* fp8 operators for compressed communication cast_to_fp8, cast_from_fp8, all_reduce_fp8 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix typo * fix scaling algorithm in FP8 casting * support fp8 communication in pipeline parallelism * add fp8_communication flag in the script * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * shardformer fp8 * fix rebase * remove all to all * fix shardformer fp8 communication training degradation * [fp8] support all-gather flat tensor (#5932) * [fp8] add fp8 comm for low level zero * [test] add zero fp8 test case * [Feature] llama shardformer fp8 support (#5938) * add llama shardformer fp8 * Llama Shardformer Parity * fix typo * fix all reduce * fix pytest failure * fix reduce op and move function to fp8.py * fix typo * [FP8] rebase main (#5963) * add SimPO * fix dataloader * remove debug code * add orpo * fix style * fix colossalai, transformers version * fix colossalai, transformers version * fix colossalai, transformers version * fix torch colossalai version * update transformers version * [shardformer] DeepseekMoE support (#5871) * [Feature] deepseek moe expert parallel implement * [misc] fix typo, remove redundant file (#5867) * [misc] fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Feature] deepseek support & unit test * [misc] remove debug code & useless print * [misc] fix typos (#5872) * [Feature] remove modeling file, use auto config. (#5884) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [Deepseek] remove redundant code (#5888) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [misc] remove redundant code * [Feature/deepseek] resolve comment. (#5889) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [misc] remove redundant code * [misc] mv module replacement into if branch * [misc] add some warning message and modify some code in unit test * [misc] fix typos --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Hoxfix] Fix CUDA_DEVICE_MAX_CONNECTIONS for comm overlap Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Feat] Diffusion Model(PixArtAlpha/StableDiffusion3) Support (#5838) * Diffusion Model Inference support * Stable Diffusion 3 Support * pixartalpha support * [HotFix] CI,import,requirements-test for #5838 (#5892) * [Hot Fix] CI,import,requirements-test --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Feature] Enable PP + SP for llama (#5868) * fix cross-PP-stage position id length diff bug * fix typo * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * use a one cross entropy func for all shardformer models --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [ShardFormer] Add Ulysses Sequence Parallelism support for Command-R, Qwen2 and ChatGLM (#5897) * add benchmark for sft, dpo, simpo, orpo. Add benchmarking result. Support lora with gradient checkpoint * fix style * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix eval * hotfix citation * [zero] support all-gather overlap (#5898) * [zero] support all-gather overlap * [zero] add overlap all-gather flag * [misc] fix typo * [zero] update api * fix orpo cross entropy loss * [Auto Parallel]: Speed up intra-op plan generation by 44% (#5446) * Remove unnecessary calls to deepcopy * Build DimSpec's difference dict only once This change considerably speeds up construction speed of DimSpec objects. The difference_dict is the same for each DimSpec object, so a single copy of it is enough. * Fix documentation of DimSpec's difference method * [ShardFormer] fix qwen2 sp (#5903) * [compatibility] support torch 2.2 (#5875) * Support Pytorch 2.2.2 * keep build_on_pr file and update .compatibility * fix object_to_tensor usage when torch>=2.3.0 (#5820) * [misc] support torch2.3 (#5893) * [misc] support torch2.3 * [devops] update compatibility ci * [devops] update compatibility ci * [devops] add debug * [devops] add debug * [devops] add debug * [devops] add debug * [devops] remove debug * [devops] remove debug * [release] update version (#5912) * [plugin] support all-gather overlap for hybrid parallel (#5919) * [plugin] fixed all-gather overlap support for hybrid parallel * add kto * fix style, add kto data sample * [Examples] Add lazy init to OPT and GPT examples (#5924) Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [ColossalChat] Hotfix for ColossalChat (#5910) * add ignore and tiny llama * fix path issue * run style * fix issue * update bash * add ignore and tiny llama * fix path issue * run style * fix issue * update bash * fix ddp issue * add Qwen 1.5 32B * refactor tokenization * [FIX BUG] UnboundLocalError: cannot access local variable 'default_conversation' where it is not associated with a value (#5931) * cannot access local variable 'default_conversation' where it is not associated with a