import logging from .utils import ( get_raw_tokenizer, map_token_indices, build_segments, create_mask_fn, distribute_segment_lengths, ) from .patches import detect_model_type, apply_patches log = logging.getLogger(__name__) def _convert_to_latent_lengths(pixel_lengths, temporal_stride, latent_frames): """Convert pixel-space segment lengths to integer latent-space lengths using the largest-remainder method. Targets the full `latent_frames` when the pixel sum looks like full coverage (within one stride of latent_frames * stride). Otherwise targets round(total_pixel / temporal_stride) so partial-coverage timelines stay partial. """ if not pixel_lengths: return [] total_pixel = sum(pixel_lengths) if total_pixel <= 0: return [1] * len(pixel_lengths) naive_total = max(1, round(total_pixel / temporal_stride)) target_total = min(latent_frames, naive_total) # Within one frame of full → user clearly intended full coverage; pin to latent_frames. if target_total >= latent_frames - 1: target_total = latent_frames exact = [p * target_total / total_pixel for p in pixel_lengths] result = [int(e) for e in exact] diff = target_total - sum(result) if diff > 0: order = sorted(range(len(exact)), key=lambda i: -(exact[i] - int(exact[i]))) for k in range(diff): result[order[k % len(order)]] += 1 # Ensure every segment has ≥ 1 latent frame (steal from the largest if needed). for i in range(len(result)): if result[i] < 1: max_idx = max(range(len(result)), key=lambda j: result[j]) if result[max_idx] > 1: result[max_idx] -= 1 result[i] = 1 return result def _encode_relay(model, clip, length, height, width, global_prompt, local_prompts, segment_lengths, epsilon, relay_options=None): locals_list = [p.strip() for p in local_prompts.split("|") if p.strip()] if not locals_list: raise ValueError("At least one local prompt is required (separate with |)") arch, patch_size, temporal_stride = detect_model_type(model) latent_frames = ((length - 1) // temporal_stride) + 1 height = height // 64 width = width // 64 tokens_per_frame = (height // patch_size[1]) * (width // patch_size[2]) parsed_lengths = None if segment_lengths.strip(): pixel_lengths = [int(x.strip()) for x in segment_lengths.split(",") if x.strip()] parsed_lengths = _convert_to_latent_lengths(pixel_lengths, temporal_stride, latent_frames) raw_tokenizer = get_raw_tokenizer(clip) full_prompt, token_ranges = map_token_indices(raw_tokenizer, global_prompt, locals_list) log.info("[PromptRelay] Global: tokens [0:%d] (%d tokens)", token_ranges[0][0], token_ranges[0][0]) for i, (s, e) in enumerate(token_ranges): log.info("[PromptRelay] Segment %d: tokens [%d:%d] (%d tokens)", i, s, e, e - s) conditioning = clip.encode_from_tokens_scheduled(clip.tokenize(full_prompt)) effective_lengths = distribute_segment_lengths(len(locals_list), latent_frames, parsed_lengths) log.info( "[PromptRelay] Latent: %d frames, %d tokens/frame, segments: %s", latent_frames, tokens_per_frame, effective_lengths, ) q_token_idx = build_segments(token_ranges, effective_lengths, epsilon, relay_options) mask_fn = create_mask_fn(q_token_idx, tokens_per_frame, latent_frames) patched = model.clone() apply_patches(patched, arch, mask_fn) return patched, conditioning