[pir_save_load] add pir for test_jit_save_load.py #63958
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
你的PR提交成功，感谢你对开源项目的贡献! |
tianshuo78520a
approved these changes
Apr 30, 2024
XieYunshen
previously approved these changes
Apr 30, 2024
changeyoung98
previously approved these changes
Apr 30, 2024
xiaoguoguo626807
dismissed stale reviews from changeyoung98 and XieYunshen
via
4e5d6e2
XieYunshen
previously approved these changes
May 6, 2024
XieYunshen
approved these changes
May 7, 2024
yinfan98
pushed a commit
to yinfan98/Paddle
that referenced
this pull request
May 7, 2024
* add jit load.train * modify backward program lost * modify * combine eval and train * modify 8 case of jit.save.load * modify jit_save_load case * rename jit_save_load * change name all * modify timeout * modify new case * modify TestJitSaveLoadMultiMethods * modify cpu tensor no holder bug
yinfan98
pushed a commit
to yinfan98/Paddle
that referenced
this pull request
May 7, 2024
* add jit load.train * modify backward program lost * modify * combine eval and train * modify 8 case of jit.save.load * modify jit_save_load case * rename jit_save_load * change name all * modify timeout * modify new case * modify TestJitSaveLoadMultiMethods * modify cpu tensor no holder bug
yinfan98
pushed a commit
to yinfan98/Paddle
that referenced
this pull request
May 7, 2024
add int4_1 int4_2 FLAGS_logging_pir_py_code (PaddlePaddle#63981) * FLAGS_logging_pir_py_code * FLAGS_logging_pir_py_code_dir --------- Co-authored-by: jiahy0825 <[email protected]> [Cleanup] Remove Flake8 config in `.editorconfig` (PaddlePaddle#64027) 【PIR Dist Op Reg No.19】 reg pull_box_sparse (PaddlePaddle#62982) * fix * fix * fix * fix * fix * fix * add test * add * fix * fix * add out * fix * codestyle * fix * fix backward * merge [Dy2St][PIR] Hold backward program in GradNode (PaddlePaddle#63694) Co-authored-by: xiongkun <[email protected]> Co-authored-by: Nyakku Shigure <[email protected]> split test.cmake: add new test_cases.cmake (PaddlePaddle#64007) [PIR] Support sparse_slice and sparse_sum in pt (PaddlePaddle#64009) * support sparse_slice and sparse_sum in pt * support sparse_slice and sparse_sum in pt * support sparse_slice and sparse_sum in pt option for WITH_CPP_TEST (PaddlePaddle#63896) * option for WITH_CPP_TEST * fix * Fix * Fix [PIR] Fix `attributes_num` of `SliceArrayOp` (PaddlePaddle#64013) [Dy2St] Use `full_graph=True` outside dy2st uts (part1) (PaddlePaddle#64058) [Dy2St] Use `full_graph=True` outside dy2st uts (part2) (PaddlePaddle#64059) fix typo (PaddlePaddle#64060) Co-authored-by: jiahy0825 <[email protected]> update (PaddlePaddle#64042) Replace paddle/fluid/platform/device/gpu/gpu_dnn.h (PaddlePaddle#63819) * Fix * Fix * Fix Clean lookup_table_v2_op.h lookup_table_v2_op.cu (PaddlePaddle#64020) * Fix * ci refine GetTensorListFromArgs (PaddlePaddle#64045) Revert "【Hackathon 6th Fundable Projects 3 No.60】Remove fluid operator chunk_…" (PaddlePaddle#64050) This reverts commit 88b1a6e. [Prim][PIR] support floor_divide op forward in prim pir (PaddlePaddle#64023) * floor-div-dev * update test [CINN] Reconstruct shape_analysis (PaddlePaddle#63790) * reconstruct shape_analysis * fix input value shape infer * fix merge bugs * fix concat and gather op InferSymbolicShape * fix merge bug * fix value_to_shape_or_data hash error and add some checks * fix set shape for null value * fix group op lazy infer * add IsStaticShape check * fix merge bug * support static dim check and set for VectorType * change auto to detail type [XPU] fix bugs in processing of attention_mask and fix_seed_offset on XPU (PaddlePaddle#64003) * [XPU] fix segmentfault caused by setting fix_seed_offset on XPU * cast attention_mask to float32 when necessary fix merge bug (PaddlePaddle#64069) 【Fix PIR Unittest No.125、147、481】Fix some 0D uts in PIR mode (part1) (PaddlePaddle#64064) [Prim][VJP]support autogen to remove unused composite in .yaml (PaddlePaddle#64054) * support autogen to remove unused composite in .yaml * fix bug [PIR] Fix typo `set_pit_tests_properties` -> `set_pir_tests_properties` (PaddlePaddle#64063) [Dy2St] Use `full_graph=True` outside dy2st uts (part3) (PaddlePaddle#64066) [PIR save/load] Open more tests for paddle.save and paddle.load (PaddlePaddle#64044) * open more