value set default value for 'default_conversation' * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix test data * refactor evaluation * remove real data path * remove real data path * Add n_fused as an input from native_module (#5894) * [FIX BUG] convert env param to int in (#5934) * [Hotfix] Fix ZeRO typo #5936 Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Feature] Add a switch to control whether the model checkpoint needs to be saved after each epoch ends (#5941) * Add a switch to control whether the model checkpoint needs to be saved after each epoch ends * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix style * fix style * fix style * [shardformer] hotfix attn mask (#5945) * [shardformer] hotfix attn mask (#5947) * [Feat] Distrifusion Acceleration Support for Diffusion Inference (#5895) * Distrifusion Support source * comp comm overlap optimization * sd3 benchmark * pixart distrifusion bug fix * sd3 bug fix and benchmark * generation bug fix * naming fix * add docstring, fix counter and shape error * add reference * readme and requirement * [zero] hotfix update master params (#5951) * [release] update version (#5952) * [Chat] Fix lora (#5946) * fix merging * remove filepath * fix style * Update README.md (#5958) * [hotfix] Remove unused plan section (#5957) * remove readme * fix readme * update * [test] add mixtral for sequence classification * [test] add mixtral transformer test * [moe] fix plugin * [test] mixtra pp shard test * [chore] handle non member group * [zero] solve hang * [test] pass mixtral shardformer test * [moe] implement transit between non moe tp and ep * [zero] solve hang * [misc] solve booster hang by rename the variable * solve hang when parallel mode = pp + dp * [moe] implement submesh initialization * [moe] add mixtral dp grad scaling when not all experts are activated * [chore] manually revert unintended commit * [chore] trivial fix * [chore] arg pass & remove drop token * [test] add mixtral modelling test * [moe] implement tp * [moe] test deepseek * [moe] clean legacy code * [Feature] MoE Ulysses Support (#5918) * moe sp support * moe sp bug solve * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [chore] minor fix * [moe] init moe plugin comm setting with sp * moe sp + ep bug fix * [moe] finalize test (no pp) * [moe] full test for deepseek and mixtral (pp + sp to fix) * [chore] minor fix after rebase * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * [chore] solve moe ckpt test failure and some other arg pass failure * [moe] remove ops * [test] fix test: test_zero1_2 * [bug] fix: somehow logger hangs the program * [moe] deepseek moe sp support * [test] add check * [deepseek] replace attn (a workaround for bug in transformers) * [misc] skip redunant test * [misc] remove debug/print code * [moe] refactor mesh assignment * Revert "[moe] implement submesh initialization" This reverts commit2f9bce6686
. * [chore] change moe_pg_mesh to private * [misc] remove incompatible test config * [misc] fix ci failure: change default value to false in moe plugin * [misc] remove useless condition * [chore] docstring * [moe] remove force_overlap_comm flag and add warning instead * [doc] add MoeHybridParallelPlugin docstring * [moe] solve dp axis issue * [chore] remove redundant test case, print string & reduce test tokens * [feat] Dist Loader for Eval (#5950) * support auto distributed data loader * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * support auto distributed data loader * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix tp error * remove unused parameters * remove unused * update inference * update docs * update inference --------- Co-authored-by: Michelle <qianranma8@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [lora] lora support hybrid parallel plugin (#5956) * lora support hybrid plugin * fix * fix * fix * fix * fp8 operators for compressed communication cast_to_fp8, cast_from_fp8, all_reduce_fp8 * fix scaling algorithm in FP8 casting * support fp8 communication in pipeline parallelism * add fp8_communication flag in the script * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * shardformer fp8 * fix rebase * remove all to all * fix shardformer fp8 communication training degradation * [fp8] support all-gather flat tensor (#5932) * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * Update low_level_optim.py --------- Co-authored-by: YeAnbang <anbangy2@outlook.com> Co-authored-by: Haze188 <haze188@qq.