|
| 1 | +# Copyright (c) 2025, NVIDIA CORPORATION. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +from typing import List |
| 16 | + |
| 17 | +import torch |
| 18 | + |
| 19 | +from dfm.src.megatron.data.common.diffusion_sample import DiffusionSample |
| 20 | +from dfm.src.megatron.data.common.diffusion_task_encoder_with_sp import DiffusionTaskEncoderWithSequencePacking |
| 21 | + |
| 22 | + |
| 23 | +class ConcreteDiffusionTaskEncoder(DiffusionTaskEncoderWithSequencePacking): |
| 24 | + """Concrete implementation for testing.""" |
| 25 | + |
| 26 | + def encode_sample(self, sample: dict) -> dict: |
| 27 | + """Simple implementation for testing purposes.""" |
| 28 | + return sample |
| 29 | + |
| 30 | + def batch(self, samples: List[DiffusionSample]) -> dict: |
| 31 | + """Simple batch implementation that returns first sample as dict.""" |
| 32 | + if len(samples) == 1: |
| 33 | + sample = samples[0] |
| 34 | + return dict( |
| 35 | + video=sample.video.unsqueeze(0), |
| 36 | + context_embeddings=sample.context_embeddings.unsqueeze(0), |
| 37 | + context_mask=sample.context_mask.unsqueeze(0) if sample.context_mask is not None else None, |
| 38 | + loss_mask=sample.loss_mask.unsqueeze(0) if sample.loss_mask is not None else None, |
| 39 | + seq_len_q=sample.seq_len_q, |
| 40 | + seq_len_q_padded=sample.seq_len_q_padded, |
| 41 | + seq_len_kv=sample.seq_len_kv, |
| 42 | + seq_len_kv_padded=sample.seq_len_kv_padded, |
| 43 | + pos_ids=sample.pos_ids.unsqueeze(0) if sample.pos_ids is not None else None, |
| 44 | + latent_shape=sample.latent_shape, |
| 45 | + video_metadata=sample.video_metadata, |
| 46 | + ) |
| 47 | + else: |
| 48 | + # For multiple samples, just return a simple dict |
| 49 | + return {"samples": samples} |
| 50 | + |
| 51 | + |
| 52 | +def create_diffusion_sample(key: str, seq_len: int, video_shape=(16, 8), embedding_dim=128) -> DiffusionSample: |
| 53 | + """Helper function to create a DiffusionSample for testing.""" |
| 54 | + return DiffusionSample( |
| 55 | + __key__=key, |
| 56 | + __restore_key__=(), |
| 57 | + __subflavor__=None, |
| 58 | + __subflavors__=["default"], |
| 59 | + video=torch.randn(seq_len, video_shape[0]), |
| 60 | + context_embeddings=torch.randn(10, embedding_dim), |
| 61 | + context_mask=torch.ones(10), |
| 62 | + loss_mask=torch.ones(seq_len), |
| 63 | + seq_len_q=torch.tensor([seq_len], dtype=torch.int32), |
| 64 | + seq_len_q_padded=torch.tensor([seq_len], dtype=torch.int32), |
| 65 | + seq_len_kv=torch.tensor([10], dtype=torch.int32), |
| 66 | + seq_len_kv_padded=torch.tensor([10], dtype=torch.int32), |
| 67 | + pos_ids=torch.arange(seq_len).unsqueeze(1), |
| 68 | + latent_shape=torch.tensor([4, 2, 4, 4], dtype=torch.int32), |
| 69 | + video_metadata={"fps": 30, "resolution": "512x512"}, |
| 70 | + ) |
| 71 | + |
| 72 | + |
| 73 | +def test_select_samples_to_pack(): |
| 74 | + """Test select_samples_to_pack method.""" |
| 75 | + # Create encoder with seq_length=20 |
| 76 | + encoder = ConcreteDiffusionTaskEncoder(seq_length=20) |
| 77 | + |
| 78 | + # Create samples with different sequence lengths |
| 79 | + samples = [ |
| 80 | + create_diffusion_sample("sample_1", seq_len=8), |
| 81 | + create_diffusion_sample("sample_2", seq_len=12), |
| 82 | + create_diffusion_sample("sample_3", seq_len=5), |
| 83 | + create_diffusion_sample("sample_4", seq_len=7), |
| 84 | + create_diffusion_sample("sample_5", seq_len=3), |
| 85 | + ] |
| 86 | + |
| 87 | + # Call select_samples_to_pack |
| 88 | + result = encoder.select_samples_to_pack(samples) |
| 89 | + |
| 90 | + # Verify result is a list of lists |
| 91 | + assert isinstance(result, list), "Result should be a list" |
| 92 | + assert all(isinstance(group, list) for group in result), "All elements should be lists" |
| 93 | + |
| 94 | + # Verify all samples are included |
| 95 | + all_samples = [sample for group in result for sample in group] |
| 96 | + assert len(all_samples) == len(samples), "All samples should be included" |
| 97 | + |
| 98 | + # Verify no bin exceeds seq_length |
| 99 | + for group in result: |
| 100 | + total_seq_len = sum(sample.seq_len_q.item() for sample in group) |
| 101 | + assert total_seq_len <= encoder.seq_length, ( |
| 102 | + f"Bin with total {total_seq_len} exceeds seq_length {encoder.seq_length}" |
| 103 | + ) |
