【Flexcheckpoint】add_get_var_mapping_chain_macro by zty-king · Pull Request #76013 · PaddlePaddle/Paddle

zty-king · 2025-10-23T15:08:26Z

PR Category

User Experience

PR Types

New features

Description

当前遇到如下aoa配置时，无法正常处理：

aoa_statements = [
            "layers.0.gate_up_fused_proj.weight -> temp_var, fused_ffn \n",
            "temp_var^T -> new_name_layers.0.gate_up_fused_proj.weight \n",
        ]
aoa_statements = [
            "layers.0.gate_up_fused_proj.weight^T -> temp_var \n",
            "temp_var -> new_name_layers.0.gate_up_fused_proj.weight,fused_ffn\n",
        ]

其主要原因在于，fused_ffn,fused_qkv相关macro均需要，且只能从src_key和dst_key中获取切分信息，然而当->左边或者右边为中间变量时，此时的切分信息就丢失了，无法获取，因此新开发了一个macro，通过解析aoa_statements，建立temp_var->src以及temp_var->dst的映射信息。
input_var在model_state和opt_state同时存在时，dtype会被覆盖，导致错误，进行修正
merge_sharded_state_dict当前不支持单卡运行，给予支持

paddle-bot · 2025-10-23T15:08:34Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

codecov-commenter · 2025-10-23T19:51:58Z

Codecov Report

❌ Patch coverage is 95.40230% with 4 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@b51d1da). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...ddle/distributed/flex_checkpoint/aoa/aoa_engine.py	88.23%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #76013   +/-   ##
==========================================
  Coverage           ?   95.40%           
==========================================
  Files              ?        4           
  Lines              ?       87           
  Branches           ?        0           
==========================================
  Hits               ?       83           
  Misses             ?        4           
  Partials           ?        0

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

xingmingyyj · 2025-10-24T12:00:02Z

+
+        For example:
+        - reverse=False: temp_var -> dst_key
+        - reverse=True: temp_var -> src_key


reverse=True: temp_var -> src_key 如何理解

这个意思是，记录temp_var是从哪个src_key转换过来了，记录它的来源信息，则当以下情景时：
aoa_statements = [
"layers.0.gate_up_fused_proj.weight^T -> temp_var \n",
"temp_var -> new_name_layers.0.gate_up_fused_proj.weight,fused_ffn\n",
]
我们调用fused_ffn，而此时箭头左侧需要找到temp_var是从哪个src_key转换过来的，从而找到src_key对应的来源信息

xingmingyyj · 2025-10-24T12:05:52Z

+            mapping_dict = self.left_var_to_right_var_mapping
+
+        while current_key in mapping_dict:
+            if current_key in visited:


这里是不是遇到了环，为什么可以直接return

这里想的是，正确的aoa配置一般不会出现环的情况，出现的情景一般如下：
aoa_statements = [
"src_key -> dist_key \n",
]
或
aoa_statements = [
"src_key^T -> A\n",
"A -> A \n",
]
即，src_key 和dist_key同名，或用户设置的中间变量和dst同名(src同理),则会出现dst_key:dst_key（或src_key:src_key）的映射，即传入的本身就是dst_key或src_key，则需要中断防止一直陷入循环，并返回传入的这个key。

xingmingyyj · 2025-10-24T12:13:38Z

+            visited.add(current_key)
+
+            mapped_vars = mapping_dict[current_key]
+            if mapped_vars and len(mapped_vars) > 0:


这里如果是 a-> b,c或者 a,b->c这种场景，切分信息还能往下传递吗

可以的，因为这里保存的value是列表，例如：
a ->b,c
则对于left_var_to_right_var_mapping保存的是:
{
a:[b,c]
}
而对于right_var_from_left_var_mapping保存的是:
{
b:[a]
c:[a]
}
若a是src_key，b，c是中间变量，需要找到b或c来源的src_key的切分信息时，访问right_var_from_left_var_mapping即可映射到；同理若a是中间变量，b，c是dst_key时，直接访问left_var_to_right_var_mapping映射，得到[b,c]列表即可，并且访问列表的第[0]个元素即可，因为二者作为dst_key在统一操作下，则他们在dst中的切分信息也应该是相同的。

xingmingyyj · 2025-10-24T12:20:10Z

+
+
+@macro(name='get_var_mapping_chain_macro', priority=3)
+def get_var_mapping_chain_macro(tokens, expression, context):


