3 Commits
Author SHA1 Message Date
376e445f0e [feat]: FastH3 OmniRef inference on consumer devices (RTX 5090 / PRO 6000 / DGX Spark / 4090) (#1980)
Co-authored-by: Aryan Kumar <aryan5v@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 14:57:36 -07:00
Davide Locatelli e1e2559359 [perf] FastH3 OmniRef: lossless speed-ups behind default-off switches (#1930)
These speed-ups for the MiniMax-H3 pipelines keep the output bit-identical. Each
one has a switch that is off by default:

- FASTVIDEO_MINIMAX_H3_EXACT_KERNELS: Triton RoPE, AdaLN modulation, gated
  residual and SwiGLU kernels that round to bf16 at the same points as the
  eager ops.
- FASTVIDEO_H3_VSA_HEADS_FIRST_TILE: VSA-H3 tiles are written directly in the
  heads-first layout of the kernel.
- FASTVIDEO_H3_VAE_TILE_PARALLEL: VAE spatial tiles are split across the
  sequence-parallel ranks.
- FASTVIDEO_H3_REF2VA_MEMO_ENTRIES: repeat reference encodes are reused.

The example basic_fasth3_omniref_pdd.py --lossless-accel turns them all on.

On GB200 with 4 GPUs, MiniMax-H3 text to video with audio (124 frames,
1344x768) gives bit-identical frames and audio with the first three switches
on and off.
2026-10-07 15:35:37 -07:00
8444c0897a [feat] FastH3 Ref2VA PDD inference with reference-video VSA sparsity (#1907)
Co-authored-by: Davids048 <jundasu@ucsd.edu>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-02 11:28:07 -07:00