Skip to content
Research Agent Skillsby Kalaris Labs
Install
ML Trainingv1.0.0AdaptedMIT

optimizing-attention-flash

Enables Flash Attention for transformer models using PyTorch native scaled_dot_product_attention (PyTorch 2.2+) or the flash-attn library, including multi-query attention, sliding window attention, and FP8 on H100 (FlashAttention-3). Covers profiling speedup, checking accuracy against a baseline, and troubleshooting install and GPU support errors. Use when training or running transformers on long sequences (over 512 tokens), when standard attention runs out of GPU memory, when attention is the inference bottleneck, when switching a PyTorch model to the flash backend, or when tuning attention on H100 GPUs. Not for CPU inference, V100 GPUs, or sequences under 256 tokens; consider xFormers for other attention variants.

Install this skill

Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.

skills.sh CLI (Recommended)
npx skills add KalarisLabs/research-agent-skills --skill optimizing-attention-flash
uvx (Python tool runner)
uvx research-agent-skills install optimizing-attention-flash

Documentation & Instructions

Flash Attention - Fast Memory-Efficient Attention

Quick start

Flash Attention provides 2-4x speedup and 10-20x memory reduction for transformer attention through IO-aware tiling and recomputation.

PyTorch native (easiest, PyTorch 2.2+):

import torch
import torch.nn.functional as F

q = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)  # [batch, heads, seq, dim]
k = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)

# Automatically uses Flash Attention if available
out = F.scaled_dot_product_attention(q, k, v)

flash-attn library (more features):

pip install flash-attn --no-build-isolation
from flash_attn import flash_attn_func

# q, k, v: [batch, seqlen, nheads, headdim]
out = flash_attn_func(q, k, v, dropout_p=0.0, causal=True)

Common workflows

Workflow 1: Enable in existing PyTorch model

Copy this checklist:

Flash Attention Integration:
- [ ] Step 1: Check PyTorch version (≥2.2)
- [ ] Step 2: Enable Flash Attention backend
- [ ] Step 3: Verify speedup with profiling
- [ ] Step 4: Test accuracy matches baseline

Step 1: Check PyTorch version

python -c "import torch; print(torch.__version__)"
# Should be ≥2.2.0

If <2.2, upgrade:

pip install --upgrade torch

Step 2: Enable Flash Attention backend

Replace standard attention:

# Before (standard attention)
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / math.sqrt(d_k), dim=-1)
out = attn_weights @ v

# After (Flash Attention)
import torch.nn.functional as F
out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)

Force Flash Attention backend:

with torch.backends.cuda.sdp_kernel(
    enable_flash=True,
    enable_math=False,
    enable_mem_efficient=False
):
    out = F.scaled_dot_product_attention(q, k, v)

Step 3: Verify speedup with profiling

import torch.utils.benchmark as benchmark

def test_attention(use_flash):
    q, k, v = [torch.randn(2, 8, 2048, 64, device='cuda', dtype=torch.float16) for _ in range(3)]

    if use_flash:
        with torch.backends.cuda.sdp_kernel(enable_flash=True):
            return F.scaled_dot_product_attention(q, k, v)
    else:
        attn = (q @ k.transpose(-2, -1) / 8.0).softmax(dim=-1)
        return attn @ v

# Benchmark
t_flash = benchmark.Timer(stmt='test_attention(True)', globals=globals())
t_standard = benchmark.Timer(stmt='test_attention(False)', globals=globals())

print(f"Flash: {t_flash.timeit(100).mean:.3f}s")
print(f"Standard: {t_standard.timeit(100).mean:.3f}s")

Expected: 2-4x speedup for sequences >512 tokens.

Step 4: Test accuracy matches baseline

# Compare outputs
q, k, v = [torch.randn(1, 8, 512, 64, device='cuda', dtype=torch.float16) for _ in range(3)]

# Flash Attention
out_flash = F.scaled_dot_product_attention(q, k, v)

# Standard attention
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / 8.0, dim=-1)
out_standard = attn_weights @ v

# Check difference
diff = (out_flash - out_standard).abs().max()
print(f"Max difference: {diff:.6f}")
# Should be <1e-3 for float16

Workflow 2: Use flash-attn library for advanced features

For multi-query attention, sliding window, or H100 FP8.

