5.4. Test Suites#
CTS (Compatibility Test Suite) - Compatibility Test Suite#
Test domain |
Coverage |
Pass standard |
|---|---|---|
Compiler compatibility |
FlagTree compilation and FLIR IR mapping correctness |
100% |
Operator compatibility |
FlagGems core operator correctness |
>=95% |
Communication compatibility |
FlagCX collective communication correctness |
100% |
Framework compatibility |
Basic FlagScale training and inference workflow |
100% |
PyTorch API compatibility |
Correctness of common aten API return values |
>=90% |
Triton API compatibility |
Triton language feature support |
>=95% |
Accuracy consistency |
Comparison with the NVIDIA baseline |
Deviation <2% |
VTS (Verification Test Suite) - Verification Test Suite#
Test domain |
Coverage |
Pass standard |
|---|---|---|
Stress test |
Continuous operator execution for 72 hours |
No crashes |
Stability test |
1,000 repeated training and inference runs |
Consistent results |
Long-duration stability |
Continuous 24/7 operation |
No memory leaks |
Resource monitoring |
GPU utilization, memory bandwidth, and temperature |
Within vendor specifications |
STS (Security Test Suite) - Security Test Suite#
Test domain |
Coverage |
Pass standard |
|---|---|---|
SBOM scanning |
Dependency vulnerability scanning |
No high-risk vulnerabilities or CVEs |
Driver security |
Driver file permission and signature validation |
Passed |
Container security |
Container image scanning |
No high-risk vulnerabilities |
Communication encryption |
FlagCX communication-link encryption validation |
Passed |
ITS (Interoperability Test Suite) - Interoperability Test Suite#
Test domain |
Coverage |
Pass standard |
|---|---|---|
Heterogeneous mixed training |
Chip and NVIDIA mixed training |
Efficiency >=81% |
Heterogeneous mixed inference |
Chip and NVIDIA mixed inference |
Accuracy deviation <2% |
Cross-vendor communication |
FlagCX cross-vendor testing |
Correctness 100% |
Multi-framework compatibility |
FlagScale, vLLM, and SGLang |
All passed |