feat(amdgpu): add hipep backend selected via profile=llm#1
Open
zz002 wants to merge 12 commits into
Open
Conversation
zz002
force-pushed
the
feat/rocm-llm-backend
branch
from
9644fc4 to
ff40ecc
Compare
onnxruntime#21) Add a declarative pattern-matching fusion engine to the DirectML plugin EP that detects raw ONNX operator sequences and fuses them into optimized DML kernels without depending on ORT internal C++ classes. Key changes: - pattern_matcher: PNode tree DSL with greedy longest-match semantics and anchor-indexed O(nodes + matches) traversal - ep_fusion_manager: anchor-indexed rule registry; FusionMatchMap keyed by node ID hash eliminates re-detection at CompileImpl time - quick_gelu_ep_fusion: detects Sigmoid+Mul (2-node) or Mul+Sigmoid+Mul (3-node); compiles via IDMLDevice1::CompileGraph with lazy init - bias_gelu_ep_fusion: detects raw 6-node Gelu expansion Add+Div+Erf+Add+Mul+Mul; uploads bias to GPU once at session init; compiles DML_ELEMENT_WISE_ADD1 with fused GELU activation - Shape changes recompile a temporary kernel per call, matching ORT DML EP behavior
zz002
force-pushed
the
feat/rocm-llm-backend
branch
from
ff40ecc to
01e7cb3
Compare
zz002
changed the title
zz002
force-pushed
the
feat/rocm-llm-backend
branch
3 times, most recently
from
2f32dad to
e78f4b1
Compare
zz002
changed the title
* Add mlss_use_specific_ops as EP option * Add putenv for getenv compatibility * Add exhaustive_tune ep option to amdgpu ep to match migraphx ep
…attern matcher (onnxruntime#27) The BiasGelu pattern tree matched nodes it couldn't compile — the rank-1 bias check ran in Compile() after nodes were already claimed via AddNodesToFuse, causing ORT_EP_FAIL and session abort. Add a Predicate input constraint kind to the pattern matching engine that runs caller-provided lambda checks inline during graph traversal. This allows semantic validation (rank, shape, initializer membership) at pattern-match time alongside structural matching. Changes: - pattern_matcher.h: Add Predicate kind with PredicateFn accepting node pointer for cross-input validation. Add PInitializerWithMaxRank helper. - pattern_matcher.cc: Pre-fetch OrtValueInfo map per node, handle Predicate in both fixed-index and scan-all branches. - bias_gelu_ep_fusion.cc: Replace post-match validation with PBiasInput predicate (rank-1 check, last-dim match). Compile reads bias name from pattern capture. Guard deferred compilation when shapes are unknown. - ep_fusion_manager.cc: Move MatchesResult before AddNodesToFuse to prevent claiming nodes no rule owns.
Add the hipep backend to the AMD GPU umbrella EP: load the pre-built hipep-backend plugin EP DLL as a third backend (alongside DirectML and MIGraphX) and route to it when the user selects profile=llm. - Profile::Llm parses "llm" / "3"; gpu_ep.cc routes profile=llm to ProviderFactory::CreateHipepBackend. - ProviderFactory loads hipep-backend and forwards CreateEp / ReleaseEp to it, mirroring the DirectML/MIGraphX backend pattern. hipep is a pre-built plugin EP DLL loaded at runtime, so no ROCm SDK is required to build amdgpu-ep. - CreateEp forwards the ORT-selected hardware device(s) + ep_metadata to the backend's CreateEp. Unlike DirectML/MIGraphX, hipep's CreateEp rejects num_devices == 0, so the umbrella must pass the selected device. DirectML and MIGraphX paths are unchanged. Co-authored-by: Cursor <cursoragent@cursor.com>
…ctML/MIGraphX The hipep backend no longer requires the umbrella to forward the ORT-selected device: its CreateEp now tolerates a null device array / num_devices == 0 (ROCm/MorphiZen#234). Drop the devices/ep_metadata parameters from ProviderFactory::CreateHipepBackend and the ExecutionProvider constructor and pass nullptr/nullptr/0, so the hipep path matches the DirectML and MIGraphX backends and the EP constructor keeps the same signature as the sibling EPs. Co-authored-by: Cursor <cursoragent@cursor.com>
Add src/hipep parallel to directml/migraphx. It builds hipep-backend.dll, a thin pass-through OrtEpFactory that loads hipep.dll, forwards CreateEpFactories, and returns the HIP EP's factory unchanged so the AMD GPU umbrella forwards to it directly. ReleaseEpFactory forwards to the HIP EP and unloads the DLL. Only a single factory is supported. Wired via a new USE_HIPEP option (forced on by USE_AMDGPU) and add_subdirectory(hipep). Co-authored-by: Cursor <cursoragent@cursor.com>
CreateEpFactories can be invoked more than once (e.g. a real session plus the device-init session in OGA). The previous single-handle logic reloaded the DLL on each call (leaking the earlier handle) and unloaded it on the first ReleaseEpFactory while later factories were still live. Load on the first factory and unload only after the last is released; roll back the refcount if the load fails so a retry re-attempts it. Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Address PR review: the backend name must not contain "ep". Rename the umbrella backend shim from hipep to hip (src/hip/, hip-ep target, hip-backend.dll, USE_HIP option, and all C++ identifiers). The shim now loads the HIP EP DLL as hipgpu (hipgpu.dll / libhipgpu.so), to be renamed on the onnx-hipdnn-ep side in a follow-up. Co-authored-by: Cursor <cursoragent@cursor.com>
zz002
force-pushed
the
feat/rocm-llm-backend
branch
from
4e9aa28 to
71c00fb
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
5 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
Add the hipep backend to the AMD GPU umbrella EP, selected via the
profile=llmprovider option. The hipep plugin EP (hipep-backend.dll) is loaded as a third backend alongside DirectML and MIGraphX, reusing the existing dynamic-backend pattern.What
Profile::Llm(gpu_info) accepts"llm"/"3".ProviderFactory::CreateHipepBackendloadshipep-backend, forwardsCreateEp/ReleaseEp, mirroring the DirectML/MIGraphX backends. hipep is a pre-built plugin EP DLL loaded at runtime, so no ROCm SDK is required to build amdgpu-ep.gpu_ep.ccroutesprofile=llmtoCreateHipepBackend.CreateEpforwards the ORT-selected hardware device(s) +ep_metadatato the backend. Unlike DirectML/MIGraphX, hipepsCreateEprejectsnum_devices == 0`, so the umbrella must pass the selected device.DirectML and MIGraphX paths are unchanged. No CMake changes - hipep is a pure code addition, always compiled into amdgpu-ep.
Test plan
Validated on gfx1151 (with the stacked build PR so DirectML/MIGraphX are off): a MatMul model run through
AMDExecutionProviderwithprofile=llmcompiles via hip-compiler and executes through hipBLASLt; output matches the CPU reference (max abs diff 0), no fallback or crash. Requires the companion hipdnn-ep factory fix (ROCm/MorphiZen#234).Notes for reviewers
USE_AMDGPUforces both on,migraphx-epneeds the MIGraphX SDK). Building/running with only hipep uses the stacked build-scaffolding PR (not for merge).hipep-backend.dll; the hipdnn-ep build currently producesonnxruntime_morphizen_ep.dll, so the output name needs to be aligned (rename the build target or ship a renamed copy).