Make a simple RGB gradient tensor in [0, 255] matching build_demo_image.

Accept explicit episode lists or simple range strings (e.g. "0-128").

Accept explicit episode lists or simple range strings (e.g. "0-128").

Optimized bilinear pos embedding interpolation on CUDA (no host sync).

GPU-friendly rotary position builder without .item()/.tolist() syncs.

Resolve an hf: checkpoint path to a local Hugging Face cache path. Format: hf:org/repo[/optional/subpath] or hf:org/repo@revision[/subpath]

Build an episode list by matching split names against provided regex patterns.

Build an episode list by matching split names against provided regex patterns.

Returns torch.sqrt(torch.max(0, x)) but with a zero sub-gradient where x is 0. Reference: https://github.com/facebookresearch/pytorch3d/b

Optional shard hook for policies that support FSDP.

Returns a boolean mask (B, T) that is True on assistant *content* tokens and False elsewhere. This parses ChatML boundaries directly

Construct a Beta distribution on the training device.

Create a deterministic RGB gradient so the script works without external assets.

Create optimizer parameter groups from a learning-rate mapping.

Compute attention weights manually to get the attention scores. Args: q: Query tensor [batch, num_heads, seq_len, head_dim]

Convert image tensor from LeRobotDataset to PIL Image. Args: img_tensor: Image tensor, shape (C, H, W) or (H, W, C), dtype float32 [

Create a new floating-point tensor with the magnitude of input and the sign of other, element-wise. Note: The implementation follows from

Manual autoregressive decoding loop with explicit KV cache updates.

Encode the observation prefix and return KV cache artifacts.

Extract attention weights from the VLM bridge. Returns: attention_weights: [batch, num_heads, seq_len, seq_len] image_mask:

Extract object name from instruction text. Supports formats like: - "put {object} on {location}" -> returns "{object}" - "put {objec

Generate bbox prediction using prefill and decode.

No-op action used to let the environment settle.

Instantiate a transform object from a spec. Accepts either an existing object with a `.compute()` method, or a dict with a `_target_` key and

Distributed-aware checkpoint load for directory checkpoints. `checkpoint` is a directory with `model.pt` and `optimizer.pt`. Returns (missing

Print local-shard sizes (MB) for params, buffers, grads, and optimizer state.

Map attention weights from token space to image space. Args: attention_weights: [batch, num_heads, seq_len, seq_len] - attention wei

Parse bbox JSON from model output. The output format is typically: [{"bbox_2d": [...], "label": "..."}, ...]<|im_end|> Args: bbo

Prefill pass that returns cache, cache_position, attention_mask, and next token.

Apply a quaternion rotation to a vector. Args: quat: The quaternion in (w, x, y, z). Shape is (..., 4). vec: The vector in (x, y,

Convert rotations given as rotation matrices to quaternions. Args: matrix: The rotation matrices. Shape is (..., 3, 3). Returns:

Convert a unit quaternion to a standard form where the real part is non-negative. Quaternion representations have a singularity since ``q`` and `

Convert 6D rotation representation back to a proper rotation matrix. Uses Gram-Schmidt orthogonalization on the two 3D row vectors as in http

Sample diffusion timesteps with a small clamp to avoid numerical issues.

Save a structured config to cfg.yaml in run_dir.

Save checkpoint as a directory with model.pt and optimizer.pt. `filename` should be the checkpoint directory name, e.g., `checkpoint_5`. If a

Normalizes a given input tensor to a range of [-1, 1]. .. note:: It uses pytorch broadcasting functionality to deal with batched input.

Initialize dist process group using env:// init and optionally build a device mesh.

Visualize attention map overlaid on the original image. Args: img_chw: Original image [3, H, W] in uint8 attention_map: Atte

Visualize bounding boxes on image and save. Args: image: PIL Image bboxes: List of dicts with keys "bbox_2d" and "label"

Block until a request is available, or timeout. Returns a dict with the request data, or None if stopped/timeout.