`maskmem_pos_enc` is the same across frames and objects, so we cache it as a constant in the inference session to reduce session stor

确保 pipe 已创建，且每个 rank 绑定自己的 GPU。

Map client-side object id to model-side object index.

Positionally encode points that are normalized to [0,1].

Add new points to a frame.

Calculates boxes in XYXY format around masks. Return [0,0,0,0] for an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x..

Computes the stability score for a batch of masks. The stability score is the IoU between the binary masks obtained by thresholding the p

Count data rows (exclude header).

JSON entry -> vid (stem, no .mp4). Supports: - {"vid": "..."} - {"video": "..."} / {"video_path": "..."} / {"name": "..."} / {"fi

imgs: [B, 3, H, W] of torch.float32. - mean: [0.48145466, 0.4578275, 0.40821073] - std: [0.26862954, 0.2613025

Forward pass through the upsampling block. Args: x (torch.Tensor): Input tensor feat_cache (list, optional):

Positionally encode points that are not normalized to [0,1].

this will perform the index select, cat the tensors, and provide the attn_bias from cache

Returns the positional encoding used to encode point prompts, applied to a dense set of points the shape of the image encoding.

Get the normalization layer. Args: norm_layer (str): The type of normalization layer. Returns: norm_layer (nn.Module):

Generates position ids for the input embeddings grid. modified from the qwen2_vl mrope.

Initialize an inference state.

Load a PNG file as a mask and its palette.

Encodes masks to an uncompressed RLE, in the format expected by pycoco tools.