Add AMDGPU execution provider (port of #2165 to rel-0.15.1) - #1
Merged
Zhaeong merged 11 commits intoAug 4, 2026
Merged
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.
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.
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.
Ports the AMDGPU execution provider work from PR microsoft#2165 onto rel-0.15.1.
Adds the AMDGPU umbrella EP with a GPU-resident KV cache and host-accessible decode inputs. Same 11 commits as microsoft#2165, adapted to the 0.15.1 DeviceInterface::GetMemoryInfo() refactor.