Copy this checklist:

flash-attn Library Setup:
- [ ] Step 1: Install flash-attn library
- [ ] Step 2: Modify attention code
- [ ] Step 3: Enable advanced features
- [ ] Step 4: Benchmark performance

Step 1: Install flash-attn library

# NVIDIA GPUs (CUDA 12.0+)
pip install flash-attn --no-build-isolation

# Verify installation
python -c "from flash_attn import flash_attn_func; print('Success')"

Step 2: Modify attention code

from flash_attn import flash_attn_func

# Input: [batch_size, seq_len, num_heads, head_dim]
# Transpose from [batch, heads, seq, dim] if needed
q = q.transpose(1, 2)  # [batch, seq, heads, dim]
k = k.transpose(1, 2)
v = v.transpose(1, 2)

out = flash_attn_func(
    q, k, v,
    dropout_p=0.1,
    causal=True,  # For autoregressive models
    window_size=(-1, -1),  # No sliding window
    softmax_scale=None  # Auto-scale
)

out = out.transpose(1, 2)  # Back to [batch, heads, seq, dim]

Step 3: Enable advanced features

Multi-query attention (shared K/V across heads):

from flash_attn import flash_attn_func

# q: [batch, seq, num_q_heads, dim]
# k, v: [batch, seq, num_kv_heads, dim]  # Fewer KV heads
out = flash_attn_func(q, k, v)  # Automatically handles MQA

Sliding window attention (local attention):

# Only attend to window of 256 tokens before/after
out = flash_attn_func(
    q, k, v,
    window_size=(256, 256),  # (left, right) window
    causal=True
)

Step 4: Benchmark performance

import torch
from flash_attn import flash_attn_func
import time

q, k, v = [torch.randn(4, 4096, 32, 64, device='cuda', dtype=torch.float16) for _ in range(3)]

# Warmup
for _ in range(10):
    _ = flash_attn_func(q, k, v)

# Benchmark
torch.cuda.synchronize()
start = time.time()
for _ in range(100):
    out = flash_attn_func(q, k, v)
    torch.cuda.synchronize()
end = time.time()

print(f"Time per iteration: {(end-start)/100*1000:.2f}ms")
print(f"Memory allocated: {torch.cuda.max_memory_allocated()/1e9:.2f}GB")

Workflow 3: H100 FP8 optimization (FlashAttention-3)

For maximum performance on H100 GPUs.

FP8 Setup:
- [ ] Step 1: Verify H100 GPU available
- [ ] Step 2: Install flash-attn with FP8 support
- [ ] Step 3: Convert inputs to FP8
- [ ] Step 4: Run with FP8 attention

Step 1: Verify H100 GPU

nvidia-smi --query-gpu=name --format=csv
# Should show "H100" or "H800"

Step 2: Install flash-attn with FP8 support

pip install flash-attn --no-build-isolation
# FP8 support included for H100

Step 3: Convert inputs to FP8

import torch

q = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
k = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)

# Convert to float8_e4m3 (FP8)
q_fp8 = q.to(torch.float8_e4m3fn)
k_fp8 = k.to(torch.float8_e4m3fn)
v_fp8 = v.to(torch.float8_e4m3fn)

Step 4: Run with FP8 attention

from flash_attn import flash_attn_func

# FlashAttention-3 automatically uses FP8 kernels on H100
out = flash_attn_func(q_fp8, k_fp8, v_fp8)
# Result: ~1.2 PFLOPS, 1.5-2x faster than FP16

When to use vs alternatives

Use Flash Attention when:

  • Training transformers with sequences >512 tokens
  • Running inference with long context (>2K tokens)
  • GPU memory constrained (OOM with standard attention)
  • Need 2-4x speedup without accuracy loss
  • Using PyTorch 2.2+ or can install flash-attn

Use alternatives instead:

  • Standard attention: Sequences <256 tokens (overhead not worth it)
  • xFormers: Need more attention variants (not just speed)
  • Memory-efficient attention: CPU inference (Flash Attention needs GPU)

Common issues

Issue: ImportError: cannot import flash_attn

Install with no-build-isolation flag:

pip install flash-attn --no-build-isolation

Or install CUDA toolkit first:

conda install cuda -c nvidia
pip install flash-attn --no-build-isolation

Issue: Slower than expected (no speedup)

Flash Attention benefits increase with sequence length:

  • <512 tokens: Minimal speedup (10-20%)
  • 512-2K tokens: 2-3x speedup
  • 2K tokens: 3-4x speedup

Check sequence length is sufficient.

Issue: RuntimeError: CUDA error

Verify GPU supports Flash Attention:

import torch
print(torch.cuda.get_device_capability())
# Should be ≥(7, 5) for Turing+

Flash Attention requires:

  • Ampere (A100, A10): ✅ Full support
  • Turing (T4): ✅ Supported
  • Volta (V100): ❌ Not supported

Issue: Accuracy degradation

Check dtype is float16 or bfloat16 (not float32):

q = q.to(torch.float16)  # Or torch.bfloat16

Flash Attention uses float16/bfloat16 for speed. Float32 not supported.

