Benign huggingface Download

AndrewThompson1233/maba-v2-architecture main

pytest test frameworkPython repeats repository clone calls

SHA-25679fc78a6fa39c44e21b779192253b3a4923e71d5991cb94f353044e12c20d349

URLhttps://huggingface.co/AndrewThompson1233/maba-v2-architecture

Evidence

Imports the pytest module lines 1–10

1
2import gc
3import pytest
4import torch
5
⋯5 lines

pytest fixture or mark reference lines 21–31

21:32… ,
22)
23
24CUDA_AVAILABLE = torch.cuda.is_available() and is_cuda_sm75_available()
25TRITON_ACTIVE = CUDA_AVAILABLE and is_triton_available()
26
27
28@pytest.fixture(autouse=True)
29def reset_dispatcher(monkeypatch):
30    monkeypatch.delenv("MABA_BACKEND", raising=False)
31    monkeypatch.delenv(" …

Python test source content lines 35–48

⋯4 lines
39class TestCentroidAutogradGradientStress:
40
41    @pytest.mark.skipif(not CUDA_AVAILABLE, reason="Requires CUDA sm_75+")
42    @pytest.mark.parametrize("B,L,d_idx", [
43        (1, 64, 64),
44        (2, 128, 64),
⋯4 lines

Checks condition using Python assert lines 61–73

17… entroids.sum()
      loss.backward()
63
      assert k.grad is not None
      assert not torch.isnan(k.grad).any()
      assert not torch.isinf(k.grad).any()
67
      min_bound = 1.0 / (2.0 * block_size) - 1e-6
      assert (k.grad >= min_bound).all(), (
          f"Positivity bound violated: min {k.grad.min().item()} < {min_bound} at shape ({B}, {L}, {d_idx})"
      )
72
  @pytest.mark.skipif(not CUDA_AVAILA …

Python repeats repository clone calls lines 154–158

50… cuda", dtype=torch.float32, requires_grad=True)
      k_tri = k_ref.detach().clone().requires_grad_(True)
156
      c_ref = ref_compute_centroids(k_ref, block_size=block_size)
      c_tri = triton_compute_ce …