|
| 1 | +# Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# |
| 3 | +# See LICENSE for license information. |
| 4 | + |
| 5 | +import torch |
| 6 | +from transformer_engine.pytorch import LayerNormMLP |
| 7 | +import pytest |
| 8 | + |
| 9 | +torch.manual_seed(1234) |
| 10 | +device = torch.device("cuda") |
| 11 | + |
| 12 | + |
| 13 | +class _Sequential(torch.nn.Sequential): |
| 14 | + """Sequential model that forwards keyword arguments to modules""" |
| 15 | + |
| 16 | + def forward(self, input_: torch.Tensor, **kwargs) -> torch.Tensor: |
| 17 | + x = input_ |
| 18 | + for module in self: |
| 19 | + x = module(x, **kwargs) |
| 20 | + return x |
| 21 | + |
| 22 | + |
| 23 | +class ModelConfig: |
| 24 | + def __init__( |
| 25 | + self, |
| 26 | + hidden_size: int = 128, |
| 27 | + ffn_hidden_size: int = 512, |
| 28 | + layers: int = 1, |
| 29 | + ): |
| 30 | + self._hidden_size = hidden_size |
| 31 | + self._ffn_hidden_size = ffn_hidden_size |
| 32 | + self._layers = layers |
| 33 | + |
| 34 | + def build(self): |
| 35 | + |
| 36 | + ln_list, sln_list = [], [] |
| 37 | + for _ in range(self._layers): |
| 38 | + ln = LayerNormMLP(self._hidden_size, self._ffn_hidden_size, checkpoint=False).to(device) |
| 39 | + sln = LayerNormMLP(self._hidden_size, self._ffn_hidden_size, checkpoint=True).to(device) |
| 40 | + with torch.no_grad(): |
| 41 | + sln.layer_norm_weight = torch.nn.Parameter(ln.layer_norm_weight.clone()) |
| 42 | + sln.layer_norm_bias = torch.nn.Parameter(ln.layer_norm_bias.clone()) |
| 43 | + sln.fc1_weight = torch.nn.Parameter(ln.fc1_weight.clone()) |
| 44 | + sln.fc2_weight = torch.nn.Parameter(ln.fc2_weight.clone()) |
| 45 | + sln.fc1_bias = torch.nn.Parameter(ln.fc1_bias.clone()) |
| 46 | + sln.fc2_bias = torch.nn.Parameter(ln.fc2_bias.clone()) |
| 47 | + ln_list.append(ln) |
| 48 | + sln_list.append(sln) |
| 49 | + |
| 50 | + ln_model = _Sequential(*ln_list) |
| 51 | + sln_model = _Sequential(*sln_list) |
| 52 | + |
| 53 | + return ln_model, sln_model |
| 54 | + |
| 55 | + |
| 56 | +config = { |
| 57 | + "small": ModelConfig(128, 512, 12), |
| 58 | + "medium": ModelConfig(512, 2048, 12), |
| 59 | + "large": ModelConfig(1024, 4096, 12), |
| 60 | + "huge": ModelConfig(2048, 8192, 12), |
| 61 | +} |
| 62 | + |
| 63 | +seq_sizes = [2**7, 2**10, 2**14, 2**16] |
| 64 | + |
| 65 | + |
| 66 | +def _warmup(model, tensor): |
| 67 | + for _ in range(3): |
| 68 | + model(tensor).sum().backward() |
| 69 | + |
| 70 | + |
| 71 | +def _run_fwd(model, tensor): |
| 72 | + |
| 73 | + torch.cuda.reset_peak_memory_stats(device) |
| 74 | + start_time, end_time = torch.cuda.Event(enable_timing=True), torch.cuda.Event( |
| 75 | + enable_timing=True |
| 76 | + ) |
| 77 | + |
| 78 | + torch.cuda.synchronize() |
| 79 | + start_mem = torch.cuda.memory_allocated(device) |
| 80 | + start_time.record() |
| 81 | + out = model(tensor) |
| 82 | + end_time.record() |
| 83 | + end_time.synchronize() |
| 84 | + elapsed = start_time.elapsed_time(end_time) |
| 85 | + peak_mem = torch.cuda.max_memory_allocated(device) |
| 86 | + mem = float(peak_mem - start_mem) |
| 87 | + |
| 88 | + return out, elapsed, mem |
| 89 | + |
| 90 | + |
| 91 | +def _run_bwd(model, out): |
| 92 | + |
| 93 | + model.zero_grad(set_to_none=False) |
| 94 | + loss = out.sum() |
| 95 | + |
| 96 | + torch.cuda.reset_peak_memory_stats(device) |
| 97 | + start_time, end_time = torch.cuda.Event(enable_timing=True), torch.cuda.Event( |
