SageAttention could speed up Sharpen / Rescue (SeedVR2) without touching the Fibo templates.
Why it might help:
SeedVR2's DiT step at 16 MP is dominated by attention: ~36 s for 7B and ~23 s for 3B on the 3090.
SageAttention (INT8 QK) is typically ~2× faster per attention op, so that step could drop to ~20–25 s.
The VAE encode/decode (~21 s) and model load aren't affected.
Fibo at ~1 MP would gain less (maybe 10–25%).
Scope:
SeedVR2 has its own attention_mode input on SeedVR2LoadDiTModel (sageattn_2), separate from ComfyUI's global --use-sage-attention.
Fibo uses core nodes and would only change if that global flag were set.
Linux (3090):
PyPI sageattention 1.0.6 is Triton-based (no compile).
It exposes sageattn_varlen with a signature compatible with what the node pack calls.
Not yet run.
SageAttention 2 needs an nvcc build; the system nvcc is 12.0 and torch is cu128.
Windows (4090):
Prebuilt SageAttention 2 wheels exist (they also use the 4090's FP8 path).
Not installed there today.
Plan:
Install on one machine.
Run Sharpen with attention_mode = sageattn_2 vs sdpa on the same photos.
Compare DiT time and quality (Sage occasionally adds faint artifacts).
If it's worth it, set it in /media/nas/image_seedvr2_*_imageapi_API.json.
SageAttention could speed up Sharpen / Rescue (SeedVR2) without touching the Fibo templates.
- **Why it might help:**
- SeedVR2's DiT step at 16 MP is dominated by attention: ~36 s for 7B and ~23 s for 3B on the 3090.
- SageAttention (INT8 QK) is typically ~2× faster per attention op, so that step could drop to ~20–25 s.
- The VAE encode/decode (~21 s) and model load aren't affected.
- Fibo at ~1 MP would gain less (maybe 10–25%).
- **Scope:**
- SeedVR2 has its own `attention_mode` input on `SeedVR2LoadDiTModel` (`sageattn_2`), separate from ComfyUI's global `--use-sage-attention`.
- Fibo uses core nodes and would only change if that global flag were set.
- **Linux (3090):**
- PyPI `sageattention` 1.0.6 is Triton-based (no compile).
- It exposes `sageattn_varlen` with a signature compatible with what the node pack calls.
- Not yet run.
- SageAttention 2 needs an nvcc build; the system nvcc is 12.0 and torch is cu128.
- **Windows (4090):**
- Prebuilt SageAttention 2 wheels exist (they also use the 4090's FP8 path).
- Not installed there today.
- **Plan:**
- Install on one machine.
- Run Sharpen with `attention_mode = sageattn_2` vs `sdpa` on the same photos.
- Compare DiT time and quality (Sage occasionally adds faint artifacts).
- If it's worth it, set it in `/media/nas/image_seedvr2_*_imageapi_API.json`.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
SageAttention could speed up Sharpen / Rescue (SeedVR2) without touching the Fibo templates.
attention_modeinput onSeedVR2LoadDiTModel(sageattn_2), separate from ComfyUI's global--use-sage-attention.sageattention1.0.6 is Triton-based (no compile).sageattn_varlenwith a signature compatible with what the node pack calls.attention_mode = sageattn_2vssdpaon the same photos./media/nas/image_seedvr2_*_imageapi_API.json.