File size: 4,840 Bytes
5fed0fc |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 |
Flash Attention Optimization Problem
=====================================
Problem Setting
---------------
Design and optimize high-performance Triton kernels for Flash Attention computation on GPU. This problem focuses on implementing efficient attention kernels with causal masking support using Triton's JIT compilation system.
The challenge involves optimizing:
- **Attention computation**: Efficient computation of scaled dot-product attention
- **Causal masking**: Handling causal attention masks efficiently
- **Memory access patterns**: Efficient loading and storing of Q, K, V tensors
- **Numerical stability**: Handling softmax operations with proper numerical stability using streaming softmax
- **Block tiling**: Optimal block sizes for GPU execution across different sequence lengths
- **Performance benchmarking**: Achieving speedup over baseline PyTorch implementations
Target
------
- **Primary**: Maximize geometric mean speedup over baseline (higher is better)
- **Secondary**: Ensure correctness across diverse sequence lengths and attention heads
- **Tertiary**: Minimize kernel launch overhead and memory usage
API Specification
-----------------
Implement a `Solution` class that returns a Triton kernel implementation:
```python
class Solution:
def solve(self, spec_path: str = None) -> dict:
"""
Returns a dict with either:
- {"code": "python_code_string"}
- {"program_path": "path/to/kernel.py"}
"""
# Your implementation
pass
```
Your kernel implementation must provide:
```python
import torch
import triton
import triton.language as tl
def flash_attn(Q: torch.Tensor, K: torch.Tensor, V: torch.Tensor, causal: bool = True) -> torch.Tensor:
"""
Flash attention computation with optional causal masking.
Args:
Q: Input tensor of shape (Z, H, M, Dq) - query tensor (float16)
K: Input tensor of shape (Z, H, N, Dq) - key tensor (float16)
V: Input tensor of shape (Z, H, N, Dv) - value tensor (float16)
causal: Whether to apply causal masking (default True)
Returns:
Output tensor of shape (Z, H, M, Dv) - attention output (float16)
"""
# Your implementation
pass
```
Input Specifications
--------------------
- **Q**: Query tensor of shape `(Z, H, M, Dq)` where:
- `Z`: Batch size (typically 1)
- `H`: Number of attention heads (typically 8)
- `M`: Query sequence length (tested with 512, 1024, 2048)
- `Dq`: Query/key feature dimension (typically 64)
- **K**: Key tensor of shape `(Z, H, N, Dq)` where `N` matches `M` for flash attention
- **V**: Value tensor of shape `(Z, H, N, Dv)` where:
- `Dv`: Value feature dimension (typically 64)
- All inputs are `torch.float16` and on CUDA device
- `causal`: Boolean flag for causal masking (default True)
Output Specifications
--------------------
- Output tensor of shape `(Z, H, M, Dv)` matching the query batch/head dimensions
- Output dtype: `torch.float16`
- Output device: Same as input (CUDA)
Correctness Requirements
-------------------------
- Numerical correctness verified against PyTorch baseline implementation
- Relative tolerance: 1e-2, Absolute tolerance: 5e-3
- All test cases must pass for any score above 0
- Causal masking must be correctly implemented when `causal=True`
Scoring (0-100)
---------------
Performance is measured against GPU baseline implementations:
```
geometric_mean_gpu_time = geometric_mean(gpu_baseline_times)
geometric_mean_answer_time = geometric_mean(answer_times)
# Linear interpolation: 0 points = 1x GPU baseline, 100 points = 10x GPU baseline
target_time_0 = geometric_mean_gpu_time # 0 points (1x GPU baseline)
target_time_100 = geometric_mean_gpu_time / 10.0 # 100 points (10x speedup over GPU)
score = 100 * (target_time_0 - geometric_mean_answer_time) / (target_time_0 - target_time_100)
```
- 0 points = 1x GPU baseline performance
- 100 points = 10x speedup over GPU baseline
- Score is linearly interpolated between these two points
Note: Correctness is verified against GPU baseline, and scoring spans from 1x GPU baseline (0 points) to 10x GPU baseline (100 points).
Evaluation Details
------------------
- Test cases: M = 512, 1024, 2048 (with N = M)
- Warmup phase: 10 iterations to stabilize GPU clocks and caches
- Random seed: Fixed seed (0) for reproducible data generation
- Strict correctness: Any test failure results in score of 0
Additional Notes
----------------
- The benchmark uses float32 for PyTorch baseline (for numerical stability) but float16 for answer evaluation
- Streaming softmax techniques are recommended for numerical stability
- Consider using block pointers (`tl.make_block_ptr`) for efficient memory access
- Causal masking requires careful attention to the masking pattern (lower triangular for causal attention)
|