mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-10-28 23:07:09 +08:00
760e3cadc0
- tests/neuron/setup_trn2.sh: one-shot setup of a fresh Trn2 (venv layered
on the DLAMI torch-neuronx venv, C++ build, official reference + tokenizer,
tiny checkpoint, CPU golden).
- tests/neuron/run_all.sh cpu|device|all [stage...]: vendor kernel tests,
tilelang_ops interface tests (basic-mode stubs reported as skips via
nie_skip), deep_ep at world size 1/2/4 (gloo) and 2/4/8 on the neuron
process group, device primitive checks, end-to-end vs golden on CPU and
on a NeuronCore, and the device diff; PASS/FAIL summary.
- tests/neuron/device/test_device_primitives.py: torch-neuronx behaviours
the port depends on or works around (fp8 strided-view indexing, uint64,
1-rank all_to_all, inference-mode version counters, ...).
- examples/neuron/device_diff.py: re-runs every kernel call of a NeuronCore
run on host copies and ranks kernels by device/host difference.
- Unit tests written during bring-up moved from scratch dirs into
tests/neuron/{deep_gemm,flash_mla,deep_ep,tile_kernels,sampling}.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
18 lines
808 B
Python
18 lines
808 B
Python
"""pytest plugin: report kernels stubbed out in Neuron basic mode as skipped, not failed.
|
|
|
|
The Neuron backend exports every asc kernel factory, but those reachable only from DSpark/MTP, ZOS, CP or
|
|
VL raise ``NotImplementedError("<name>: not supported on Neuron (basic mode)")``. The upstream interface
|
|
tests still parametrize over them; with ``-p nie_skip`` they show up as skips.
|
|
"""
|
|
|
|
import pytest
|
|
|
|
|
|
@pytest.hookimpl(hookwrapper=True)
|
|
def pytest_runtest_makereport(item, call):
|
|
outcome = yield
|
|
report = outcome.get_result()
|
|
if call.excinfo is not None and call.excinfo.errisinstance(NotImplementedError) and "basic mode" in str(call.excinfo.value):
|
|
report.outcome = "skipped"
|
|
report.longrepr = (str(item.path), item.location[1] or 0, f"Neuron basic mode: {call.excinfo.value}")
|