tests for paddle.save and paddle.load * fix API Improvement for paddle.nn.functional.group_norm and paddle.nn.GroupNorm (PaddlePaddle#63881) * update group_norm * update trt plugin * update trt plugin * fix trt plugin * fix trt plugin * fix test * fix test * fix ci windows inference * update kernel function names and add v2 test * fix * fix fp16 test Revert "【Hackathon 6th Fundable Projects 3 No.81】Remove fluid operators ctc_a…" (PaddlePaddle#64049) This reverts commit 2134ead. Clean paddle/fluid/operators/fused/attention_layer_norm.h (PaddlePaddle#64051) * Fix * Fix Replace operators::math to phi::math in fluid/operators (PaddlePaddle#63854) [CINN]Clean usless loop_reorder_aligment tactic (PaddlePaddle#63998) * [CINN]Clean usless loop_reorder_aligment tactic * fix source 【Hackathon 6th Fundable Projects 3 No.396】fluid operator yolo_box_head (PaddlePaddle#63783) * Fix * Fix * Fix * Fix * Fix 【Hackathon 6th Fundable Projects 3 No.240】fluid operator moe (PaddlePaddle#63929) 【Hackathon 6th Fundable Projects 3 No.82】fluid operator cudnn_lstm (PaddlePaddle#63936) * Fix * Fix * Fix * Fix [CINN] Remove useless log (PaddlePaddle#64052) [pir_save_load] add pir for test_jit_save_load.py (PaddlePaddle#63958) * add jit load.train * modify backward program lost * modify * combine eval and train * modify 8 case of jit.save.load * modify jit_save_load case * rename jit_save_load * change name all * modify timeout * modify new case * modify TestJitSaveLoadMultiMethods * modify cpu tensor no holder bug Flashattention support qkvpacked and varlen (PaddlePaddle#63289) * Flashattention support qkvpacked and varlen * fix codestyle * fix codestyle * FlashAttention kvReduceGQA Performance Optimization * Fix problem with windows * code clean * update third_party/flashattn * update errormsg and docs * update api * update doc * update doctest * update doc, test=document_fix * update doc, test=document_fix * Update python/paddle/nn/functional/flash_attention.py Co-authored-by: zachary sun <[email protected]> * Update python/paddle/nn/functional/flash_attention.py Co-authored-by: zachary sun <[email protected]> * update doc --------- Co-authored-by: zachary sun <[email protected]> 【PIR Dist Op Reg No.20】 reg global_gather (PaddlePaddle#63867) * reg global_gather * reg global_gather * reg_global_gather * fix * fix * fix * fix conflict * fix conflict * Update ops_api_gen.py * Update ops_api_gen.py Fix backward program kwargs error when process inplace value (PaddlePaddle#63939) 【Hackathon 6th No.35】support kwargs for recompute when use_reentrant == True fix (PaddlePaddle#63880) * support kwargs for recompute when use_reentrant == True * recover third party merge main lint delete printf change flash attn version
co63oc
pushed a commit
to co63oc/Paddle
that referenced
this pull request
May 10, 2024
* add jit load.train * modify backward program lost * modify * combine eval and train * modify 8 case of jit.save.load * modify jit_save_load case * rename jit_save_load * change name all * modify timeout * modify new case * modify TestJitSaveLoadMultiMethods * modify cpu tensor no holder bug
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR Category
others
PR Types
others
Description
pcard-67164
适配test_jit_save_load.py 25/28个class 测试.
a. 解决装饰器问题，保证外部环境是动态图，save中切换静态图test_with_dygraph_pir。 todo ：paddle.load 测试。
b. 多输出精度不对，原因是save 时加scale 加到了一个变量上，导致剪枝了一层linear的parameter
c. input_specs_compatible(flatten(input_spec), flatten(self._function_spec.input_spec)在新IR下判断为false,因为datatype判断不同，input_spec 是DataType.FLOAT32, 原来的是paddle.float32, 原因是在layer中加动转静装饰器，在切换pir 环境前进行动转静，导致数据类型不匹配。
d. feed_var 是input_spec的时候需要找到对应的dataop拿到value.
e. 补充pirtranslatelayer 的input_spec,output_spec的接口。
f. net.forward.input[0].shape 旧IR下返回了（）， pir下是[]
g. 适配save根据output_spec 进行输出过滤。对input_spec 进行处理
h.pir下output_spec 传递的是tensor 无法进行检查，因为value 没有名称，无法通过name检查，只能直接判断var 是否在 output中，var 需要是value, 只能通过动转静之后的网络中找到value.(添加了check逻辑）
i. 多次load 需要给parameter重命名，且parameter如果需要训练,用优化器更新，返回的是EagerParamBase， 否则是tensor
j. 适配combine_param save 接口中相关代码。
k. TestJitSaveLoadMultiMethods 原始动转静save 路径有问题导致load 不出其他层； 输出=输入的情况下，value拥有两个名字，执行器获取shadow_output 的名字得不到正确的tensor, 会报错tensor no holder。
TODO
TestJitSaveLoadSaveWithoutRunning 含控制流
TestJitSaveLoadFinetuneLoad 原始动转静报错待修复， 含控制流