GitLab

* [shardformer] embedding support inplace sharding

* [shardformer] linear support inplace sharding

* [shardformer] layernorm support inplace sharding

* [shardformer] qkv support inplace sharding

* [test] update shardformer layer test

* [shardformer] fix shared param sharding

* [shardformer] fix bert policy

* [shardformer] fix bloom policy

* [shardformer] fix llama policy

* [shardformer] fix opt policy

* [shardformer] fix t5 policy

* [shardformer] fix fused qkv linear

* [shardformer] fix bugs

* force sync

* [test] fix bugs

* [test] fix transformer version

[pipeline] support shardformer for GPT2ForQuestionAnswering & complete pipeline support for GPT2 (#4245)

* change for transformers loggers

* add forward for GPT2ForQuestionAnswering

* fix assert

* fix torchrec test

* hot fix

* hot fx tracer

* bloom policy

* llama pipeline forward and tests

* fix the output and attention_mask

* fix name

* bind argument to policy

* finish bloom model

* test shard gpt2

* clear cache

* support all bloom models

* add bloom models policies

* finish bloom pipeline and tests

* add set pipeline

* finish bloom

* bloom policy

* llama pipeline forward and tests

* fix the output and attention_mask

* fix name

* bind argument to policy

* Revert "bloom policy"

This reverts commit 8dee68a0a22568dbeed6d4563372b25e1e825fb0.

This policy should be revert and copied to feature/bloom

* revert the bloom changes

* cancel unneeded inputs

* gpt

* finish llama

* causal lm and sequence classification

* revision

* add pure pipeline test

* finish some bert models

* finish all bert models

* finish bert tests

* fix bugs

* fix bugs

* fix test pipeline

* fix data gen for qa

* update the set pipeline forward

* shared params

* fix bugs

* add forward for GPTLMHeadModel

* add test for gpt_lm

* arranging get_held_layers method

* arrange forward replacement

* add forward for GPT2ForTokenClassification

* add forward for GPT2ForSequenceClassification

* fix test_shard_gpt2.py

* add GPT2DoubleHeadsmodel & fix bugs

* add id checking in get_shared_params

* * fix typehint & docstring in sharder.py

* * update pipeline forward for GPT2Model

* * add test for pipeline forward of GPT2Model

* * add cache cleaning in gpt2 test

* * change assert to raise command

* bloom policy

* llama pipeline forward and tests

* fix the output and attention_mask

* fix name

* bind argument to policy

* finish bloom model

* test shard gpt2

* clear cache

* bloom policy

* llama pipeline forward and tests

* fix the output and attention_mask

* fix name

* bind argument to policy

* Revert "bloom policy"

This reverts commit 8dee68a0a22568dbeed6d4563372b25e1e825fb0.

This policy should be revert and copied to feature/bloom

* revert the bloom changes

* cancel unneeded inputs

* gpt

* finish llama

* causal lm and sequence classification

* revision

* bloom policy

* llama pipeline forward and tests

* fix the output and attention_mask

* fix name

* bind argument to policy

* Revert "bloom policy"

This reverts commit 8dee68a0a22568dbeed6d4563372b25e1e825fb0.

