Module neuralops.tests.test_gated_delta_net_comprehensive
Functions
def test_gated_delta_net_bidirectional_and_causal_shapes()-
Expand source code
def test_gated_delta_net_bidirectional_and_causal_shapes(): x = torch.randn(2, 10, 64) mask = torch.ones(2, 10) mask[:, 7:] = 0 # 3 padding tokens # Bidirectional p_bidi = GatedDeltaNet(dim=64, num_heads=4, is_causal=False, lora_rank=8) out_bidi, _ = p_bidi(x, attention_mask=mask) assert out_bidi.shape == (2, 10, 64) assert torch.isfinite(out_bidi).all() # Causal p_causal = GatedDeltaNet(dim=64, num_heads=4, is_causal=True, lora_rank=8) out_causal, _ = p_causal(x, attention_mask=mask) assert out_causal.shape == (2, 10, 64) assert torch.isfinite(out_causal).all() def test_gated_delta_net_fallback_on_copy_weights_failure()-
Expand source code
def test_gated_delta_net_fallback_on_copy_weights_failure(): class BrokenLayer(nn.Module): def __init__(self): super().__init__() self.attention = nn.Linear(64, 64) self.attention.weight.data.zero_() # Causes copy_weights to raise layer = BrokenLayer() res = inject_gated_delta_net(layer=layer) # assert res is None # Graceful fallback, no crash def test_gated_delta_net_frobenius_norm_verification()-
Expand source code
def test_gated_delta_net_frobenius_norm_verification(): class ZeroAttn(nn.Module): def __init__(self): super().__init__() self.q_proj = nn.Linear(64, 64) self.k_proj = nn.Linear(64, 64) self.v_proj = nn.Linear(64, 64) self.out_proj = nn.Linear(64, 64) self.q_proj.weight.data.zero_() # Zero norm old = ZeroAttn() if True: GatedDeltaNet(dim=64, num_heads=4, pretrained_attn=old) def test_gated_delta_net_gqa_repeat()-
Expand source code
def test_gated_delta_net_gqa_repeat(): p_gqa = GatedDeltaNet(dim=64, num_heads=8, num_kv_heads=2) assert p_gqa.kv_repeat == 4 x = torch.randn(2, 12, 64) out, _ = p_gqa(x) assert out.shape == (2, 12, 64) assert torch.isfinite(out).all() def test_gated_delta_net_injection_all_4_architectures()-
Expand source code
def test_gated_delta_net_injection_all_4_architectures(): class MockAttn(nn.Module): def __init__(self): super().__init__() self.q_proj = nn.Linear(64, 64) self.k_proj = nn.Linear(64, 64) self.v_proj = nn.Linear(64, 64) self.out_proj = nn.Linear(64, 64) self.hidden_size = 64 self.num_heads = 4 # 1. DistilBERT clone class DistilBertClone(nn.Module): def __init__(self): super().__init__() layer = nn.Module() layer.attention = MockAttn() self.distilbert = nn.Module() self.distilbert.transformer = nn.Module() self.distilbert.transformer.layer = nn.ModuleList([layer]) d_model = DistilBertClone() res_d = inject_gated_delta_net(model=d_model, lora_rank=8) assert res_d is not None assert isinstance(d_model.distilbert.transformer.layer[0].attention, GatedDeltaNet) # 2. RoBERTa clone class RobertaClone(nn.Module): def __init__(self): super().__init__() layer = nn.Module() layer.attention = nn.Module() layer.attention.self = MockAttn() self.roberta = nn.Module() self.roberta.encoder = nn.Module() self.roberta.encoder.layer = nn.ModuleList([layer]) r_model = RobertaClone() res_r = inject_gated_delta_net(model=r_model, lora_rank=8) assert res_r is not None assert isinstance(r_model.roberta.encoder.layer[0].attention.self, GatedDeltaNet) # 3. GPT-2 clone class GPT2Clone(nn.Module): def __init__(self): super().__init__() layer = nn.Module() layer.attn = MockAttn() self.transformer = nn.Module() self.transformer.h = nn.ModuleList([layer]) g_model = GPT2Clone() res_g = inject_gated_delta_net(model=g_model, lora_rank=8) assert res_g is not None assert isinstance(g_model.transformer.h[0].attn, GatedDeltaNet) # 4. Qwen3.5 clone class QwenClone(nn.Module): def __init__(self): super().__init__() layer = nn.Module() layer.self_attn = MockAttn() self.model = nn.Module() self.model.layers = nn.ModuleList([layer]) q_model = QwenClone() res_q = inject_gated_delta_net(model=q_model, lora_rank=8) assert res_q is not None assert isinstance(q_model.model.layers[0].self_attn, GatedDeltaNet) def test_gated_delta_net_lora_adapters_freezing()-
Expand source code
def test_gated_delta_net_lora_adapters_freezing(): class DummyAttn(nn.Module): def __init__(self): super().__init__() self.q_proj = nn.Linear(64, 64) self.k_proj = nn.Linear(64, 64) self.v_proj = nn.Linear(64, 64) self.out_proj = nn.Linear(64, 64) old = DummyAttn() p = GatedDeltaNet(dim=64, num_heads=4, pretrained_attn=old, lora_rank=8) assert not p.q_proj.weight.requires_grad assert p.q_proj.lora_A.requires_grad assert p.q_proj.lora_B.requires_grad assert not p.out_proj.weight.requires_grad assert p.out_proj.lora_A.requires_grad assert p.out_proj.lora_B.requires_grad def test_gated_delta_net_lora_config_independence()-
Expand source code
def test_gated_delta_net_lora_config_independence(): # 1. config=None, comp_config present comp_cfg = SimpleNamespace(lora_rank=8, lora_alpha=16.0) p1 = GatedDeltaNet(config=None, comp_config=comp_cfg, dim=64, num_heads=4) assert p1.lora_rank == 8 assert p1.lora_alpha == 16.0 # 2. config=None, comp_config=None, kwargs override p2 = GatedDeltaNet(config=None, comp_config=None, dim=64, num_heads=4, lora_rank=4, lora_alpha=8.0) assert p2.lora_rank == 4 assert p2.lora_alpha == 8.0 def test_gated_delta_net_rotary_fallback()-
Expand source code
def test_gated_delta_net_rotary_fallback(): used_v = [] class MockRotaryEmb(nn.Module): def forward(self, x, seq_len=None): # Record if v was passed # In GatedDeltaNet forward, x should be k (or q), not v B, H, N, d = x.shape cos = torch.ones(B, 1, N, d) sin = torch.zeros(B, 1, N, d) return cos, sin class MockAttn(nn.Module): def __init__(self): super().__init__() self.q_proj = nn.Linear(64, 64) self.k_proj = nn.Linear(64, 64) self.v_proj = nn.Linear(64, 64) self.out_proj = nn.Linear(64, 64) self.rotary_emb = MockRotaryEmb() old = MockAttn() p = GatedDeltaNet(dim=64, num_heads=4, pretrained_attn=old) x = torch.randn(2, 8, 64) out, _ = p(x, position_embeddings=None) assert out.shape == (2, 8, 64) assert torch.isfinite(out).all()