[PIR+CINN]Part-2 Pybind IrParser.ParseProgram and Polish UT into check_run #59449
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
add unittest fix UT not take effect [PIR+CINN]Pybind IrParser.ParseProgram and Polish UT into check_run fix cmake flasgs remove VLOG fix code comment
Aurelius84
force-pushed
the
ir_cinn_test1
branch
from
f7e9694
to
85eb60c
Compare
| self.assertEqual(len(outs), 0) | ||
|
|
||
|
|
||
| # class TestSaveInferProg(TestSaveFwdBwdProg): |
There was a problem hiding this comment.
这个依赖动转静SOT一个BUG Fix 的PR，待依赖PR合入后，单独打开
Aurelius84
changed the title
Aurelius84
force-pushed
the
ir_cinn_test1
branch
from
1a9b4fa
to
0c4ad9f
Compare
| need_fetch_info = [] | ||
| for i, fetch_var in enumerate(fetch_targets): | ||
| if isinstance(fetch_var, str): | ||
| if fetch_var not in fetch_info[1]: | ||
| raise Exception( | ||
| f"Found fetch_target[{i}] is type(str) and doesn't have fetch op." | ||
| ) | ||
| elif fetch_var not in fetch_info[0]: | ||
| need_fetch_info.append(fetch_var) | ||
|
|
||
| return need_fetch_info |
There was a problem hiding this comment.
这段逻辑删除了对之前fetch op数量报错以及警告的检查，改为补充缺失的fetch op，会不会存在这样的问题，用户多次调用executor run，后续run可能和前边的run并没有太大关系，但是fetch op依然滞留到了program里，导致跑后续run的过程中，依然会fetch并不需要fetch的数据，由于fetch会进行copy操作，会造成隐形性能开销，这里是不是还是拦截或者提示一下相关信息比较好
There was a problem hiding this comment.
之前的逻辑这里也会给program添加缺失的fetch_ops，这个行为是没有改变的，而且之前的add_pir_fetch_ops是会对所有的fetch_list 添加，即使之前部分fetch_var已经有fetch_op了。
关于多次run且彼此之间的fetch_list不一样的问题，其实在run()接口里应该要先clone program，然后add_feed_fetch_ops，然后缓存起来。后续优先根据feed/fetch 信息来查缓存program，这样每次run就是独立的，互不影响。
这个cache策略是已有的，在上层做的，add_feed_fetch_ops 是不需要关心这个缓存逻辑的
PR types
New features
PR changes
Others
Description
Pcard-67164
前置PR见：#59353
paddle.pir.parse_program(str)接口，支持前端从string里反序列化得到Program对象test_subgraph_exporter单测，由静态check优化为加载file文件，反序列化并执行Program