This policy should be revert and copied to feature/bloom

* revert the bloom changes

* cancel unneeded inputs

* gpt

* add pipeline forward

* complete pipeline forward check

* fix bert forward without pipeline

* fix comments

* discard useless line

* add todo

* clean prints

* fix distribute layers

* [shardformer] support lazy init

* [shardformer] linear support lazy init

* [shardformer] embedding support lazy init

* [shardformer] norm support lazy init

* [shardformer] fused linear support lazy init

* [test] update shardformer test layer

* [test] shardformer with lazy init fit ddp

* [lazy] hotfix deepcopy of param

* [shardformer] fix bert policy and update test

* [shardformer] fix bloom policy and update test

* [shardformer] fix opt policy and update test

* [shardformer] fix t5 policy and update test

* [shardformer] fix gpt2 policy and update test

* [shardformer] fix llama policy and update test

* move bert related pipeline components to shardformer

* fix bugs

* revision

* fix bert model tests

* fix bert_lm_head model tests

* fix tests

* fix tests

* done checks

* skip bloom

* add pipeline policy and bert forward to be done

* add bertmodel pipeline forward and make tests

* add Bert_Policy and test for policy

* update formatting

* update formatting

* update the code

* fix bugs

* fix name confilt

* add bloom model and policy ,revise the base class of policy

* revise

* revision

* add bert_for_pretraining

* add bert_for_pretraining forward and policy

* fix typos

* cancel warning

* change the imediate output to default dict

* change the default output of get_shared_params

* add pipeline policy and bert forward to be done

* add bertmodel pipeline forward and make tests

* add Bert_Policy and test for policy

* update formatting

* update formatting

* update the code

* fix bugs

* fix name confilt

* add bloom model and policy ,revise the base class of policy

* revise

* revision

* add bert_for_pretraining

* add pipeline policy and bert forward to be done

* add bertmodel pipeline forward and make tests

* add Bert_Policy and test for policy

* update formatting

* update formatting

* update the code

* fix bugs

* fix name confilt

* [pipeline] add stage manager

* [test] add pipeline stage manager test

* [pipeline] add docstring for stage manager

* add pipeline policy and bert forward to be done

* add bertmodel pipeline forward and make tests

* add Bert_Policy and test for policy

* update formatting

* update formatting

* update the code

* fix bugs

* fix name confilt

* [api] update optimizer wrapper to fit pipeline

* [pipeline] add base schedule

* [pipeline] add 1f1b schedule

* [test] add pipeline schedule utils test

* [pipeline] fix import

* [pipeline] add p2p communication

* [test] add p2p communication test

* [test] add rerun decorator

* [test] rename to avoid conflict

* [pipeline] add stage manager

* [test] add pipeline stage manager test

* [pipeline] add docstring for stage manager

* [cluster] add process group mesh

* [test] add process group mesh test

* force sync

* improve stablility of zero

* fix wrong index

* add record stream

Updated coloattention tests of checking outputs and gradients

fixed an import error

Improved ColoAttention interface to support flash attention 2. Solved #4322 

* [test] remove legacy zero test

* [test] remove lazy distribute test

* [test] remove outdated checkpoint io

* support shard optimizer of zero

* polish code

* support sync grad manually

* add state dict for zero

* fix unit test

* polish

* allow passing process group to zero12

* union tp-zero and normal-zero

* polish code

* support no sync for zero1 plugin

* polish

* polish

* refactor low level zero

* fix zero2 and support cpu offload

* avg gradient and modify unit test

* refactor grad store, support layer drop

* refactor bucket store, support grad accumulation

* fix and update unit test of zero and ddp

* compatible with tp, ga and unit test

* fix memory leak and polish

* add zero layer drop unittest

* polish code

* fix import err in unit test

* support diffenert comm dtype, modify docstring style

* polish code

* test padding and fix

* fix unit test of low level zero

* fix pad recording in bucket store

* support some models

* polish

* sharded optimizer checkpoint for gemini plugin

* modify test to reduce testing time

* update doc

* fix bug when keep_gatherd is true under GeminiPlugin

* [lazy] support init on cuda

* [test] update lazy init test

* [test] fix transformer version

* added softmax kernel

* added qkv_kernel

* added ops

* adding tests

* upload tets

* fix tests

* debugging

* debugging tests

* debugging

* added

* fixed errors

* added softmax kernel

* clean codes

* added tests

* update tests

* update tests

* added attention

* add

* fixed pytest checking

* add cuda check

* fix cuda version

* fix typo

* [checkpointio] unsharded optimizer checkpoint for Gemini plugin

* [checkpointio] unsharded optimizer checkpoint for Gemini using all_gather

Co-authored-by: github-actions <github-actions@github.com>

* [shardformer] made tensor parallelism configurable

* polish code