核函数NPU上板精度验证

本场景以FlashAttentionScore算子为例，假设输入数据和标杆数据是用户自行提供的bin文件，核函数NPU上板调测过程如下。请根据自身实际情况，按需修改示例代码。

import torch
import numpy as np
import ascendebug

# 设置和清理日志文件
ascendebug.set_log_file('ops_adv.log', clean=True)
# 根据实际情况设置路径
CANN_INSTALL_PATH = "/usr/local/Ascend/"
DATA_PATH = '/user_data_path/'
REPO_PATH = "ops_adv_code_path/"

# 1.导入输入/标杆数据，构建算子信息
debug_op = ascendebug.create_debug_op('FlashAttentionScore', 'MixCore', 'Ascendxxx') \
    .custom_input('query', 'float16', [24, 144, 1280], os.path.join(DATA_PATH, 'q.bin')) \
    .custom_input('key', 'float16', [24, 144, 1280], os.path.join(DATA_PATH, 'k.bin')) \
    .custom_input('value', 'float16', [24, 144, 1280], os.path.join(DATA_PATH, 'v.bin')) \
    .custom_input('real_shift', 'float16', None, None, ['optional']) \
    .custom_input('drop_mask', 'uint8', [1244160], os.path.join(DATA_PATH, 'drop_mask.bin'), ['optional']) \
    .custom_input('padding_mask', 'float16', None, None, ['optional']) \
    .custom_input('atten_mask', 'bool', None, None, ['optional']) \
    .custom_input('prefix', 'int64', None, None, ['optional']) \
    .custom_input('actual_seq_qlen', 'int64', None, None, ['optional']) \
    .custom_input('actual_seq_kvlen', 'int64', None, None, ['optional']) \
    .custom_input('q_start_idx', 'int64', None, None, ['optional']) \
    .custom_input('kv_start_idx', 'int64', None, None, ['optional']) \
    .custom_output('softmax_max', 'float32', [24, 20, 144, 8], None) \
    .custom_output('softmax_sum', 'float32', [24, 20, 144, 8], None) \
    .custom_output('softmax_out', 'float16', [24, 20, 144, 144], None) \
    .custom_output('attention_out', 'float16', [24, 20, 144, 64], os.path.join(DATA_PATH, 'attention_out.bin')) \
    .attr('scale_value', 'float', 1.0) \
    .attr('keep_prob', 'float', 0.8) \
    .attr('pre_tockens', 'int', 2147483647) \
    .attr('next_tockens', 'int', 2147483647) \
    .attr('head_num', 'int', 20) \
    .attr('input_layout', 'string', 'BSH') \
    .attr('inner_precise', 'int', 0)

# 2.创建调试对象并初始化工作空间
op_executor = ascendebug.create_op_executor(debug_op=debug_op, install_path=CANN_INSTALL_PATH)

# 3.调用Tiling调测接口（可选，若已有Tiling bin文件可跳过本步骤）
# 如需本地调试Tiling，推荐使用方式1，否则使用方式2
# 方式1：基于本地ops_adv仓编译Tiling so，再执行Tiling计算
tiling_so = op_executor.compile_ops_adv_tiling(REPO_PATH)    
tiling_info = op_executor.run_tiling(tiling_so)
# 方式2：工具自动从CANN包获取Tiling so并进行Tililng计算
# tiling_info = op_executor.run_ops_adv_tiling()
print(tiling_info.tiling_bin, tiling_info.tiling_workspace, tiling_info.block_num, tiling_info.tiling_key)

# 4.调用NPU编译接口
compile_npu_options = ascendebug.CompileNpuOptions()
name, kernel_file, extern = op_executor.compile_ops_adv_npu(REPO_PATH, tiling_info.tiling_key, compile_npu_options)

# 5.调用NPU运行接口，完成上板精度比对
run_npu_options = ascendebug.RunNpuOptions()
compile_info = ascendebug.NpuCompileInfo(syncall=extern['cross_core_sync'], task_ration=extern['task_ration'])
op_executor.run_npu(kernel_file, run_npu_options, npu_compile_info=compile_info, tiling_info=tiling_info)

算子在NPU板端运行调测的精度比对结果示例可以参见“NPU调测功能 > 调测产物”。