flash-prefill

Accelerates long-context LLM prefilling by identifying sparse attention patterns without expensive scoring, using block-level approximations and dynamic thresholding. Achieves 27.78x speedup at 256K t

SKILL.md

GitHub

FlashPrefill: Achieving 27x Speedup via Instant Sparse Attention Pattern Discovery

Long-context LLM inference suffers from quadratic attention complexity during prefilling. Computing full attention over sequences of 100K+ tokens incurs enormous memory and compute costs. Existing sparse attention methods like Top-k and Top-p require expensive score computation, sorting operations, and handle long-tail distributions poorly, creating additional overhead that partially cancels efficiency gains.

FlashPrefill eliminates this overhead through instantaneous pattern discovery: identifying which attention blocks matter without materializing full score matrices. By using block-level proxies and max-based thresholding, it achieves 27.78x speedup on 256K sequences while maintaining near-identical accuracy.

Core Concept

Rather than computing all Q-K dot products then sorting (O(L²) or worse), FlashPrefill:

  1. Instant Pattern Recognition: Uses uniformly distributed query probes to simultaneously identify vertical (column-sparse), slash (diagonal), and block-sparse patterns from block-level statistics alone
  2. Block Approximation: Computes block-pair importance via fused 2D-reduction kernels, reducing memory traffic from O(L²/B) to O((L/B)²)
  3. Dynamic Thresholding: Replaces sorting with single-pass max reduction to determine pruning thresholds, avoiding cumulative summation overhead

The key insight: attention patterns are often predictable at block granularity without computing token-level scores. Vertical patterns (attending to few positions) and slash patterns (attending to sliding windows) can be identified from approximate block interactions.

Architecture Overview

  • Block-Level Proxy Computation: Compute block-level attention scores using average-pooled keys/queries within each block
  • Three Pattern Types Detected: Vertical (few columns), Slash (diagonal bands), Block (rectangular dense regions)
  • Fused Kernel Implementation: Single-pass kernel computes all block interactions without intermediate materialization
  • Physical Index Jumping: Use identified sparse block indices to implement block-sparse attention efficiently

Implementation Steps

The algorithm operates in three stages executed sequentially in a fused manner.

Stage 1: Block-Level Attention Approximation

Compute approximate block importance scores using pooled features. For a sequence of length L with block size B, create L/B "block queries" by averaging queries in each block.

# Compute block-level attention scores
def block_attention_scores(Q, K, block_size):
    """
    Q: query tensor [seq_len, dim]
    K: key tensor [seq_len, dim]
    block_size: B (e.g., 64 or 128)

    Returns: score matrix [num_blocks, num_blocks]
    """
    seq_len = Q.shape[0]
    num_blocks = (seq_len + block_size - 1) // block_size

    # Pool queries and keys by block
    Q_blocks = []
    K_blocks = []

    for i in range(num_blocks):
        start = i * block_size
        end = min((i + 1) * block_size, seq_len)

        # Average pool within block
        Q_blocks.append(Q[start:end].mean(dim=0, keepdim=True))
        K_blocks.append(K[start:end].mean(dim=0, keepdim=True))

    Q_blocks = torch.cat(Q_blocks, dim=0)  # [num_blocks, dim]
    K_blocks = torch.cat(K_blocks, dim=0)  # [num_blocks, dim]

    # Compute block-level attention scores
    block_scores = torch.matmul(Q_blocks, K_blocks.t()) / math.sqrt(Q.shape[-1])
    return block_scores  # [num_blocks, num_blocks]

Stage 2: Pattern Identification and Dynamic Thresholding

Identify which blocks to attend to using a single-pass max reduction for the threshold.

# Identify sparse pattern via dynamic thresholding
def identify_sparse_pattern(block_scores, threshold_percentile=80):
    """
    block_scores: [num_blocks, num_blocks]
    threshold_percentile: cutoff for sparsity (e.g., 80 means keep top 20%)

    Returns: boolean mask [num_blocks, num_blocks] indicating which blocks to compute
    """
    # Single-pass max to establish dynamic threshold
    max_score = block_scores.max()
    min_score = block_scores.min()

    # Dynamic threshold: alpha * max_score
    # alpha often set to 0.1-0.3 depending on desired sparsity
    threshold = 0.15 * max_score + 0.85 * min_score

    # Binary mask: attend to blocks above threshold
    pattern_mask = block_scores > threshold

    return pattern_mask

Stage 3: Fused Block-Sparse Attention

Execute attention only over identified sparse blocks. Materialize full token-level attention only for selected block pairs.

# Compute sparse attention efficiently
def sparse_block_attention(Q, K, V, pattern_mask, block_size):
    """
    Q, K, V: [seq_len, dim] tensors
    pattern_mask: [num_blocks, num_blocks] boolean mask
    block_size: B

    Returns: attended values [seq_len, dim]
    """
    seq_len = Q.shape[0]
    num_blocks = pattern_mask.shape[0]

    # Initialize output
    output = torch.zeros_like(Q)
    attention_sum = torch.zeros(seq_len, 1, device=Q.device)

    # Iterate only over unmasked (attended) block pairs
    for i in range(num_blocks):
        for j in range(num_blocks):
            if not pattern_mask[i, j]:
                continue  # Skip masked blocks

            # Compute full attention for this block pair only
            q_start, q_end = i * block_size, min((i + 1) * block_size, seq_len)
            k_start, k_end = j * block_size, min((j + 1) * block_size, seq_len)

            Q_block = Q[q_start:q_end]
            K_block = K[k_start:k_end]
            V_block = V[k_start:k_end]

            # Standard scaled dot-product attention for this block pair
            scores = torch.matmul(Q_block, K_block.t()) / math.sqrt(Q.shape[-1])
            attn_weights = torch.softmax(scores, dim=-1)
            output[q_start:q_end] += torch.matmul(attn_weights, V_block)
            attention_sum[q_start:q_end] += attn_weights.sum(dim=-1, keepdim=True)

    # Normalize by attention weights
    output = output / (attention_sum + 1e-10)
    return output

Practical Guidance

Hyperparameters:

  • Block size: Typically 64-256 tokens (larger blocks = faster but less precise pattern discovery)
  • Threshold α: 0.1-0.3 depending on desired sparsity; 0.15 is a good default
  • Pattern types: All three (vertical, slash, block) should be checked; enable/disable based on profiling

When to Apply:

  • Long-context inference (100K+ tokens) where attention is the bottleneck
  • Prefilling phase specifically—not decoding, where sparse patterns are less predictable
  • Models with uniform attention patterns (e.g., many retrieval-based tasks)

When NOT to Apply:

  • Short sequences (<4K tokens) where full attention is already fast
  • Decoding phase (generation token-by-token) where causal masking changes pattern structure
  • Tasks requiring dense attention across all positions (e.g., some translation tasks)

Key Pitfalls:

  • Setting block_size too large eliminates fine-grained patterns; too small creates overhead
  • Threshold too high = missing important attention; too low = no speedup
  • Pattern types not aligned with workload (e.g., slash patterns matter for retrieval but not generation)
  • Not accounting for causal masking if using in autoregressive settings

Integration Notes: Works as a drop-in replacement for standard attention in prefilling phases; requires minimal code changes to attention kernels and integrates with Flash-Attention style implementations.

Evidence: Achieves 27.78x speedup at 256K token lengths on Llama-3 models; maintains <1% accuracy drop across diverse vision-language and text-only models; largest gains on retrieval-augmented generation tasks.

Reference: https://arxiv.org/abs/2603.06199