메뉴
Accelerate video diffusion models using sparse radial attention that exploits energy decay patterns. Achieves 3.7× speedup on long videos while maintaining quality through O(n log n) complexity instead of O(n²).
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill radial-attention-video명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | radial-attention-video |
| title | Radial Attention: O(nlog n) Sparse Attention with Energy Decay for Long Video Generation |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2506.19852 |
| keywords | ["Sparse Attention","Video Generation","Long Sequence","Computational Efficiency","Diffusion Models"] |
| description | Accelerate video diffusion models using sparse radial attention that exploits energy decay patterns. Achieves 3.7× speedup on long videos while maintaining quality through O(n log n) complexity instead of O(n²). |
Video diffusion models need to attend over thousands of spatial-temporal tokens to generate coherent videos. Standard full attention requires O(n²) memory and computation—infeasible for long videos (256 or more frames). The challenge is finding a sparse pattern that preserves video quality while reducing complexity.
Radial Attention exploits a discovered property of video diffusion: attention scores naturally decay as distance between tokens increases, following an exponential pattern similar to physical signal decay. By converting this energy decay into a static sparse attention mask, you achieve O(n log n) complexity while capturing the meaningful long-range dependencies videos need.
The key insight is that dense attention weights are concentrated locally. In a 256-frame video:
Rather than computing all n² attention pairs, radial attention creates attention bands:
This pattern requires only O(n log n) total comparisons while preserving the exponential decay structure the model learned.
Radial attention modifies the attention mechanism:
Step 1: Compute the radial attention mask
Create a sparse attention pattern that encodes exponential decay.
import torch
import torch.nn.functional as F
math
():
mask = torch.zeros(seq_length, seq_length, dtype=torch., device=device)
spatial_tokens_per_frame = num_spatial_tokens
num_frames = num_temporal_frames
token_idx (seq_length):
frame_i = token_idx // spatial_tokens_per_frame
spatial_i = token_idx % spatial_tokens_per_frame
other_idx (seq_length):
frame_j = other_idx // spatial_tokens_per_frame
spatial_j = other_idx % spatial_tokens_per_frame
temporal_distance = (frame_i - frame_j)
spatial_distance = spatial_distance_on_grid(spatial_i, spatial_j,
grid_size=(math.sqrt(spatial_tokens_per_frame)))
attend = should_attend_radial(temporal_distance, spatial_distance)
attend:
mask[token_idx, other_idx] =
mask
():
x1, y1 = pos1 // grid_size, pos1 % grid_size
x2, y2 = pos2 // grid_size, pos2 % grid_size
(x1 - x2) + (y1 - y2)
():
combined_distance = temporal_distance * + spatial_distance
combined_distance == :
band = math.floor(math.log2((combined_distance, )))
density = decay_factor ** band
attend_probability = (density, )
combined_distance <= ( / attend_probability)
():
attn_weights = attn_weights.masked_fill(~mask, ())
attn_weights
Step 2: Integrate radial attention into video diffusion model
Replace standard scaled dot-product attention with masked attention using the radial pattern.
class RadialAttention(torch.nn.Module):
"""
Multi-head attention with radial sparse masking.
Replaces standard attention in video diffusion models.
"""
def __init__(self, hidden_dim, num_heads=8, num_frames=32,
spatial_grid_size=16):
super().__init__()
self.hidden_dim = hidden_dim
self.num_heads = num_heads
self.head_dim = hidden_dim // num_heads
self.scale = self.head_dim ** -0.5
# Pre-compute mask based on expected sequence structure
self.num_frames = num_frames
self.spatial_grid_size = spatial_grid_size
self.seq_length = (spatial_grid_size ** 2) * num_frames
# Precompute and cache the mask
self.register_buffer(
'radial_mask',
create_radial_attention_mask(
seq_length=self.seq_length,
num_spatial_tokens=spatial_grid_size ** 2,
num_temporal_frames=num_frames
),
persistent=False
)
# Linear projections
self.q_proj = torch.nn.Linear(hidden_dim, hidden_dim)
self.k_proj = torch.nn.Linear(hidden_dim, hidden_dim)
self.v_proj = torch.nn.Linear(hidden_dim, hidden_dim)
self.out_proj = torch.nn.Linear(hidden_dim, hidden_dim)
def forward(self, x):
"""
x: [batch, seq_length, hidden_dim]
"""
batch_size, seq_length, _ = x.shape
# Project to Q, K, V
Q = self.q_proj(x) # [batch, seq_length, hidden_dim]
K = self.k_proj(x)
V = self.v_proj(x)
# Reshape for multi-head attention
# [batch, seq_length, num_heads, head_dim] -> [batch, num_heads, seq_length, head_dim]
Q = Q.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
K = K.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
V = V.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
# Compute attention scores
scores = torch.matmul(Q, K.transpose(-2, -1)) * self.scale
# [batch, num_heads, seq_length, seq_length]
# Apply radial mask: zero out positions not in the sparse pattern
# Adapt mask size if sequence length differs
if seq_length != self.seq_length:
# Create new mask for this sequence length
mask = create_radial_attention_mask(seq_length, device=x.device)
else:
mask = self.radial_mask
# Broadcast mask to batch and heads
mask = mask.unsqueeze(0).unsqueeze(0) # [1, 1, seq_length, seq_length]
scores = apply_radial_attention_mask(scores, mask)
# Softmax (positions masked to -inf become 0)
attn_weights = torch.softmax(scores, dim=-1)
# Apply attention to values
attn_output = torch.matmul(attn_weights, V)
# [batch, num_heads, seq_length, head_dim]
# Concatenate heads
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(batch_size, seq_length, self.hidden_dim)
# Final projection
output = self.out_proj(attn_output)
return output
Step 3: Benchmark speedup and quality
Measure inference time and verify video generation quality remains high.
