviceOrtInit builds also carries the profile; otherwise the umbrella has no backend to create for that init session and device/allocator setup fails. Co-authored-by: Cursor <cursoragent@cursor.com>
aditya-dl
force-pushed
the
amd/dev/adilohia/migraphx_support
branch
from
August 7, 2026 00:13
1e76280 to
05297cc
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Makes the AMDGPU umbrella EP work end-to-end through OGA's own
model_benchmarkwithprofile=hip. Three OGA-side fixes insrc/amdgpu; no umbrella EP change required.Why
Testing with OGA's bundled C
model_benchmark(which has no--ep_library/-p/--profileflag; profile is passed viagenai_config.jsonprovider options) surfaced three gaps:OrtEpDeviceis only discoverable afterRegisterExecutionProviderLibrary. Legacy in-proc EPs skip this, but the umbrella does not, soAppendExecutionProvider_V2finds no device and allocation fails ("no requested allocator available"). The RyzenAI interface already self-registers its DLL; the AMDGPU path did not.profile) from session-config entries prefixedep.amdgpuexecutionprovider., not theAppendExecutionProvider_V2ep_optionschannel, soprofile=hipnever selected the hipgpu backend.EnsureDeviceOrtInitbuilds a trivial device-init session to set up the device allocator. After the Route init-session provider-option shaping through DeviceInterface microsoft/onnxruntime-genai#2232 refactor, its provider options come solely from the per-EPShapeInitSessionProviderOptionshook (default no-op), so the umbrella had no backend to create for that session.What
EnsureUmbrellaEpRegistered()insrc/amdgpu/session_options.cpp— self-registersamdgpu-ep.dllon theOrtEnv, mirroring the RyzenAI pattern (resolve the DLL next to the genai/ort module or the executable; re-registration is benign).ForwardUmbrellaProviderOptions()insrc/amdgpu/session_options.cpp— bridges the genai_config provider options intoep.amdgpuexecutionprovider.*config entries so the umbrella selects the requested backend.ShapeInitSessionProviderOptions()override insrc/amdgpu/interface.cpp— forwards the user's provider options to the trivial init session so it carries the profile too.Test plan
model_benchmark.exe+onnxruntime-genai.dlllocally with the umbrellaamdgpu-ep.dll+hip-backend.dll.profile=hipruns end-to-end on Llama-3.1-8B (int4-awq) viamodel_benchmark.exeon gfx1151 (StxHalo); prefill + decode complete, output correct.Notes for reviewers