所有macro看似被这个macro分成了两类，在这个macro之前是不是所有匹配src 或 dst中的key的macros都要展开？如果是这样是不是在代码上加一下限制比较好，不然容易出错

对的，这里就是通过priority来控制的，让此macro在所有展开操作的macro之后做，priority应该是4，这里当时修改掉了，后续会改正

zty-king · 2025-10-26T05:15:12Z

/re-run all-failed

xingmingyyj · 2025-10-27T02:45:18Z

LGTM

… get_var_mapping_chain_macro

zty-king · 2025-10-28T03:58:09Z

/re-run all-failed

zty-king · 2025-10-28T08:03:22Z

/re-run all-failed

zty-king · 2025-10-28T13:23:32Z

/re-run all-failed

xingmingyyj · 2025-10-28T14:04:59Z

/re-run all-failed

xingmingyyj · 2025-10-28T14:05:13Z

LGTM

zty-king · 2025-10-28T14:07:19Z

/re-run all-failed

zty-king · 2025-10-28T14:15:49Z

/re-run all-failed

… get_var_mapping_chain_macro

zty-king · 2025-10-29T13:26:34Z

/re-run all-failed

zty-king · 2025-10-29T16:03:49Z

/re-run all-failed

… get_var_mapping_chain_macro

zty-king · 2025-10-30T13:53:54Z

/re-run all-failed

zty-king · 2025-10-31T02:09:14Z

/re-run all-failed

* add_get_var_mapping_chain_macro * add note * fix the bug input_vars and resolve_mapping_chain * fix the code style * fit the dtype assert bug * fix the bug * fix the merge_sharded_state_dict bug

…#76252) * 【FlexCheckpoint】fix_the_layer_id_macro (#75556) * fix_the_layer_id_macro * fix the ctest * add expert_id_macro * fix the assert bug * fix the code style * Pr support load hf checkpoint (#75928) * support hf checkpoint fix support cast add id macro fix * add test and fix some bug * fix full param bug * add full param cast test --------- Co-authored-by: xingmingyyj <[email protected]> * 【Flexcheckpoint】add_get_var_mapping_chain_macro (#76013) * add_get_var_mapping_chain_macro * add note * fix the bug input_vars and resolve_mapping_chain * fix the code style * fit the dtype assert bug * fix the bug * fix the merge_sharded_state_dict bug * fix aoa transpose corner case (#76234) --------- Co-authored-by: Tianyu Zheng <[email protected]>

….2 (#76249) * 【FlexCP】merge_sharded_state_dict support distribute merge (#75005) * fix data is nullptr * add dist merge * change test * change test * 【FlexCP】add Skip param param for merge_shard_state_dict (#75061) * fix data is nullptr * add dist merge * change test * change test * add skip optimizer param * [Flex CP]Fix merge_sharded_state_dict with aoa and offload (#75062) * fix merge_state_dict with aoa and offload * add tests * refine * fix * fix * add log * fix * fix * 【FlexCheckpoint】Upgrade some macros and optimize load_state_dict communication (#75282) * upgrad macros and load_state_dict comm task fix fix support 0-d tensor fix balance save and fix * fix test * Add the test about the sharded_state_dict of optimizer (#75067) * fix the share_weight_bug * add note * add the unit test * set the timeout * add more test * Trigger CI rebuild * fix the CmakeLists * handle_missing_edge_cases_in_fc (#75413) * up_grade fc (#75613) fix and add test fix fix fix fix cmakelists add notion * 【FlexCheckpoint】fix_the_layer_id_macro (#75556) * fix_the_layer_id_macro * fix the ctest * add expert_id_macro * fix the assert bug * fix the code style * Pr support load hf checkpoint (#75928) * support hf checkpoint fix support cast add id macro fix * add test and fix some bug * fix full param bug * add full param cast test --------- Co-authored-by: xingmingyyj <[email protected]> * 【Flexcheckpoint】add_get_var_mapping_chain_macro (#76013) * add_get_var_mapping_chain_macro * add note * fix the bug input_vars and resolve_mapping_chain * fix the code style * fit the dtype assert bug * fix the bug * fix the merge_sharded_state_dict bug * fix aoa transpose corner case (#76234) --------- Co-authored-by: xiaoguoguo626807 <[email protected]> Co-authored-by: Chen Zhiyang <[email protected]> Co-authored-by: Tianyu Zheng <[email protected]>

