[AMP] support OD level for static #53768
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
❌ The PR is not created using PR's template. You can refer to this Demo. |
AnnaTrainingG
force-pushed
the
static_amp_OD
branch
from
fa29c2f
to
2b78304
Compare
AnnaTrainingG
changed the title
AnnaTrainingG
changed the title
| amp_lists = AutoMixedPrecisionLists(dtype=amp_dtype) | ||
|
|
||
| amp_lists.white_list = {"conv2d", "matmul_v2"} | ||
| amp_lists.black_list = amp_lists.all_list - amp_lists.white_list |
There was a problem hiding this comment.
880～881和887～888可以合并下，OD level下，直接用默认名单。
884行的warning还需要一个条件，当用户设置了amp_lists，才需要给warning提示。
| @@ -116,13 +116,13 @@ def _get_sys_unsupported_list(dtype): | |||
| } | |||
| sys_unsupported_list -= supported_fp16_list | |||
|
|
|||
| return device, sys_unsupported_list | |||
| return device, sys_unsupported_list, sys_unsupported_list | |||
There was a problem hiding this comment.
这里为啥返回了2个sys_unsupported_list
|
增加一个单测到test/amp/test_amp_api.py中，里面目前有动态图的单测，可以加个静态图的，静态图可以参考test/amp/test_amp_promote.py用collect_operator_stats收集OP信息。 |
| @@ -101,7 +101,7 @@ def _get_sys_unsupported_list(dtype): | |||
| device = 'NPU' | |||
| else: | |||
| device = 'GPU' | |||
| _, _, sys_unsupported_list = core.op_supported_infos(device, var_type) | |||
| _, _, sys_supported_list = core.op_supported_infos(device, var_type) | |||
There was a problem hiding this comment.
我看op_supported_infos这个接口的返回值是all_ops、supported_ops和unsupported_ops
| dtype = get_low_precision_dtypestr(dest_type) | ||
| amp_lists = AutoMixedPrecisionLists(dtype) | ||
|
|
||
| amp_lists.white_list = {"conv2d", "matmul_v2"} |
There was a problem hiding this comment.
这里其实应该是用框架默认白名单，也就是617行得到的amp_list中的白名单实际就是默认名单。需要修改的就是黑名单。
| level, | ||
| ) | ||
|
|
||
| def test_static_amp_o1(self): |
| use_promote=True, | ||
| expected_op_calls=expected_fp16_calls, | ||
| ) | ||
| paddle.disable_static() |
There was a problem hiding this comment.
这个单测可能存在一个问题就是build_conv_model得到的program，在O1下大概也是87行的结果，可能会看不出差异。你可以把网络换成25行那个动态图的模型写法，这样能确保O1和OD是有差异的
|
ok |
PR types
Others
PR changes
APIs
Description
Pcard-70458