Add script to generate tripy integration test stablehlo ir #490
Conversation
| for test_name, asm in test_assemblies.items(): | ||
| f.write(f' "{test_name}": """{asm}""",\n') | ||
|
|
There was a problem hiding this comment.
How large is the generated code usually? Wondering if we might need to split them across tests instead of just test files?
There was a problem hiding this comment.
I have not generated the code for the complete test suit (WIP). I will update once done.
There was a problem hiding this comment.
Generated code is about 30k lines and takes around 15 minutes to run end to end in MLIR-TRT. May be we can perform some deduplication i.e. just include first test case in a class. or may be just include a single test per module.
There was a problem hiding this comment.
Does a naive deduplication help at all? e.g. just comparing the generated StableHLO textual representation directly
| module_match = re.search(r'<Module (.*?)>', relevant_output) | ||
| module_name = module_match.group(1) if module_match else '' | ||
| class_match = re.search(r'<Class (.*?)>', relevant_output) | ||
| class_name = class_match.group(1) if class_match else '' |
There was a problem hiding this comment.
This might be incorrect if we have more than one class per module.
This script can be run periodically (or during CI/CD) to generate Tripy integration test IRs. For now, we generate stablehlo IR but script itself is IR agonistic. This will generate a file
test_tripy_integration.pywhich can be copied tomlir-tensorrt/compiler/test/python/IntegrationTests/test_tripy_integration.pyand committed to mitigating Tripy related regressions in MLIR-TensorRT.