Closed GreatV closed 1 year ago
任务编号 | 任务内容 | 认领人 | PR链接 |
---|---|---|---|
4 | 清理 is_compiled_with_mlu 相关代码 |
@KimBioInfoStudio | #52378 |
8 | 清理 is_compiled_with_npu 相关代码 |
@KimBioInfoStudio | #52385 |
11 | 清理 paddle_build.sh 脚本中所有的 WITH_MLU WITH_ASCEND* 关键字 |
@KimBioInfoStudio | #52373 |
12 | 清理 WITH_CNCL 相关代码 | @KimBioInfoStudio | #52434 |
14 | 清理 dockerfile | @KimBioInfoStudio | #52474 |
./python/paddle/distributed/fleet/layers/mpu/mp_ops.py
有残留./python/paddle/distributed/fleet/meta_optimizers/common.py
有残留./python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py
有残留./python/paddle/distributed/fleet/utils/hybrid_parallel_inference.py
./python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_utils.py
./python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_utils.py
./python/paddle/fluid/tests/unittests/c_embedding_op_base.py
./paddle/fluid/pybind/pybind.cc
大概率失去 profiling 功能
./python/paddle/nn/functional/conv.py
./python/paddle/fluid/dygraph/varbase_patch_methods.py
./python/paddle/fluid/framework.py
./python/paddle/fluid/__init__.py
./python/paddle/fluid/optimizer.py
./python/paddle/fluid/executor.py
./python/paddle/amp/grad_scaler.py
./python/paddle/static/nn/common.py
./python/paddle/static/amp/decorator.py
./python/paddle/static/amp/fp16_lists.py
./python/paddle/static/amp/amp_nn.py
./python/paddle/device/__init__.py
@GreatV distributed 相关的部分代码不能直接删除,需要修改, 目前 PaddleCustomDevice 一些 distributed 相关功能并没有完全 decouple cc @luotao1 @qili93
CC: @KevinYuK @yangulei
seems cncl is mlu ccl, also need clean up @GreatV CC: @luotao1 @qili93
@KimBioInfoStudio 确实缺少了WITH_CNCL
的清理
认领 2
@KimBioInfoStudio 确实缺少了
WITH_CNCL
的清理
@GreatV 同理, 还有 hccl
@KimBioInfoStudio hccl这些应该会在清理 PADDLE_WITH_ASCEND_CL
被清理掉吧
paddle/fluid/platform/device/mlu/mlu_collective_helper.cc
paddle/fluid/platform/device/mlu/cncl_helper.h
paddle/fluid/platform/device/mlu/device_context_allocator.h
paddle/fluid/platform/device/mlu/device_context_test.cc
第一项任务中,这四个文件是不是也要删除
paddle/fluid/platform/device/mlu/mlu_collective_helper.cc paddle/fluid/platform/device/mlu/cncl_helper.h paddle/fluid/platform/device/mlu/device_context_allocator.h paddle/fluid/platform/device/mlu/mlu_collective_helper.cc
第一项任务中,这四个文件是不是也要删除
应该是的
任务编号 | 任务内容 | 认领人 | PR链接 |
---|---|---|---|
3 | 清理 WITH_MLU 相关代码 |
@GreatV | #52546 |
paddle/fluid/platform/device/mlu/mlu_collective_helper.cc paddle/fluid/platform/device/mlu/cncl_helper.h paddle/fluid/platform/device/mlu/device_context_allocator.h paddle/fluid/platform/device/mlu/device_context_test.cc
第一项任务中,这四个文件是不是也要删除
是的,现在 mlu 走的是 decvice / custom device, 所以 mlu 整个 dir 都可以删除
@KimBioInfoStudio hccl这些应该会在清理
PADDLE_WITH_ASCEND_CL
被清理掉吧
应该不行,有些代码是在 Python 里面,WITH_ASCEND_CL 大部分都是在 cmake 或者 cc 里面
should we also mv the tests at python/paddle/fluid/tests/unittests/mlu
and python/paddle/fluid/tests/unittests/npu
to the CustomDevice repo?
paddle/fluid/platform/device/mlu/mlu_collective_helper.cc paddle/fluid/platform/device/mlu/cncl_helper.h paddle/fluid/platform/device/mlu/device_context_allocator.h paddle/fluid/platform/device/mlu/device_context_test.cc
第一项任务中,这四个文件是不是也要删除
是的,paddle/fluid/platform/device/mlu 删除即可
should we also mv the tests at
python/paddle/fluid/tests/unittests/mlu
andpython/paddle/fluid/tests/unittests/npu
to the CustomDevice repo?
just clean up
任务编号 | 任务内容 | 认领人 | PR链接 |
---|---|---|---|
13 | 清理 HCCL 相关代码 | @Liyulingyue | #52535 |
@luotao1 Pls add a subtask for docs change thx
@luotao1 @qili93 @ronny1996 do we need rem graphcore ipu this time?
