| 1 | import unittest |
| 2 | |
| 3 | import torch |
| 4 | from compressed_tensors.compressors.pack_quantized.helpers import pack_to_int32 |
| 5 | from compressed_tensors.quantization.lifecycle.forward import dequantize |
| 6 | |
| 7 | from train import FrozenEmbedding, FrozenLinear, NeuronDelta, batch_loss |
| 8 | from evaluate import SteeredMLP |
| 9 | |
| 10 | |
| 11 | class Training(unittest.TestCase): |
| 12 | @classmethod |
| 13 | def setUpClass(cls): |
| 14 | torch.set_num_threads(1) |
| 15 | |
| 16 | def test_frozen_linear_forward_and_input_gradient_match_dequantized_reference(self): |
| 17 | torch.manual_seed(42) |
| 18 | for output_size in [16, 65540]: |
| 19 | for asymmetric in [False, True]: |
| 20 | values = torch.randint(-8, 8, (output_size, 16), dtype=torch.int8) |
| 21 | scale = torch.rand(output_size, 2) + 0.1 |
| 22 | zero = torch.randint(-4, 4, (output_size, 2), dtype=torch.int8) if asymmetric else None |
| 23 | module = FrozenLinear(pack_to_int32(values, 4), scale, values.shape, |
| 24 | pack_to_int32(zero, 4, packed_dim=0) if zero is not None else None) |
| 25 | weight = dequantize(values, scale, zero) |
| 26 | x = torch.randn(2, 3, 16, requires_grad=True) |
| 27 | result = module(x) |
| 28 | self.assertTrue(torch.allclose(result, x @ weight.T, atol=1e-5, rtol=1e-5)) |
| 29 | gradient = torch.randn_like(result) |
| 30 | result.backward(gradient) |
| 31 | self.assertTrue(torch.allclose(x.grad, gradient @ weight, atol=0.004, rtol=0.0001)) |
| 32 | self.assertEqual(list(module.parameters()), []) |
| 33 | |
| 34 | def test_embedding_selects_only_requested_rows(self): |
| 35 | values = torch.randint(-128, 128, (32, 16), dtype=torch.int8) |
| 36 | scale = torch.rand(32, 2) + 0.1 |
| 37 | module = FrozenEmbedding(pack_to_int32(values, 8), scale, values.shape, bits=8) |
| 38 | ids = torch.tensor([[7, 1, 7], [0, 31, 9]]) |
| 39 | self.assertTrue(torch.equal(module(ids), dequantize(values, scale)[ids])) |
| 40 | |
| 41 | def test_zero_neuron_delta_preserves_output_and_only_three_vectors_train(self): |
| 42 | class MLP(torch.nn.Module): |
| 43 | def __init__(self): |
| 44 | super().__init__() |
| 45 | for name in ["gate_proj", "up_proj", "down_proj"]: |
| 46 | values = torch.randint(-8, 8, (16, 16), dtype=torch.int8) |
| 47 | setattr(self, name, FrozenLinear(pack_to_int32(values, 4), torch.ones(16, 2), values.shape)) |
| 48 | self.act_fn = torch.nn.SiLU() |
| 49 | |
| 50 | def forward(self, x): |
| 51 | return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) |
| 52 | |
| 53 | mlp = MLP() |
| 54 | module = NeuronDelta(mlp, 7) |
| 55 | x = torch.randn(1, 3, 16) |
| 56 | self.assertTrue(torch.equal(module(x), mlp(x))) |
| 57 | module(x).square().mean().backward() |
| 58 | self.assertEqual(sum(p.numel() for p in module.parameters()), 48) |
| 59 | self.assertTrue(all(p.grad is not None and torch.isfinite(p.grad).all() for p in module.parameters())) |
| 60 | |
| 61 | with torch.no_grad(): |
| 62 | for parameter in module.parameters(): |
| 63 | parameter.normal_(std=0.01) |
| 64 | base = torch.nn.Module() |
| 65 | base.gate_up_proj = lambda x: (torch.cat([mlp.gate_proj(x), mlp.up_proj(x)], -1), None) |
| 66 | base.act_fn = lambda x: torch.nn.functional.silu(x[..., :16]) * x[..., 16:] |
| 67 | base.down_proj = lambda x: (mlp.down_proj(x), None) |
| 68 | base.forward = lambda x: base.down_proj(base.act_fn(base.gate_up_proj(x)[0]))[0] |
