FlashRT Spatiotemporal Layout

This package provides FlashRT spatiotemporal layout helpers for Hugging Face Kernel Hub.

It is intended for VLA, video, diffusion, and world-model pipelines that move between latent (B, C, T, H, W) tensors, token matrices, temporal unshuffle layouts, and short latent caches.

Kernels

• ncdhw_to_blc_bf16: convert NCDHW BF16 latents to BLC token matrices.
• patch_im2col_bf16: materialize patch rows for transformer input.
• time_unshuffle2_bf16: convert (B, 2C, T, H, W) to (B, C, 2T, H, W).
• add_bias_ncdhw_bf16: in-place BF16 channel-bias add for NCDHW latents.
• update_cache2_ncdhw_bf16: maintain a two-frame NCDHW latent cache.
• channel_to_space3d_bf16: decode channel-packed spatial/temporal output.
• pack_causal_cache3_nhwc_bf16: pack previous/current causal cache rows.
• avg_pool3d_channels_bf16: grouped spatiotemporal downsample.
• ndhwc_to_ncdhw_bf16, ndhwc_to_ncdhw_bias_bf16, ndhwc_to_ncdhw_add_bf16: BF16 layout producer variants.
• ncdhw_quantize_fp8_static_ndhwc_bf16 and upsample2x_quantize_fp8_static_nhwc_bf16: fused layout-to-FP8 producers.

Hardware

• CUDA 12.8+
• BF16-capable NVIDIA GPUs

Current local source validation is on RTX 5090. Broader hardware rows should be added after installed-artifact validation.

Upstream

The serving source of truth remains FlashRT:

https://github.com/LiangSu8899/FlashRT