| 104 | + |
| 105 | + # Verify that bins are non-empty |
| 106 | + assert all(len(group) > 0 for group in result), "No bin should be empty" |
| 107 | + |
| 108 | + print(f"✓ Successfully packed {len(samples)} samples into {len(result)} bins") |
| 109 | + print(f" Bin sizes: {[sum(s.seq_len_q.item() for s in group) for group in result]}") |
| 110 | + |
| 111 | + |
| 112 | +def test_pack_selected_samples(): |
| 113 | + """Test pack_selected_samples method.""" |
| 114 | + encoder = ConcreteDiffusionTaskEncoder(seq_length=100) |
| 115 | + |
| 116 | + # Create multiple samples to pack |
| 117 | + sample_1_length = 10 |
| 118 | + sample_2_length = 15 |
| 119 | + sample_3_length = 8 |
| 120 | + sample_1 = create_diffusion_sample("sample_1", seq_len=sample_1_length) |
| 121 | + sample_2 = create_diffusion_sample("sample_2", seq_len=sample_2_length) |
| 122 | + sample_3 = create_diffusion_sample("sample_3", seq_len=sample_3_length) |
| 123 | + |
| 124 | + samples_to_pack = [sample_1, sample_2, sample_3] |
| 125 | + |
| 126 | + # Pack the samples |
| 127 | + packed_sample = encoder.pack_selected_samples(samples_to_pack) |
| 128 | + |
| 129 | + # Verify the packed sample is a DiffusionSample |
| 130 | + assert isinstance(packed_sample, DiffusionSample), "Result should be a DiffusionSample" |
| 131 | + |
| 132 | + # Verify __key__ is concatenated |
| 133 | + expected_key = "sample_1,sample_2,sample_3" |
| 134 | + assert packed_sample.__key__ == expected_key, f"Key should be '{expected_key}'" |
| 135 | + |
| 136 | + # Verify video is concatenated along dim 0 |
| 137 | + expected_video_len = 10 + 15 + 8 |
| 138 | + assert packed_sample.video.shape[0] == expected_video_len, f"Video should have length {expected_video_len}" |
| 139 | + |
| 140 | + # Verify context_embeddings is concatenated |
| 141 | + expected_context_len = 10 * 3 # 3 samples with 10 embeddings each |
| 142 | + assert packed_sample.context_embeddings.shape[0] == expected_context_len, ( |
| 143 | + f"Context embeddings should have length {expected_context_len}" |
| 144 | + ) |
| 145 | + |
| 146 | + # Verify context_mask is concatenated |
| 147 | + assert packed_sample.context_mask.shape[0] == expected_context_len, ( |
| 148 | + f"Context mask should have length {expected_context_len}" |
| 149 | + ) |
| 150 | + |
| 151 | + # Verify loss_mask is concatenated |
| 152 | + assert packed_sample.loss_mask.shape[0] == expected_video_len, f"Loss mask should have length {expected_video_len}" |
| 153 | + |
| 154 | + # Verify seq_len_q is concatenated |
| 155 | + assert packed_sample.seq_len_q.shape[0] == 3, "seq_len_q should have 3 elements" |
| 156 | + assert torch.equal( |
| 157 | + packed_sample.seq_len_q, torch.tensor([sample_1_length, sample_2_length, sample_3_length], dtype=torch.int32) |
| 158 | + ), "seq_len_q values incorrect" |
| 159 | + |
| 160 | + assert packed_sample.seq_len_q_padded.shape[0] == 3, "seq_len_q_padded should have 3 elements" |
| 161 | + assert torch.equal( |
| 162 | + packed_sample.seq_len_q_padded, |
| 163 | + torch.tensor([sample_1_length, sample_2_length, sample_3_length], dtype=torch.int32), |
| 164 | + ), "seq_len_q_padded values incorrect" |
| 165 | + |
| 166 | + assert packed_sample.seq_len_kv.shape[0] == 3, "seq_len_kv should have 3 elements" |
| 167 | + assert torch.equal(packed_sample.seq_len_kv, torch.tensor([10, 10, 10], dtype=torch.int32)), ( |
| 168 | + "seq_len_kv values incorrect" |
| 169 | + ) |
| 170 | + |
| 171 | + assert packed_sample.seq_len_kv_padded.shape[0] == 3, "seq_len_kv_padded should have 3 elements" |
| 172 | + assert torch.equal(packed_sample.seq_len_kv_padded, torch.tensor([10, 10, 10], dtype=torch.int32)), ( |
| 173 | + "seq_len_kv_padded values incorrect" |
| 174 | + ) |
| 175 | + |
| 176 | + assert packed_sample.latent_shape.shape[0] == 3, "latent_shape should have 3 rows" |
| 177 | + assert isinstance(packed_sample.video_metadata, list), "video_metadata should be a list" |
| 178 | + assert len(packed_sample.video_metadata) == 3, "video_metadata should have 3 elements" |
| 179 | + |
| 180 | + print(f"✓ Successfully packed {len(samples_to_pack)} samples") |
| 181 | + print(f" Packed video shape: {packed_sample.video.shape}") |
| 182 | + print(f" Packed context embeddings shape: {packed_sample.context_embeddings.shape}") |
0 commit comments