paddle-bot Bot added the contributor External developers label Oct 23, 2025

add_get_var_mapping_chain_macro

efcdf60

zty-king force-pushed the get_var_mapping_chain_macro branch from 6b57f39 to efcdf60 Compare October 23, 2025 15:26

xingmingyyj reviewed Oct 24, 2025

View reviewed changes

add note

2dbe3f9

zty-king added 2 commits October 27, 2025 14:01

fix the bug input_vars and resolve_mapping_chain

a29a160

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

6ab2a37

… get_var_mapping_chain_macro

zty-king force-pushed the get_var_mapping_chain_macro branch from f390065 to 96db255 Compare October 27, 2025 15:41

fix the code style

234ac3e

zty-king force-pushed the get_var_mapping_chain_macro branch from 96db255 to 234ac3e Compare October 27, 2025 15:45

fit the dtype assert bug

22d44b3

zty-king added 2 commits October 29, 2025 08:48

fix the bug

5abaf36

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

aa58dde

… get_var_mapping_chain_macro

zty-king added 2 commits October 29, 2025 16:38

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

59c2ace

… get_var_mapping_chain_macro

fix the merge_sharded_state_dict bug

2293d95

zty-king force-pushed the get_var_mapping_chain_macro branch from fe96e7a to 2293d95 Compare October 30, 2025 08:22

zty-king changed the title ~~add_get_var_mapping_chain_macro~~ 【Flexcheckpoint】add_get_var_mapping_chain_macro Oct 31, 2025

From00 approved these changes Oct 31, 2025

View reviewed changes

From00 merged commit 11fff57 into PaddlePaddle:develop Oct 31, 2025
93 of 96 checks passed

xingmingyyj mentioned this pull request Nov 5, 2025

【Cherry-pick】Cherry-pick FlexCheckpoint PRs from develop to release/3.2 #76249

Merged

xingmingyyj mentioned this pull request Nov 5, 2025

【Cherry-pick】Cherry-pick FlexCheckpoint PRs from develop to fleety_12 #76252

Merged

zty-king deleted the get_var_mapping_chain_macro branch November 23, 2025 13:59



		@macro(name='get_var_mapping_chain_macro', priority=3)
		def get_var_mapping_chain_macro(tokens, expression, context):

Conversation

zty-king commented Oct 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

paddle-bot Bot commented Oct 23, 2025

Uh oh!

codecov-commenter commented Oct 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zty-king commented Oct 26, 2025

Uh oh!

xingmingyyj commented Oct 27, 2025

Uh oh!

zty-king commented Oct 28, 2025

Uh oh!

zty-king commented Oct 28, 2025

Uh oh!

zty-king commented Oct 28, 2025

Uh oh!

xingmingyyj commented Oct 28, 2025

Uh oh!

xingmingyyj commented Oct 28, 2025

Uh oh!

zty-king commented Oct 28, 2025

Uh oh!

zty-king commented Oct 28, 2025

Uh oh!

zty-king commented Oct 29, 2025

Uh oh!

zty-king commented Oct 29, 2025

Uh oh!

zty-king commented Oct 30, 2025

Uh oh!

zty-king commented Oct 31, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

zty-king commented Oct 23, 2025 •

edited

Loading

codecov-commenter commented Oct 23, 2025 •

edited

Loading