| # Tests | |
| | Marker | File | Needs | What it checks | | |
| |---|---|---|---| | |
| | `sim` | `test_sim.py` | `nki`, `numpy`, `ml_dtypes` | each kernel under `nki.simulate` vs a NumPy fp32 reference (ragged padding, LNC 1/2, query split, padding isolation) | | |
| | `cpu` | `test_pack_cpu.py` | `torch`, `transformers` | host-side weight packing reproduces the Hugging Face `NewLayer` attention half in fp32 (`GTE_TEST_MODEL` picks the model) | | |
| | `sim` + `cpu` | `test_patch_sim.py` | all of the above | `patch_model` end to end on a 2-layer model with the kernels run by the simulator | | |
| | `neuron` | `test_device.py` | trn2 + PyTorch Native | full patched model vs CPU fp32 and vs stock PyTorch Native: score Pearson >= 0.9999, top-1 agreement >= stock, padding isolation bit-exact, fused ops actually used | | |
| ```bash | |
| python -m pytest -m sim # local, no device | |
| python -m pytest -m cpu # local, downloads the model | |
| NEURON_LOGICAL_NC_CONFIG=1 python -m pytest -m neuron -s tests/test_device.py # on trn2, once per LNC | |
| NEURON_LOGICAL_NC_CONFIG=2 python -m pytest -m neuron -s tests/test_device.py | |
| ``` | |
| A simulator pass is not device evidence. The simulator does not model SBUF/PSUM capacity, and some DMA | |
| access patterns behave differently on hardware (for example, one strided DMA spanning several | |
| 32-partition blocks is exact in the simulator but copies only the first block on trn2). | |