Capability Matrix¶
This page is generated by scripts/generate_capability_docs.py.
- Runtime capability source:
docs/runtime_capabilities.json - Tested environment source:
docs/tested_environments.json
Regenerate:
python scripts/generate_capability_docs.py --refresh-runtime-snapshot
Runtime Capability Matrix (Auto-generated)¶
- Snapshot generated at (UTC):
2026-03-31T15:54:26+00:00 - Active backend at snapshot time:
cpu - Active memory model:
host-managed-compat - Note:
Availableis host-dependent. Regenerate the snapshot on your target machine for exact values. - Generated by:
python scripts/generate_capability_docs.py --refresh-runtime-snapshot
| Device | Built | Available | Compute | Memory | Sync | Profiling | Trace | Sync Policy | Memory Model | Query |
|---|---|---|---|---|---|---|---|---|---|---|
| metal | Yes | No | Yes | Yes | Yes | No | Yes | Yes | host-managed-compat | ok |
| cpu | Yes | Yes | Yes | No | No | No | Yes | Yes | host-managed-compat | ok |
| cuda | No | No | No | No | No | No | Yes | Yes | native-device | ok |
Runtime Trace / Capability API Surface (Auto-generated)¶
| Runtime API Surface | Available | Notes |
|---|---|---|
| runtime_trace_enable | Yes | Enable/disable trace capture |
| runtime_trace_events | Yes | Raw runtime event list |
| runtime_trace_timeline | Yes | Sorted/grouped timeline report |
| runtime_trace_clear | Yes | Clear ring buffer events |
| runtime_backend_capabilities | Yes | Per-backend contract query (metal/cpu/cuda) |
| runtime_active_backend_capabilities | Yes | Capability contract of current active backend |
Tensor Contract Matrix (Auto-generated, Frozen)¶
- Source of truth:
docs/tensor_contract_matrix.json - Contract regressions are hard-gated by CI (
ci-contract-tests.yml).
| Category | Contract | Invariant | C++ Gate | Python Gate |
|---|---|---|---|---|
| shape | Tensor shape must be non-empty and each dim > 0. | Invalid shape metadata is rejected by contract validators. | test_tensor_contract_freeze, test_tensor | tests/test_python_tensor_contract_smoke.py |
| layout | Contiguous tensors keep canonical strides; slice views are strided with bounded offsets. | View metadata (shape/strides/layout/offset) must pass bounds checks against storage. | test_tensor_contract_freeze, test_tensor, test_graph_ir | tests/test_python_tensor_contract_smoke.py |
| lifetime | fromHost/toHost data exchange is copy-based and must not alias caller-owned buffers. | Caller buffer mutation after transfer cannot affect tensor/device-owned state. | test_tensor_contract_freeze, test_graph_ir | tests/test_python_tensor_contract_smoke.py |
| alias | TensorView is metadata alias over tensor storage, with strict same-device contract. | Valid same-device view reads reflect latest tensor storage; cross-device view contract fails. | test_tensor_contract_freeze, test_tensor | tests/test_python_tensor_contract_smoke.py |
Tested Environment Matrix (Auto-generated)¶
| Date | Scope | Hardware / OS | Python | Torch | Status | Notes |
|---|---|---|---|---|---|---|
| 2026-03-30 | Local benchmark snapshot | Apple Silicon macOS (Metal enabled) | 3.14 | 2.11.0 | validated | README benchmark snapshot run (ai_model_all_bench.py / ml_all_bench.py / dl_all_bench.py). |
| 2026-04-01 | CI contract tests | GitHub Actions macos-14 | 3.12 | n/a | validated | CMake + CTest quality gate workflow. |
| 2026-04-01 | CI wheel build and publish | GitHub Actions macos-14 | 3.12 | n/a | validated | Wheel build + distribution validation in publish workflow. |