Load VAE based on vae_type configuration. Supported types: - "tensorrt": TensorRT-accelerated VAE (requires pre-built engines)

Validate a requested sub-region against server MR and return registration metadata. This server uses one pre-registered MR, so sub-regions ar

Apply RoPE using explicit frame indices (for memory-aware attention). Rather than assuming sequential frame positions 0..F-1, this uses the p

Args: index: [0, 1, 2, 3] 1D Tensor representing the position index of the token

Args: index: [0, 1, 2, 3] 1D Tensor representing the position index of the token

Args: index: [0, 1, 2, 3] 1D Tensor representing the position index of the token

Scaled dot-product attention Args: Q: Query tensor [batch_size, num_heads, seq_len, d_k] K: Key tensor [batch_size, num_head

Fuse Wan AdaLN modulation with MXFP8 quantization. This is a narrow Wan transformer helper, not a generic modulate op. It expects a 2D BF16 a

Upsample video latent using the upsampler with proper normalization. Aligned with ltx_core.model.upsampler.model.upsample_video.

Create cumulative sequence lengths tensor for attention. Args: seq_len: Sequence length Returns: Cu

通用的模型过滤和格式化函数

Core forward implementation for DPT head. Args: token_list: List of transformer tokens from each layer, [B, S, N, C]

Get adaptive values from scale-shift table (no batch dimension).

Get keypoint predictions from score maps. Note: batch_size: N num_keypoints: K heatmap height: H heatmap width: W

Standard text branch post-attention processing.

将输入张量从 [seq_len/N, heads, hidden_dims] 转换为 [seq_len, heads/N, hidden_dims] 的格式。 参数: input (torch.Tensor): 输入张量，形状为 [seq_len/N, heads

Create a normalization layer based on the normalization type. Args: in_channels: Number of input channels num_groups: Number

Consume one config from ring buffer and advance head by rdma_faa.

Create pinned memory tensor buffers on CPU for lazy loading scenario. Args: base_attrs: [(name, attr_name, transpose), ...] Configuration

Create tensor buffers and move them to CUDA device (specified by AI_DEVICE). Args: base_attrs: [(name, attr_name, transpose), ...] List o