@luotao1 @qili93 @ronny1996 do we need rem graphcore ipu this time?
ipu不需要删除哈
@luotao1 Pls add a subtask for docs change thx
docs repo暂时不用删哈
目前 develop branch 还能通过下面的方法找到部分 device 相关残留 | CMD | f1b6a76b56 |
e0c14fdf25 |
06ecc6d2d3 |
adc2b7459f |
8f5eae47cd |
b428e8f35c |
---|---|---|---|---|---|---|---|
grep -rni npu \| grep -vi -P "input\|onpush\|onpull\|unpub" |
744 | 744 | 677 | 649 | 568 | 255 | |
grep -rni mlu |
131 | 35 | 35 | 3 | 3 | 0 | |
grep -rni hccl |
8 | 8 | 0 | 0 | 0 | 0 | |
grep -rni ascend\|grep -vi "ascending" |
107 | 107 | 36 | 23 | 23 | 23 | |
grep -rni cncl |
0 | 0 | 0 | 0 | 0 | 0 |
任务编号 | 任务内容 | 认领人 | PR链接 |
---|---|---|---|
16 | 清理 ASCEND 相关代码 | @GreatV | #53046 |
@ronny1996 @qili93 seems we should remove cnpai for mlu profiler, add redesign a new api for all custom_device cc @luotao1
@luotao1 @Liyulingyue in latest develop adc2b7459f
there are still 3 mlu
keyword in code
tools/infer_prune_patches/op_registry.h.patch:97: REGISTER_OP_KERNEL(op_type, MLU, ::paddle::platform::MLUPlace, __VA_ARGS__)
tools/infer_prune_patches/op_registry.h.patch:131: #define REGISTER_OP_MLU_KERNEL_FUNCTOR(op_type, ...) \
tools/infer_prune_patches/op_registry.h.patch:133: op_type, MLU, ::paddle::platform::MLUPlace, DEFAULT_TYPE, \
in latest develop adc2b7459f there are still 3 mlu keyword in code
@KimBioInfoStudio @Liyulingyue 这个patch先不用改,是推理侧为了裁剪库体积写的脚本,不是简单删除即可。
in latest develop adc2b74 there are still 3 mlu keyword in code
@KimBioInfoStudio @Liyulingyue 这个patch先不用改,是推理侧为了裁剪库体积写的脚本,不是简单删除即可。
Okay, 这个应该是需要重新取diff了,之前一致有个疑惑,删除custom device 代码会导致infer lib size 增加,现在大致猜测是因为 diff apply 不上去导致的
https://github.com/PaddlePaddle/Paddle/search?q=npu 还剩下最后一项【清理 NPU 相关代码】,还有不少npu代码。
@KimBioInfoStudio @Liyulingyue 整个 tools/infer_prune_patches
目录都可以删除,https://github.com/PaddlePaddle/Paddle/pull/53193#issuecomment-1534447963 已经使用另一种方式在裁剪预测库了。
CMD | b428e8f35c |
---|---|
grep -rni npu \| grep -vi -P "input\|onpush\|onpull\|unpub" |
255 |
grep -rni mlu |
0 |
grep -rni hccl |
0 |
grep -rni ascend\|grep -vi "ascending" |
23 |
grep -rni cncl |
0 |
感谢 @KimBioInfoStudio 统计
grep -rni ascend|grep -vi "ascending"
ascend 关键字已检查完毕,剩余的都不需要删除
grep -rni npu | grep -vi -P "input|onpush|onpull|unpub"
昇腾和寒武纪相关代码退场已全部完成,感谢参与的小伙伴们!
感谢 @GreatV 拆分任务形成 Tracking issue; @KimBioInfoStudio 多次统计需要清理的内容; @Liyulingyue 完成最后的收尾工作,以及所有的小伙伴们!
按 merge 的时间顺序,排名不分先后: @KimBioInfoStudio (5) @mrcangye (2) @jjyaoao (3) @Young-Flash (1) @Liyulingyue (3) @engineer1109 (1) @GreatV (2)
欢迎继续参与快乐开源的其他任务!
一、 项目说明 用于跟踪 昇腾和寒武纪相关代码退场 并记录该计划下各个子任务的计划和进度。
二、 主要工作 (18/18)
paddle/fluid/platform/device/mlu
目录✅(2023/3/31)paddle/fluid/operators/*_mlu.*
文件✅(2023/4/3)WITH_MLU
相关代码✅(2023/4/13)is_compiled_with_mlu
相关代码✅(2023/4/3)paddle/fluid/operators/*_npu.*
、清理WITH_ASCEND_INT64
相关代码✅(2023/4/11)/paddle/fluid/platform/device/npu
目录✅(2023/4/1)WITH_ASCEND_CL PADDLE_WITH_ASCEND_CL WITH_ASCEND_CXX11
相关代码✅(2023/4/13)is_compiled_with_npu
相关代码✅(2023/4/6)PADDLE_WITH_ASCEND
相关代码✅(2023/4/13)WITH_ASCEND
相关代码✅(2023/4/10)paddle_build.sh
脚本中所有的WITH_MLU
WITH_ASCEND*
关键字✅(2023/3/31)CNCL
相关代码✅(2023/4/17)HCCL
相关代码✅(2023/4/20)is_mlu_place
is_npu_place
is_npu_pinned_place
SupportNPU
SupportMLU
相关代码✅(2023/4/28)需要清理的具体文件可参考此链接。