def benchmark_radial_attention(model_standard, model_radial, test_prompts,
num_frames=256, num_inference_steps=50):
"""
Compare standard full attention vs radial sparse attention.
Measure latency, memory, and quality metrics.
"""
results = {'standard': {}, 'radial': {}}
# Warmup
with torch.no_grad():
_ = model_standard(test_prompts[0], num_frames=num_frames)
_ = model_radial(test_prompts[0], num_frames=num_frames)
# Benchmark standard attention
latencies_standard = []
peak_memory_standard = 0
for prompt in test_prompts:
torch.cuda.reset_peak_memory_stats()
torch.cuda.synchronize()
start_time = time.time()
with torch.no_grad():
output_standard = model_standard(prompt, num_frames=num_frames,
num_inference_steps=num_inference_steps)
torch.cuda.synchronize()
elapsed = time.time() - start_time
latencies_standard.append(elapsed)
memory = torch.cuda.max_memory_allocated() / (1024**2) # MB
peak_memory_standard = max(peak_memory_standard, memory)
# Benchmark radial attention
latencies_radial = []
peak_memory_radial = 0
for prompt in test_prompts:
torch.cuda.reset_peak_memory_stats()
torch.cuda.synchronize()
start_time = time.time()
with torch.no_grad():
output_radial = model_radial(prompt, num_frames=num_frames,
num_inference_steps=num_inference_steps)
torch.cuda.synchronize()
elapsed = time.time() - start_time
latencies_radial.append(elapsed)
memory = torch.cuda.max_memory_allocated() / (1024**2)
peak_memory_radial = max(peak_memory_radial, memory)
# Compute metrics
speedup = np.mean(latencies_standard) / np.mean(latencies_radial)
memory_reduction = (peak_memory_standard - peak_memory_radial) / peak_memory_standard
# Quality comparison: LPIPS, FVD on video frames
# (Details depend on your quality evaluation suite)
quality_standard = evaluate_video_quality(output_standard)
quality_radial = evaluate_video_quality(output_radial)
quality_retention = quality_radial / quality_standard # Should be > 0.95
return {
'speedup': speedup,
'memory_reduction': memory_reduction,
'quality_retention': quality_retention,
'latency_standard_ms': np.mean(latencies_standard) * 1000,
'latency_radial_ms': np.mean(latencies_radial) * 1000
}
Step 4: Fine-tune with LoRA for longer sequences
Enable efficient adaptation to sequences longer than training length using Low-Rank Adapters.
from peft import get_peft_model, LoraConfig
def add_lora_for_sequence_extension(model_with_radial_attention,
target_seq_length=512):
"""
Add LoRA adapters to enable efficient fine-tuning for longer sequences.
LoRA has low rank (r=32-64) so training is cheap.
"""
# Target attention layers for LoRA
lora_config = LoraConfig(
r=32, # Low rank
lora_alpha=64,
target_modules=['q_proj', 'v_proj'], # Key and value projections
lora_dropout=0.05,
bias='none',
task_type='CAUSAL_LM'
)
# Apply LoRA to attention layers
model_lora = get_peft_model(model_with_radial_attention, lora_config)
# Generate a few examples with longer sequences
# Fine-tune on these examples
def create_long_sequence_examples(num_examples=10):
examples = []
for _ in range(num_examples):
# Create synthetic long-sequence data
video_prompt = "Generate a smooth video transition"
examples.append(video_prompt)
return examples
long_examples = create_long_sequence_examples()
# Fine-tune with LoRA
optimizer = torch.optim.AdamW(model_lora.parameters(), lr=1e-4)
for epoch in range(3):
for prompt in long_examples:
# Generate with target_seq_length
# (requires updated mask creation for new length)
loss = model_lora.forward_with_loss(
prompt,
num_frames=int(target_seq_length / 256)
)
optimizer.zero_grad()
loss.backward()
optimizer.step()
return model_lora
| Hyperparameter | Recommended Value | Notes |
|---|---|---|
| Decay factor | 0.5-0.7 | Controls sparsity pattern; 0.5 is standard |
| Temporal weight | 2.0-3.0 | Temporal distance matters more than spatial |
| Minimum density | 0.05 | Prevent total attention starvation |
| Band width | 1 | Logarithmic bands; change rarely |
| LoRA rank | 32-64 | Higher rank = more capacity but slower |
When to use Radial Attention:
When NOT to use Radial Attention:
Common pitfalls:
Radial Attention: O(nlog n) Sparse Attention with Energy Decay for Long Video Generation https://arxiv.org/abs/2506.19852