随机生成pytorch算子测试序列且保证算子参数合法
背景:
1.一些对维度进行操作的算子的单算子测试,结果正常,但多个算子组合在一起,结果就不对。是否能给一个算子列表,随机生成它们的组合呢
功能描述:
1.此程序用于在 CUDA 环境中生成随机张量并对其施加一系列随机选择的操作
2.程序首先随机生成张量的形状和内容，然后随机选择一个操作（如 reshape、transpose、matmul 等），并生成适当的参数以执行该操作
3.最终输出变换后的张量并打印相关操作信息
4.整个过程在100次不同的种子下每次进行10次操作，以保证操作的多样性和结果的随机性
markup
通过LLM多轮对话生成pytorch算子组合测试用例 小结
初衷: 给一个算子列表,自动生成列表中算子的随机组合测试,可以覆盖不同的shape,支持任意多个算子的组合
原计划(LLM全自动生成):
1.测试了qwen-max、kimi moonshot-v1-128k、ERNIE-4.0-8K、sparkai(3.5)、yi-large这几个模型(各家最新的模型)
2.这几个模型都能按要求生成单元测试用例,但几乎所有的代码运行都会出错(95%以上的错都是shape不匹配)
3.一些模型通过几次交互能修复bug,但整体上效果不理想
4.也许LLM对pytorch算子的约束不太了解,可以尝试将算子的接口文档告诉LLM,采用few shot的方式,是否有所改善。
妥协的方案:
1.于是将这个需求细化,与GPT-4o多次交互,生成了这个功能模块的代码,功能正常。之后加到ut测试中代码
python
import torch
import random
from functools import reduce
from operator import mul
import numpy as np
max_size = 4096 # 每个维度的最大大小
max_tensor_elements = 1*4096*4096 # 张量中元素的总数限制
min_dim_size = 1 # 最小维度大小
max_dim_size = max_size # 扩大这个范围可以更快生成符合要求的大小
def generate_random_shape(dim, max_attempts=10):
for _ in range(max_attempts):
shape = [random.randint(min_dim_size, max_dim_size) for _ in range(dim)]
if reduce(mul, shape, 1) <= max_tensor_elements:
return tuple(shape)
# 兜底策略，防止尝试次数用尽：再遍历生成的随机形状，逐个将维度缩小直到符合限制
shape = [random.randint(1, max_size) for _ in range(dim)]
current_elements = reduce(mul, shape, 1)
while current_elements > max_tensor_elements:
for i in range(len(shape)):
if shape[i] > 1:
shape[i] //= 2
current_elements = reduce(mul, shape, 1)
if current_elements <= max_tensor_elements:
break
return tuple(shape)
def generate_random_input(shape):
return torch.randn(shape).to("cuda").half()
def generate_random_operator(input_shape):
operators = ['unsqueeze', 'repeat', 'permute', 'transpose', 'reshape', 'expand', 'contiguous', 'matmul', 'mul', 'concat',"view"]
return random.choice(operators)
def generate_random_reshape(input_shape):
# 计算输入张量的总元素数
total_elements = np.prod(input_shape)
divisors = []
# 找到 total_elements 的所有约数
for i in range(1, int(np.sqrt(total_elements)) + 1):
if total_elements % i == 0:
divisors.append(i)
if i != total_elements // i:
divisors.append(total_elements // i)
dimensions = []
remaining_elements = total_elements
# 随机选择新的维度并且保证元素数量不变
while remaining_elements > 1 and len(dimensions) < len(input_shape):
divisor = np.random.choice(divisors)
dimensions.append(divisor)
remaining_elements //= divisor
divisors = [d for d in divisors if remaining_elements % d == 0]
if remaining_elements > 1:
dimensions.append(remaining_elements)
np.random.shuffle(dimensions)
return tuple(dimensions)
def generate_reshape_params(tensor):
return generate_random_reshape(tensor.shape)
def random_transpose_params(tensor):
return random.sample(range(tensor.dim()), 2)
def generate_repeat_params(input_shape):
while True:
repeats = [random.randint(1, 4) for _ in input_shape]
if reduce(mul, [dim * repeat for dim, repeat in zip(input_shape, repeats)], 1) <= max_tensor_elements:
return tuple(repeats)
def generate_expand_params(input_shape):
expanded_shape = []
while True:
expanded_shape = [random.randint(min(2,dim), dim*2) if dim == 1 else dim for dim in input_shape]
if reduce(mul, expanded_shape, 1) <= max_tensor_elements:
break
return expanded_shape
def generate_random_operator_parameters(input_shape, operator, input_tensor):
if operator == 'unsqueeze':
return (random.randint(0, len(input_shape) - 1),)
if operator == 'repeat':
return generate_repeat_params(input_shape)
if operator == 'permute':
return random.sample(range(len(input_shape)), len(input_shape))
if operator == 'transpose':
return random_transpose_params(input_tensor)
if operator in ['reshape',"view"]:
return generate_reshape_params(input_tensor)
if operator == 'expand':
return generate_expand_params(input_shape)
if operator == 'matmul':
if input_tensor.dim() == 1:
return ()
return (input_tensor.size(-1), random.randint(1, max_size))
if operator in ['contiguous','mul']:
return ()
if operator == 'concat':
return (random.randint(0, len(input_shape) - 1),)
