Ethanfel dd61ae8d1f Bundle sparse_sage Triton kernel for block-sparse attention ...

Without sparse attention, the model uses full (dense) attention which
attends to distant irrelevant information, causing ghosting artifacts.
The FlashVSR paper explicitly requires block-sparse attention.

Vendored from SageAttention team (Apache 2.0), pure Triton (no CUDA C++).
Import chain: local sparse_sage → external sageattn.core → SDPA fallback.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-13 19:22:40 +01:00

..

configs

Add FlashVSR support: diffusion-based 4x video super-resolution (Wan 2.1-1.3B)

2026-02-13 15:12:33 +01:00

models

Bundle sparse_sage Triton kernel for block-sparse attention

2026-02-13 19:22:40 +01:00

pipelines

Fix FlashVSR attention mask and output quality

2026-02-13 18:41:43 +01:00

schedulers

Add FlashVSR support: diffusion-based 4x video super-resolution (Wan 2.1-1.3B)

2026-02-13 15:12:33 +01:00

vram_management

Add FlashVSR support: diffusion-based 4x video super-resolution (Wan 2.1-1.3B)

2026-02-13 15:12:33 +01:00

__init__.py

Add FlashVSR support: diffusion-based 4x video super-resolution (Wan 2.1-1.3B)

2026-02-13 15:12:33 +01:00