| 1 | """ |
| 2 | Gemma Text Encoder Wrapper for DMD distillation. |
| 3 | |
| 4 | Provides a simple interface for text encoding without prompt enhancement. |
| 5 | Just pure text -> context embedding conversion. |
| 6 | """ |
| 7 | |
| 8 | from typing import List, Dict, Any, Optional |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | |
| 12 | from ltx_core.loader.registry import Registry |
| 13 | |
| 14 | |
| 15 | class GemmaTextEncoderWrapper(nn.Module): |
| 16 | """ |
| 17 | Wrapper for Gemma text encoder to provide DMD-compatible interface. |
| 18 | |
| 19 | This wrapper: |
| 20 | - Takes raw text prompts (no enhancement needed) |
| 21 | - Returns conditional_dict with video_context and audio_context |
| 22 | - Handles batched encoding |
| 23 | """ |
| 24 | |
| 25 | def __init__( |
| 26 | self, |
| 27 | text_encoder, |
| 28 | embeddings_processor, |
| 29 | device: torch.device = None, |
| 30 | dtype: torch.dtype = torch.bfloat16, |
| 31 | ): |
| 32 | """ |
| 33 | Args: |
| 34 | text_encoder: GemmaTextEncoder instance |
| 35 | embeddings_processor: EmbeddingsProcessor instance |
| 36 | device: Target device |
| 37 | dtype: Model dtype |
| 38 | """ |
| 39 | super().__init__() |
| 40 | self.text_encoder = text_encoder |
| 41 | self.embeddings_processor = embeddings_processor |
| 42 | self.device = device |
| 43 | self.dtype = dtype |
| 44 | |
| 45 | @torch.no_grad() |
| 46 | def forward( |
| 47 | self, |
| 48 | text_prompts: List[str], |
| 49 | padding_side: str = "left", |
| 50 | ) -> Dict[str, Optional[torch.Tensor]]: |
| 51 | """ |
| 52 | Encode text prompts to conditioning embeddings. |
| 53 | |
| 54 | Args: |
| 55 | text_prompts: List of text prompts (already processed, no enhancement) |
| 56 | padding_side: Padding side for tokenizer |
| 57 | |
| 58 | Returns: |
| 59 | Dictionary containing: |
| 60 | - video_context: [B, seq_len, dim] video conditioning |
| 61 | - audio_context: [B, seq_len, dim] audio conditioning |
| 62 | - attention_mask: [B, seq_len] attention mask |
| 63 | """ |
| 64 | batch_size = len(text_prompts) |
| 65 | |
| 66 | # Encode each prompt |
| 67 | video_contexts = [] |
| 68 | audio_contexts = [] |
| 69 | attention_masks = [] |
| 70 | |
| 71 | for prompt in text_prompts: |
| 72 | # 1) Run Gemma LLM to get raw hidden states + attention mask |
| 73 | hidden_states, attn_mask = self.text_encoder.encode(prompt, padding_side=padding_side) |
| 74 | # 2) Process hidden states to obtain final embeddings |
| 75 | output = self.embeddings_processor.process_hidden_states( |
| 76 | hidden_states, attn_mask, padding_side=padding_side |
| 77 | ) |
| 78 | |
| 79 | video_contexts.append(output.video_encoding) |
| 80 | audio_contexts.append(output.audio_encoding) |
| 81 | attention_masks.append(output.attention_mask) |
| 82 | |
| 83 | # Stack batch |
| 84 | video_context = torch.cat(video_contexts, dim=0) if len(video_contexts) > 0 else None |
| 85 | # Handle optional audio connector (may be None depending on config) |
| 86 | if any(ac is None for ac in audio_contexts): |
| 87 | audio_context = None |
| 88 | else: |
| 89 | audio_context = torch.cat(audio_contexts, dim=0) |
| 90 | attention_mask = torch.cat(attention_masks, dim=0) if len(attention_masks) > 0 else None |
| 91 | |
| 92 | return { |
| 93 | "video_context": video_context, |
| 94 | "audio_context": audio_context, |
| 95 | "attention_mask": attention_mask, |
| 96 | } |
| 97 | |
| 98 | def encode_batch( |
| 99 | self, |
| 100 | text_prompts: List[str], |
| 101 | ) -> Dict[str, torch.Tensor]: |
| 102 | """Alias for forward() with default padding.""" |
| 103 | return self.forward(text_prompts) |
| 104 | |
| 105 | |
| 106 | def create_text_encoder_wrapper( |
| 107 | checkpoint_path: str, |
| 108 | gemma_path: str, |
| 109 | device: torch.device, |
| 110 | dtype: torch.dtype = torch.bfloat16, |
| 111 | registry: Registry | None = None, |
| 112 | ) -> GemmaTextEncoderWrapper: |
| 113 | """ |
| 114 | Factory function to create GemmaTextEncoderWrapper from checkpoint. |
| 115 | |
| 116 | Args: |
| 117 | checkpoint_path: Path to LTX-2 checkpoint |
| 118 | gemma_path: Path to Gemma text encoder |
| 119 | device: Target device |
| 120 | dtype: Model dtype |
| 121 | |
| 122 | Returns: |
| 123 | Configured GemmaTextEncoderWrapper |
| 124 | """ |
| 125 | from ltx_pipelines.utils.model_ledger import ModelLedger |
| 126 | |
| 127 | # Load to CPU first to avoid safetensors device issues |
| 128 | ledger = ModelLedger( |
| 129 | dtype=dtype, |
| 130 | device=torch.device("cpu"), |
| 131 | checkpoint_path=checkpoint_path, |
| 132 | gemma_root_path=gemma_path, |
| 133 | registry=registry, |
| 134 | ) |
| 135 | |
| 136 | text_encoder = ledger.text_encoder().to(device=device, dtype=dtype) |
| 137 | embeddings_processor = ledger.gemma_embeddings_processor().to(device=device, dtype=dtype) |
| 138 | |
| 139 | wrapper = GemmaTextEncoderWrapper( |
| 140 | text_encoder=text_encoder, |
| 141 | embeddings_processor=embeddings_processor, |
| 142 | device=device, |
| 143 | dtype=dtype, |
| 144 | ) |
| 145 | |
| 146 | return wrapper |
| 147 |