ComfyUI Attention Optimizer

Automatically benchmark and optimize the attention mechanism in diffusion models for maximum generation speed.

Why This Matters

The Problem

Modern diffusion models (SDXL, Flux, WAN, LTX-V, Hunyuan Video) are based on transformer architecture. The core operation - attention - computes relationships between all elements in the image/video latent space. This is:

• The most expensive operation - attention takes 40-70% of total generation time
• O(n²) complexity - cost grows quadratically with resolution/frames
• GPU-dependent - different GPUs perform best with different implementations

The Solution

Multiple optimized attention backends exist:

• PyTorch SDPA - built-in, always available
• Flash Attention - CUDA kernels, memory efficient
• SageAttention - INT8 quantization, up to 2-4x faster
• xFormers - memory efficient attention

But which one is fastest for YOUR specific GPU and model?

This plugin benchmarks all available backends and automatically applies the fastest one.

Real-World Speedups

Tested on RTX 4090 with head_dim=128 (SDXL, Flux):

| Backend | Time | Speedup | |---------|------|---------| | PyTorch SDPA | 5.0ms | 1.0x (baseline) | | Flash Attention | 5.4ms | 0.93x | | SageAttention | 2.7ms | 1.9x |

Result: 1.9x faster generation just by switching attention backend.

For video models (WAN, Hunyuan) with longer sequences, speedups can reach 2-4x.

Installation

Option 1: ComfyUI Manager (Recommended)

1. Open ComfyUI Manager
2. Click "Install via Git URL"
3. Paste: https://github.com/D-Ogi/ComfyUI-Attention-Optimizer.git
4. Restart ComfyUI

Option 2: Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/D-Ogi/ComfyUI-Attention-Optimizer.git

Restart ComfyUI.

Optional: Install Optimized Backends

The plugin works out-of-the-box with PyTorch SDPA. For better performance, install additional backends:

# SageAttention - recommended for RTX 30xx/40xx (1.5-2x speedup)
pip install sageattention

# Flash Attention - alternative for Ampere+ GPUs
pip install flash-attn

# xFormers - memory efficient option
pip install xformers

Note: On Windows, Flash Attention requires building from source or using prebuilt wheels. SageAttention is easier to install and often faster on consumer GPUs.

Usage

Basic Usage

1. Add "Attention Optimizer" node to your workflow (category: model_patches)
2. Connect your model to the model input
3. Run - it benchmarks once, caches results, and auto-applies the fastest backend

How It Works

┌─────────────────┐     ┌──────────────────────────┐     ┌─────────────┐
│ Load Checkpoint │────▶│ Attention Optimizer      │────▶│ KSampler    │
└─────────────────┘     │                          │     └─────────────┘
                        │ 1. Detect model params   │
                        │ 2. Check cache           │
                        │ 3. Benchmark (if needed) │
                        │ 4. Clone model & apply   │
                        │    attention override    │
                        └──────────────────────────┘

First run: Benchmarks all backends (~5-10 seconds), saves to cache. Subsequent runs: Loads from cache (instant), applies optimal backend.

Node Inputs

| Input | Type | Default | Description | |-------|------|---------|-------------| | model | MODEL | required | The diffusion model to optimize | | attention_backend | dropdown | auto | auto = benchmark & select best, or force specific backend | | force_refresh | bool | False | Re-run benchmark even if cached | | auto_apply | bool | True | Apply the selected backend to this model | | seq_len | int | 8192 | Sequence length for benchmark | | num_heads | int | 24 | Number of attention heads |

Node Outputs

| Output | Type | Description | |--------|------|-------------| | model | MODEL | Cloned model with optimized attention applied | | best_attention | STRING | Name of applied backend | | kjnodes_mode | STRING | Compatible mode for KJNodes PatchSageAttention | | impl_type | STRING | Implementation type (cuda/triton/pytorch) | | speedup | FLOAT | Speedup vs PyTorch SDPA baseline | | time_ms | FLOAT | Time per attention call in milliseconds | | head_dim | INT | Detected head dimension from model | | report | STRING | Full benchmark report text |