| 69 | wrapped = SteeredMLP(base, {name: getattr(module, name).detach() for name in ["gate", "up", "down"]}, |
| 70 | {"neuron": "7"}) |
| 71 | self.assertTrue(torch.equal(wrapped(x), mlp(x))) |
| 72 | scaled = NeuronDelta(mlp, 7) |
| 73 | for strength in [0.0, 0.5, 1.0]: |
| 74 | tensors = {name: getattr(module, name).detach() * strength for name in ["gate", "up", "down"]} |
| 75 | with torch.no_grad(): |
| 76 | for name, tensor in tensors.items(): |
| 77 | getattr(scaled, name).copy_(tensor) |
| 78 | wrapped = SteeredMLP(base, tensors, {"neuron": "7"}) |
| 79 | wrapped.enabled = True |
| 80 | self.assertTrue(torch.allclose(wrapped(x), scaled(x), atol=0.0001, rtol=0.0001)) |
| 81 | |
| 82 | def test_checkpointed_hybrid_model_backpropagates_across_multiple_recurrent_chunks(self): |
| 83 | from transformers import Qwen3_5ForCausalLM, Qwen3_5TextConfig |
| 84 | |
| 85 | config = Qwen3_5TextConfig( |
| 86 | hidden_size=64, intermediate_size=128, num_hidden_layers=4, |
| 87 | vocab_size=128, num_attention_heads=4, num_key_value_heads=2, head_dim=16, |
| 88 | linear_num_key_heads=2, linear_num_value_heads=2, |
| 89 | linear_key_head_dim=16, linear_value_head_dim=16, |
| 90 | layer_types=["linear_attention"] * 3 + ["full_attention"], |
| 91 | rope_parameters={"rope_type": "default", "rope_theta": 10000000.0, |
| 92 | "partial_rotary_factor": 1.0, "mrope_section": [3, 3, 2]}, |
| 93 | use_cache=False, pad_token_id=0, bos_token_id=1, eos_token_id=2, |
| 94 | ) |
| 95 | config._attn_implementation = "sdpa" |
| 96 | model = Qwen3_5ForCausalLM(config).float() |
| 97 | model.requires_grad_(False) |
| 98 | mlp = model.model.layers[0].mlp |
| 99 | for name in ["gate_proj", "up_proj", "down_proj"]: |
| 100 | values = torch.randint(-8, 8, getattr(mlp, name).weight.shape, dtype=torch.int8) |
| 101 | setattr(mlp, name, FrozenLinear(pack_to_int32(values, 4), |
| 102 | torch.full((len(values), values.shape[1] // 16), 0.01), values.shape)) |
| 103 | module = NeuronDelta(mlp, 7) |
| 104 | model.model.layers[0].mlp = module |
| 105 | model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False}) |
| 106 | model.train() |
| 107 | ids = torch.randint(0, 128, (1, 160)) |
| 108 | model(input_ids=ids, labels=ids, use_cache=False).loss.backward() |
| 109 | trained = {name for name, p in model.named_parameters() if p.requires_grad} |
| 110 | self.assertEqual(trained, {"model.layers.0.mlp.gate", "model.layers.0.mlp.up", "model.layers.0.mlp.down"}) |
| 111 | self.assertTrue(all(p.grad is not None and torch.isfinite(p.grad).all() for p in module.parameters())) |
| 112 | |
| 113 | rows = [(ids[0, :9].tolist(), ids[0, 9:73].tolist()), |
| 114 | (ids[0, :17].tolist(), ids[0, 17:].tolist())] |
| 115 | module.zero_grad(set_to_none=True) |
| 116 | batched = batch_loss(model, rows) |
| 117 | batched.backward() |
| 118 | gradients = [p.grad.clone() for p in module.parameters()] |
| 119 | module.zero_grad(set_to_none=True) |
| 120 | separate = torch.stack([batch_loss(model, [row]) for row in rows]).mean() |
| 121 | separate.backward() |
| 122 | self.assertTrue(torch.allclose(batched, separate, atol=1e-6, rtol=1e-5)) |
| 123 | for parameter, expected in zip(module.parameters(), gradients, strict=True): |
| 124 | self.assertTrue(torch.allclose(parameter.grad, expected, atol=1e-6, rtol=0.001)) |
| 125 | |
| 126 | |
| 127 | if __name__ == "__main__": |
| 128 | unittest.main() |