On GB10 the light H3 VAE's INT8 ConvRot overlay decodes 832x480x124 in
42 s and gains nothing from tile batching. The dense weights take 24 s
with one tile per decoder call, 14 s with 12 tiles per call and 9.8 s
compiled, at higher fidelity (77.7 vs 63.4 dB PSNR against an FP32
decode). FASTVIDEO_H3_VAE_INT8_OVERLAY=0 now skips the overlay (default
on, so other recipes are unchanged), and the single-Spark recipes use it
with FASTVIDEO_H3_VAE_TILE_BATCH=12 and a compiled VAE decoder.
FlashInfer's CUTLASS mm_fp4 is 2.7-3x slower than its cuDNN backend for
the H3 FFN projections on GB10 with identical outputs, so the recipes
select FASTVIDEO_NVFP4_MM_BACKEND=cudnn. FlashInfer autotuning was not
adopted: it made repeated calls non-deterministic.
One Spark, 832x480x124, seed 1234, warm calls: FastH3 Trim 121-122 s to
68 s (denoise 78 -> 58 s, decode 43 -> 9.5 s); FastH3 V2 139-141 s to
79 s. Repeated calls are bit-identical, which the previous V2 recipe was
not.
FP4 attention (FASTVIDEO_H3_VSA_FP4=1, built for sm_121a since #1598) is
documented for GB10 as opt-in and stays off: it is 24-33% faster in
denoise on V2 but changes the output (LPIPS 0.39-0.55 against BF16
attention), and the checkpoints were not trained for it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.
The distributed_setup test fixture now sets MASTER_ADDR and MASTER_PORT when
they are not set, with a new port for each test, and restores them after the
test. pyproject.toml now declares orjson and python-multipart, which the
OpenAI-compatible server needs. Two documents about the SSIM reference
bootstrap are corrected.
For Wan VSA inference with 64-token tiles, each attention layer scattered Q, K,
V and the gate into a tile buffer and then made four transposed copies. On
SM100, one Triton kernel (tile_to_bhsd) now does the scatter and the transposes
in one pass. Training, 256-token tiles, other GPUs and builds without the kernel
use the previous path. FASTVIDEO_DISABLE_VSA64_FUSED_LAYOUT=1 turns the fused
path off.
On GB200 the layout and attention step goes from 15.7 ms to 8.6 ms. FastWan 1.3B
frames are bit-exact with the previous path at sequence parallel 1 and 2.
Adds preprocess_hunyuan15_overfit.py, which writes VAE latents plus Qwen and ByT5 text embeddings in the dual-text parquet schema (pyarrow_schema_t2v_dual_text) that main's HunyuanVideo 1.5 training (#1662) reads; main's guide and overfit config pointed at this script, but nothing in-tree produced text_embedding_2 before.
Reconciled with #1662 during maintainer review (Swarm orders #723 and #777): main's schema and collate are the base. On top of them, both text streams now share one CFG-dropout decision per row also when _sample_index is absent, and a row missing the primary text_embedding raises instead of training unconditioned. The PR's widening of the shared t2v schema and its writer changes were dropped. docs/training/hunyuan15_overfit.md now describes the in-tree script.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
