HyperFlow H3 Flash: 31B Video+Audio Diffuzion Modelini 14B Gacha Kompressiya Qilindi

Video Rebirth kompaniyasi 31 milliard parametrli video va audio diffuzion modelini 14 milliard parametrga qisqartirib, 6 qadamda ishlaydigan HyperFlow H3 Flash ni taqdim etdi. Bu yangi model, MiniMax-H3 ning kompressiya qilingan versiyasi, video va sinxronlashtirilgan stereo audio ishlab chiqarish uchun mo'ljallangan.

HyperFlow H3 Flash: Asosiy Xususiyatlari

HyperFlow H3 Flash, 31 milliard parametrli MiniMax-H3 modelini 14 milliard parametrga qisqartirib, tezroq va samarali ishlash imkonini beradi. U 5.2 dan 15 sekundgacha bo'lgan videolarni 24 fps tezlikda ishlab chiqaradi. Video o'lchami 1344x768 bo'lsa-da, foydalanuvchi o'zgartirishlari mumkin. Audio trek 32 kHz stereo formatida taqdim etiladi.

pip install git+https://github.com/Video-Rebirth/hyperflow-h3-flash
pip install "hyperflow-h3-flash[hopper] @ git+https://github.com/Video-Rebirth/hyperflow-h3-flash"  # H100 / H200: FlashAttention-3
pip install "hyperflow-h3-flash[fp8] @ git+https://github.com/Video-Rebirth/hyperflow-h3-flash"     # --text-encoder-precision fp8

Ishlash Tezligi va Samaradorligi

HyperFlow H3 Flash ning ishlash tezligi turli grafik karta (GPU) turlariga qarab farqlanadi. Masalan, 5.2 sekundlik video klipni 6 qadamda ishlab chiqarish uchun 4x B200 GPU da 8.7 sekund, 1x H200 GPU da 36 sekund, va 1x RTX 5090 GPU da 92 sekund vaqt ketadi. NVIDIA SOL-H3 ning 4x B-seriya GPU larida esa real vaqt rejimida ishlaydi.

# Text to video + audio (t2va)
hyperflow-h3-flash --prompt "A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window." \
    --output ramen.mp4

# First keyframe, or first + last keyframe (fl2va)
hyperflow-h3-flash --prompt "..." --image first.png --output clip.mp4
hyperflow-h3-flash --prompt "..." --image first.png --last-image last.png --output clip.mp4

# Reference images, videos (with soundtrack), and audio clips in prompt order (ref2va)
hyperflow-h3-flash --prompt "<Picture 1> walks through the scene of <Picture 2> ..." \
    --reference person.png --reference scene.png --output clip.mp4

# 15 s clip, context-parallel across 4 GPUs (or --gpus auto for up to 4 visible GPUs)
hyperflow-h3-flash --prompt "..." --num-frames 360 --gpus 4

# 4 steps (speed), 6 steps (default), or 8 steps (quality)
hyperflow-h3-flash --prompt "..." --steps 4

Modelning o'lchami 13.7 milliard parametr (fl2va) yoki 13.9 milliard parametr (ref2va) bo'lib, bu MiniMax-H3 ning yarmini tashkil qiladi. Modelning eng kichik o'lchami Q4_K_M 9.1 GB (Q8_0 14.9 GB) bo'lib, 24 GB va 32 GB GPU larida ishlaydi. 16 GB GPU larida ishlash imkoniyati hali o'rganilmagan.

import torch
from diffusers.utils.export_utils import encode_video
from hyperflow_h3_flash import load_pipeline

pipe = load_pipeline("t2va")  # "fl2va": pass image=... (and last_image=...); "ref2va": pass references=[...]
out = pipe(
    prompt="A chef slides a steaming bowl of ramen across a wooden counter.",
    num_frames=124,
    num_inference_steps=6,  # 4, 6, or 8
    generator=torch.Generator().manual_seed(0),
    output=["videos", "audio", "sampling_rate"],
)
encode_video(
    out["videos"][0],
    fps=24,
    output_path="ramen.mp4",
    audio=out["audio"][0],
    audio_sample_rate=out["sampling_rate"],
)