| 98 | + enable_timing=True |
| 99 | + ) |
| 100 | + |
| 101 | + torch.cuda.synchronize() |
| 102 | + start_mem = torch.cuda.memory_allocated(device) |
| 103 | + start_time.record() |
| 104 | + loss.backward() |
| 105 | + end_time.record() |
| 106 | + end_time.synchronize() |
| 107 | + elapsed = start_time.elapsed_time(end_time) |
| 108 | + peak_mem = torch.cuda.max_memory_allocated(device) |
| 109 | + mem = float(peak_mem - start_mem) |
| 110 | + |
| 111 | + param_grads = _collect_param_grads(model) |
| 112 | + return param_grads, elapsed, mem |
| 113 | + |
| 114 | + |
| 115 | +def _max_diff(ref, other): |
| 116 | + """Return max absolute difference between two tensors or collections.""" |
| 117 | + if ref is None or other is None: |
| 118 | + return 0.0 |
| 119 | + if isinstance(ref, (list, tuple)): |
| 120 | + diffs = [_max_diff(r, o) for r, o in zip(ref, other)] |
| 121 | + return max(diffs) if diffs else 0.0 |
| 122 | + return torch.max(torch.abs(ref.detach() - other.detach())).item() |
| 123 | + |
| 124 | + |
| 125 | +def _collect_param_grads(model): |
| 126 | + grads = {} |
| 127 | + for name, param in model.named_parameters(): |
| 128 | + if param.grad is None: |
| 129 | + continue |
| 130 | + key = _param_key(name) |
| 131 | + if key is not None: |
| 132 | + grads[key] = param.grad.detach().clone() |
| 133 | + return grads |
| 134 | + |
| 135 | + |
| 136 | +def _param_key(name): |
| 137 | + return name.split(".")[-1] |
| 138 | + |
| 139 | + |
| 140 | +@pytest.mark.parametrize("size", config.keys()) |
| 141 | +@pytest.mark.parametrize("seq_size", seq_sizes) |
| 142 | +def test_selective_activation_checkpoint(size, seq_size): |
| 143 | + |
| 144 | + ln_model, sln_model = config[size].build() |
| 145 | + data = torch.randn((seq_size, config[size]._hidden_size), device=device) |
| 146 | + |
| 147 | + _warmup(ln_model, data) |
| 148 | + ln_fwd_out, ln_fwd_time, ln_fwd_mem = _run_fwd(ln_model, data) |
| 149 | + ln_grads, ln_bwd_time, ln_bwd_mem = _run_bwd(ln_model, ln_fwd_out) |
| 150 | + |
| 151 | + _warmup(sln_model, data) |
| 152 | + sln_fwd_out, sln_fwd_time, sln_fwd_mem = _run_fwd(sln_model, data) |
| 153 | + sln_grads, sln_bwd_time, sln_bwd_mem = _run_bwd(sln_model, sln_fwd_out) |
| 154 | + |
| 155 | + assert ln_fwd_mem > 6 * sln_fwd_mem, ( |
| 156 | + "selective activation checkpointing does not reduce forward memory by 6X, only by" |
| 157 | + f" {ln_fwd_mem/sln_fwd_mem}!" |
| 158 | + ) |
| 159 | + assert ln_bwd_time < sln_bwd_time, ( |
| 160 | + "selective activation activation checkpointing backward pass is NOT slower than native!" |
| 161 | + f" got Native LayerNormMLP Backward Time: {ln_bwd_time} ms and Selective Activation" |
| 162 | + f" Checkpointed LayerNormMLP Backward Time: {sln_bwd_time} ms" |
| 163 | + ) |
| 164 | + diff = _max_diff(ln_fwd_out, sln_fwd_out) |
| 165 | + assert diff == 0.0, f"outputs are not equal! maximum difference {diff}" |
| 166 | + for key in [ |
| 167 | + "layer_norm_weight", |
| 168 | + "layer_norm_bias", |
| 169 | + "fc1_weight", |
| 170 | + "fc1_bias", |
| 171 | + "fc2_weight", |
| 172 | + "fc2_bias", |
| 173 | + ]: |
| 174 | + diff = _max_diff(ln_grads[key], sln_grads[key]) |
| 175 | + assert diff == 0.0, f"gradients for {key} are not equal! maximum difference: {diff}" |
![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&size=40){kind=avatar}
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?v=4&size=40){kind=avatar}
0 commit comments