Advanced topics

Integration with HuggingFace Transformers: See references/transformers-integration.md for enabling Flash Attention in BERT, GPT, Llama models.

Performance benchmarks: See references/benchmarks.md for detailed speed and memory comparisons across GPUs and sequence lengths.

Algorithm details: See references/algorithm.md for tiling strategy, recomputation, and IO complexity analysis.

Advanced features: See references/advanced-features.md for rotary embeddings, ALiBi, paged KV cache, and custom attention masks.

Hardware requirements

  • GPU: NVIDIA Ampere+ (A100, A10, A30) or AMD MI200+
  • VRAM: Same as standard attention (Flash Attention doesn't increase memory)
  • CUDA: 12.0+ (11.8 minimum)
  • PyTorch: 2.2+ for native support

Not supported: V100 (Volta), CPU inference

Resources

Agent operating procedure

  1. Check the environment. Check GPU type, memory and driver/CUDA versions (nvidia-smi), framework versions, and dataset location and size.
  2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
  3. Run a small version first. Do a smoke run: tiny model or subset, few steps, and confirm loss decreases and checkpoints save.
  4. Execute the full task using the instructions and references above.
  5. Validate the result. Track metrics on held-out data, compare against a baseline, and record seeds, configs and hardware.
  6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.
If this happens Do this
CUDA out-of-memory Reduce batch size, enable gradient accumulation/checkpointing or mixed precision, or shard the model.
Loss is NaN or diverges Lower the learning rate, check data for invalid values, and enable gradient clipping.
A function, flag or endpoint in these instructions is missing in the installed version Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous Ask the user, or state the assumption explicitly before running.

Integrity rules

  • Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
  • Never claim training results without logs; estimate compute cost before launching large jobs.
  • Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
  • Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

  • rwkv-architecture: Covers the RWKV (Receptance Weighted Key Value) architecture, an RNN/Transformer hybrid with O(n) inference and no KV cache, including RWKV…
  • llama-cpp: Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware.
  • tensorrt-llm: Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency.

Trigger Prompts & Evaluation Corpus

These evaluation prompts test that the agent correctly selects optimizing-attention-flash for relevant tasks and leaves it out for near misses.

Should trigger (3)

  • should trigger

    “Can I make my PyTorch training loop use flash attention? Sequences are 4096 tokens.”

    expects Checks PyTorch version (2.2+), replaces manual attention with F.scaled_dot_product_attention (optionally forcing the flash backend via sdpa_kernel), uses fp16/bf16, and verifies speedup and output equivalence against the baseline.

  • should trigger

    “I'm fine-tuning Llama-3-8B on an A100 80GB with 16k context using my train.py and keep getting 'CUDA out of memory' inside the attention forward pass. I'm on PyTorch 2.3 and the HF model loads with the default eager attention. How do I fix this without shrinking the context?”

    expects Recommends switching to the flash/SDPA attention implementation (e.g. attn_implementation='flash_attention_2' or 'sdpa'), ensuring bf16/fp16 dtype, and confirming the GPU supports it. Then runs a small smoke test, measures memory and speed, and checks the loss or outputs match the eager baseline. Mentions gradient checkpointing as a complementary option and doesn't invent APIs.

  • should trigger

    “We have a few H100 nodes and attention is the main bottleneck for our long-document model at inference. I'd like to squeeze out the most throughput using lower precision for the attention math, and also have each token attend only to a local window of neighbors. What should I do?”

    expects Identifies FlashAttention-3 with FP8 on H100 and the flash-attn library's sliding window option. Verifies the GPU is Hopper (nvidia-smi / compute capability), installs flash-attn with the right CUDA/PyTorch versions, and benchmarks. Checks accuracy against a bf16 baseline since FP8 can degrade quality, and states uncertainty about version-specific API details.

Near misses (2)

  • near miss

    “I need to run a 7B chat model on my MacBook M2 with no NVIDIA GPU. How do I get it running fast with quantization?”

    expects Does not use the flash attention skill, since it requires an NVIDIA/AMD GPU. Recommends llama.cpp with GGUF quantization for Apple Silicon.

  • near miss

    “Explain how RWKV-7 gets linear-time inference without a KV cache, and show how to run it in RNN mode for text generation.”

    expects Explains the RWKV recurrent state mechanism and RNN-mode inference using the RWKV architecture skill. It does not turn to flash attention, since RWKV has no attention softmax to optimize.