Ishlash Usullari va Qo'llab-Quvvatlanadigan Platformalar

HyperFlow H3 Flash turli platformalarda ishlash uchun mo'ljallangan. Masalan, GGUF PC larida 64 GB dan kam RAM bo'lganida ishlatiladi. INT8 convrot esa data-center GPU larida yoki 24 GB dan ortiq RAM bo'lgan RTX GPU larida ishlatiladi. Barcha vaznlar VideoRebirthAI/hyperflow-h3-flash da mavjud.

make install   # create .venv and install editable package with dev tools
make test      # run unit tests on tiny models (CUDA tests skip without a GPU)
make lint      # ruff check + format check

Model turli vaznlarni qo'llab-quvvatlaydi, jumladan Q3_K_M (7.2 GB) dan Q8_0 (15.2 GB) gacha. Ushbu vaznlar turli GPU va RAM konfiguratsiyalariga moslashish uchun mo'ljallangan.

.
├── src/hyperflow_h3_flash/        # Python package (INT8 runtime)
│   ├── cli.py                     # `hyperflow-h3-flash` CLI entry point
│   ├── __main__.py                # `python -m hyperflow_h3_flash` entry point
│   ├── hostmem.py                 # Host RAM estimate before loading
│   ├── pipeline.py                # `load_pipeline`: loads models, trims unused text-encoder layers, configures offload/multi-GPU
│   ├── blocks.py                  # Modular Pipeline blocks swapped into the official MiniMax-H3 workflows
│   ├── transformer.py             # `HyperFlowH3FlashTransformer3DModel`: dropped blocks, low-rank linears, two-time (t, r) conditioning
│   ├── int8.py                    # Hadamard-rotated W8A8 linear layers (Triton, cuBLASLt, CUTLASS, and B300 BF16 kernels)
│   ├── schedule.py                # 4-, 6-, and 8-step sigma grids and per-token (t, r) pairs
│   ├── offload.py                 # Single-GPU memory placement (`none`, `model`, `stream`)
│   └── distributed.py             # Multi-GPU execution (torchrun launch, Ulysses attention, FSDP2 / rank-0 text encoder, parallel VAE)
├── integrations/                  # GGUF runtimes for consumer GPUs
│   ├── README.md                  # Step-by-step setup guide for ComfyUI, Unsloth Studio, and stable-diffusion.cpp
│   ├── ComfyUI-HyperFlow-Flash/   # Custom nodes (`loader_nodes.py`, `model_patch.py`, `gguf_io.py`; GPL-3.0)
│   │   └── example_workflows/     # Four UI workflows (`hyperflow_t2v`, `hyperflow_i2v`, `hyperflow_flf2v`, `hyperflow_ref2v`) and API versions in `api/`
│   └── stable-diffusion.cpp/      # `hyperflow-two-time.patch` for upstream commit `3be2378` (MIT)
├── examples/                      # `benchmark.py` (per-stage timings and peak memory)
├── tests/                         # Unit tests on tiny models (CUDA tests skip without a GPU)
├── Makefile                       # `make install`, `make test`, and `make lint`
├── pyproject.toml                 # Package metadata and optional extras (`hopper`, `sgl`, `fp8`, `dev`)
├── LICENSE                        # Apache License 2.0
└── THIRD_PARTY_NOTICES.md         # Notices for components adapted from `diffusers`

Quyidagilarni O'rgangan

HyperFlow H3 Flash ning ishlash tezligi va samaradorligi turli vaznlar va GPU konfiguratsiyalariga qarab farqlanadi. Masalan, Q4_K_M 9.1 GB vazni 24 GB GPU larida ishlatish uchun tavsiya etiladi. Q8_0 esa to'liq aniqlikni ta'minlaydi, agar VRAM va sistem RAM mos keladigan bo'lsa.

Model turli integratsiyalarni qo'llab-quvvatlaydi, jumladan ComfyUI, Unsloth Studio va stable-diffusion.cpp. Ushbu integratsiyalar modelni turli platformalarda ishlatish imkonini beradi.

Yakuniy So'z

HyperFlow H3 Flash, video va audio ishlab chiqarish sohasida yangi bosqichni ochdi. U tezroq, samarali va turli platformalarda ishlash imkonini beradi. Bu model, turli vaznlar va konfiguratsiyalarni qo'llab-quvvatlaydi, shuning uchun foydalanuvchilar o'z ehtiyojlariga qarab tanlashlari mumkin.

Asl manba: github.com