def execute_operator(input_tensor, operator, operator_parameters):
if operator == 'unsqueeze':
return input_tensor.unsqueeze(*operator_parameters)
if operator == 'repeat':
return input_tensor.repeat(operator_parameters)
if operator == 'permute':
return input_tensor.permute(operator_parameters)
if operator == 'transpose':
return input_tensor.transpose(*operator_parameters)
if operator == 'reshape':
return input_tensor.reshape(operator_parameters)
if operator == 'view':
return input_tensor.view(operator_parameters)
if operator == 'expand':
return input_tensor.expand(operator_parameters)
if operator == 'contiguous':
return input_tensor.contiguous()
if operator == 'matmul':
if input_tensor.dim() ==1:
return input_tensor
other = torch.randn(*operator_parameters).to(input_tensor.device).type_as(input_tensor)
return torch.matmul(input_tensor, other)
if operator == 'mul':
return input_tensor * input_tensor
if operator == 'concat':
return torch.cat((input_tensor, input_tensor), dim=operator_parameters[0])
def main():
for seed in range(2):
random.seed(seed)
np.random.seed(seed)
torch.random.manual_seed(seed)
for i in range(10):
input_shape = generate_random_shape(random.randint(2, 5))
input_tensor = generate_random_input(input_shape)
operator = generate_random_operator(input_shape)
operator_parameters = generate_random_operator_parameters(input_shape, operator, input_tensor)
output_tensor = execute_operator(input_tensor, operator, operator_parameters)
print(f"seed:{seed:03d} seq:{i:02d} {operator:<10} input:{str(input_shape):<32} param:{str(operator_parameters):<32} output:{str(output_tensor.shape):<32}")
print(output_tensor.cpu().numpy().reshape(-1)[:8])
torch.cuda.empty_cache()
if __name__ == '__main__':
main()输出
bash
seed:000 seq:00 repeat input:(7, 42, 26, 36, 56) param:(1, 1, 1, 1, 1) output:torch.Size([7, 42, 26, 36, 56])
seed:000 seq:01 view input:(248, 227, 276) param:(92, 908, 186) output:torch.Size([92, 908, 186])
seed:000 seq:02 view input:(18, 21, 51, 32, 17) param:(17, 4536, 136) output:torch.Size([17, 4536, 136])
seed:000 seq:03 reshape input:(2548, 3565) param:(644, 65, 217) output:torch.Size([644, 65, 217])
seed:000 seq:04 reshape input:(46, 42, 14, 57, 7) param:(28, 266, 3, 483) output:torch.Size([28, 266, 3, 483])
seed:000 seq:05 contiguous input:(222, 100, 597) param:() output:torch.Size([222, 100, 597])
seed:000 seq:06 view input:(15, 27, 56, 8, 59) param:(3, 3, 20160, 1, 59) output:torch.Size([3, 3, 20160, 1, 59])
seed:000 seq:07 view input:(1461, 1161) param:(188469, 9) output:torch.Size([188469, 9])
seed:000 seq:08 reshape input:(19, 29, 19, 17, 54) param:(31407, 1, 3, 17, 6, 1) output:torch.Size([31407, 1, 3, 17, 6, 1])
seed:000 seq:09 transpose input:(12, 126, 46, 157) param:[2, 3] output:torch.Size([12, 126, 157, 46])
[-0.581 0.568 1.187 2.46 -0.1392 -0.3362 0.2076 -0.662 ]
seed:001 seq:00 view input:(119, 354, 236) param:(4, 1, 17, 146202) output:torch.Size([4, 1, 17, 146202])
seed:001 seq:01 reshape input:(60, 961, 178) param:(3, 3421160) output:torch.Size([3, 3421160])
seed:001 seq:02 expand input:(16, 10, 34, 37, 58) param:[16, 10, 34, 37, 58] output:torch.Size([16, 10, 34, 37, 58])
seed:001 seq:03 concat input:(12, 44, 12, 26, 55) param:(1,) output:torch.Size([12, 88, 12, 26, 55])
seed:001 seq:04 expand input:(48, 9, 28, 20, 68) param:[48, 9, 28, 20, 68] output:torch.Size([48, 9, 28, 20, 68])
seed:001 seq:05 repeat input:(16, 16, 162, 233) param:(1, 1, 1, 1) output:torch.Size([16, 16, 162, 233])
seed:001 seq:06 expand input:(25, 426, 19, 63) param:[25, 426, 19, 63] output:torch.Size([25, 426, 19, 63])
seed:001 seq:07 permute input:(153, 153, 380) param:[2, 1, 0] output:torch.Size([380, 153, 153])
seed:001 seq:08 permute input:(3091, 1445) param:[1, 0] output:torch.Size([1445, 3091])
seed:001 seq:09 mul input:(142, 254, 388) param:() output:torch.Size([142, 254, 388])
[3.31 0.3372 0.2354 0.1373 0.594 2.326 0.7344 2.16 ]