float tensor (B, L, D) -> Tuple of locs, scales, and weights

Pre-sampled latents T.Tensor (B, L, Z) -> float tensor (B, L, D)

Inverse of `codes_to_indices`.

Send temperature settings to server

Bound `z`, an array of shape (..., d).

Converts a `code` to an index in the codebook.

cache: (batch, seq_len, 2, kv_heads, head_dim)

Converts indices to indices at each level, perhaps needed for a transformer with factorized embeddings

Quantizes z, returns quantized zhat, same shape as z.

Args ---- q : torch.Tensor Embedded query tensor, expected size is B x S x H x Eh k : torch.Tensor

Round with straight through gradients.

Send temperature settings to server