Supported Backends

| Backend | Implementation | Best For | |---------|---------------|----------| | pytorch | PyTorch SDPA | Always available, baseline | | xformers | xFormers CUDA | Memory efficiency | | sage_auto | SageAttention auto | General use (auto-selects best variant) | | sage_cuda | SageAttention CUDA | RTX 30xx/40xx | | sage_triton | SageAttention Triton | When CUDA kernel unavailable | | sage_fp8_cuda | SageAttention FP8 | Maximum speed, slight quality trade-off | | sage_fp8_cuda_fast | SageAttention FP8++ | Even faster FP8 | | sage3 | SageAttention 3 | RTX 50xx (Blackwell) only | | flash | Flash Attention 2 | H100, A100, RTX 30xx/40xx |

Model Compatibility

| Model | Status | Notes | |-------|--------|-------| | SDXL | ✅ Full | head_dim=128, SageAttention optimal | | SD 1.5 | ✅ Full | head_dim=64 | | SD 3 | ✅ Full | | | Flux | ✅ Full | Per-model attention override | | LTX-V | ✅ Full | head_dim=160 | | WAN 2.1/2.2 | ✅ Full | Per-model attention override | | Hunyuan Video | ✅ Full | Per-model attention override | | Cosmos | ✅ Full | Per-model attention override | | SeedVR2 | ❌ N/A | Uses own attention system, not affected |

GPU Recommendations

| GPU | Recommended Backend | Expected Speedup | |-----|---------------------|------------------| | RTX 4090/4080 | sage_auto or sage_fp8_cuda_fast | 1.5-2.0x | | RTX 3090/3080 | sage_auto or flash | 1.3-1.8x | | RTX 50xx (Blackwell) | sage3 | 2-4x | | H100/A100 | flash | 1.5-2.0x | | AMD (ROCm) | pytorch | 1.0x (baseline) |

Example Benchmark Report

=================================================================
BENCHMARK REPORT
=================================================================
dtype: float16 | head_dim: 128 | seq_len: 8192 | CUDA: 12.4 | Triton: 3.0.0
SageAttention: v2.1.1

>>> BEST: sage_fp8_cuda_fast (1.89x speedup) <<<
    impl: cuda | kjnodes mode: sageattn_qk_int8_pv_fp8_cuda++

Results (fastest first):
-----------------------------------------------------------------
 [v] sage_fp8_cuda_fast       2.671ms   1.89x  (cuda) <<<
 [v] sage_auto                2.679ms   1.88x  (auto)
 [v] sage_fp8_cuda            3.100ms   1.63x  (cuda)
 [v] sage_triton              3.446ms   1.47x  (triton)
 [v] sage_cuda                3.947ms   1.28x  (cuda)
 [v] pytorch                  5.049ms   1.00x  (pytorch)
 [v] xformers                 5.194ms   0.97x  (cuda/triton)
 [v] flash                    5.430ms   0.93x  (cuda)
 [ ] sage3                    ---       (N/A) Not installed
-----------------------------------------------------------------
[v] = validated (tested underlying library directly)
=================================================================

Technical Details

Why Different Backends?

PyTorch SDPA uses cuDNN/cuBLAS - general purpose, always works.

Flash Attention fuses operations into single CUDA kernel, reducing memory bandwidth. Great for long sequences.

SageAttention quantizes Q/K to INT8, reducing memory and compute. Works best for head_dim ≤ 128.

xFormers similar to Flash Attention, good memory efficiency.

head_dim Matters

Models have different attention head dimensions:

• SD 1.5: head_dim=64
• SDXL, Flux: head_dim=128
• LTX-V: head_dim=160

SageAttention works best with head_dim ≤ 128. For larger dimensions, SDPA or Flash Attention may be faster.

Cache System

Benchmark results are cached in benchmark_db.json based on:

• Model hash (architecture + weights)
• head_dim
• seq_len / num_heads parameters

Cache is per-machine - different GPUs will have different optimal backends.

Troubleshooting

"Backend X not available"

Install the missing package:

pip install sageattention  # for sage_*
pip install flash-attn     # for flash
pip install xformers       # for xformers

No speedup observed

1. Check if auto_apply is enabled
2. Try force_refresh=True to re-benchmark
3. Check console for [Benchmark] Applied: X message

Model not affected

Some models (like SeedVR2) use their own attention implementation and won't be affected by this plugin. Check the compatibility table above.

License

MIT License - see LICENSE

Credits

• SageAttention - THU-ML
• Flash Attention - Dao-AILab
• xFormers - Meta