Expand description
Finite Scalar Quantization (FSQ) — the codebook-free discrete token structure from
“Finite Scalar Quantization: VQ-VAE Made Simple” (Mentzer et al., 2023). Each latent
dimension is independently rounded to one of L evenly-spaced levels spanning a bounded
range [lo, hi]; the implicit codebook is the Cartesian product of the per-dim level grids,
so no codebook needs to be stored or learned. This is pure quantizer math — real and
buildable now. The neural encoder that produces the (bounded) latents is a separate,
NeedsWeights concern.
Per-dimension level counts may differ (e.g. [8, 5, 5, 5]), matching the FSQ convention.
On-grid values (lo + k·step) round-trip exactly; off-grid values snap to the nearest level.
Zero-heap: caller supplies the level table and all buffers.
Functions§
- fsq_
dequantize - Reconstruct latent components from FSQ level indices:
out[d] = lo + tokens[d]·step_d. - fsq_
quantize - Quantize each component of
inputto its per-dimension level grid, writing level indices.