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Edenzzzz <wenxuan.tan@wisc.edu> Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: Runyu Lu <77330637+LRY89757@users.noreply.github.com> Co-authored-by: Guangyao Zhang <xjtu521@qq.com> Co-authored-by: YeAnbang <44796419+YeAnbang@users.noreply.github.com> Co-authored-by: Hongxin Liu <lhx0217@gmail.com> Co-authored-by: Stephan Kö <stephankoe@users.noreply.github.com> Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com> Co-authored-by: Tong Li <tong.li352711588@gmail.com> Co-authored-by: zhurunhua <1281592874@qq.com> Co-authored-by: Insu Jang <insujang@umich.edu> Co-authored-by: Gao, Ruiyuan <905370712@qq.com> Co-authored-by: hxwang <wang1570@e.ntu.edu.sg> Co-authored-by: Michelle <qianranma8@gmail.com> Co-authored-by: Wang Binluo <32676639+wangbluo@users.noreply.github.com> Co-authored-by: HangXu <hangxu0304@gmail.com> * [fp8]support all2all fp8 (#5953) * support all2all fp8 * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [fp8] add fp8 linear (#5967) * [fp8] add fp8 linear * [test] fix fp8 linear test condition * [test] fix fp8 linear test condition * [test] fix fp8 linear test condition * [fp8] support fp8 amp for hybrid parallel plugin (#5975) * [fp8] support fp8 amp for hybrid parallel plugin * [test] add fp8 hook test * [fp8] fix fp8 linear compatibility * fix (#5976) * [Feature]: support FP8 communication in DDP, FSDP, Gemini (#5928) * support fp8_communication in the Torch DDP grad comm, FSDP grad comm, and FSDP params comm * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * implement communication hook for FSDP params all-gather * added unit test for fp8 operators * support fp8 communication in GeminiPlugin * update training scripts to support fsdp and fp8 communication * fixed some minor bugs observed in unit test * add all_gather_into_tensor_flat_fp8 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add skip the test if torch < 2.2.0 * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add skip the test if torch < 2.2.0 * add skip the test if torch < 2.2.0 * add fp8_comm flag * rebase latest fp8 operators * rebase latest fp8 operators * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [test ci]Feature/fp8 comm (#5981) * fix * fix * fix * [fp8] support gemini plugin (#5978) * [fp8] refactor hook * [fp8] support gemini plugin * [example] add fp8 option for llama benchmark * [fp8] use torch compile (torch >= 2.3.0) (#5979) * [fp8] use torch compile (torch >= 2.4.0) * [fp8] set use_fast_accum in linear * [chore] formal version check * [chore] fix sig * [fp8]Moe support fp8 communication (#5977) * fix * support moe fp8 * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix fix fi * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [fp8] support hybrid parallel plugin (#5982) * support fp8 comm for qwen2 model * support fp8 comm for qwen2 model * support fp8 comm for qwen2 model * fp8 * fix * bert and bloom * chatglm and command * gpt2,gptj,bert, falcon,blip2 * mistral,opy,sam,t5,vit,whisper * fix * fix * fix * [fp8] refactor fp8 linear with compile (#5993) * [fp8] refactor fp8 linear with compile * [fp8] fix linear test * [fp8] fix linear test * [fp8] support asynchronous FP8 communication (#5997) * fix * fix * fix * support async all2all * support async op for all gather * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [fp8] update torch.compile for linear_fp8 to >= 2.4.0 (#6004) * [fp8] linear perf enhancement * [fp8]update reduce-scatter test (#6002) * fix * fix * fix * fix * [fp8] add use_fp8 option for MoeHybridParallelPlugin (#6009) * [fp8] zero support fp8 linear. (#6006) * fix * fix * fix * zero fp8 * zero fp8 * Update requirements.txt * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix the merge * fix the merge * fix the merge * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix the merge * fix * fix * fix the merge * fix * fix * fix * fix * fix * fix the merge * fix * fix * fix * fix * [fp8] Merge feature/fp8_comm to main branch of Colossalai (#6016) * add SimPO * fix dataloader * remove debug code * add orpo * fix style * fix colossalai, transformers version * fix colossalai, transformers version * fix colossalai, transformers version * fix torch colossalai version * update transformers version * [shardformer] DeepseekMoE support (#5871) * [Feature] deepseek moe expert parallel implement * [misc] fix typo, remove redundant file (#5867) * [misc] fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Feature] deepseek support & unit test * [misc] remove debug code & useless print * [misc] fix typos (#5872) * [Feature] remove modeling file, use auto config. (#5884) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [Deepseek] remove redundant code (#5888) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [misc] remove redundant code * [Feature/deepseek] resolve comment. (#5889) * [misc] fix typos * [Feature] deepseek support via auto model, remove modeling file * [misc] delete useless file * [misc] fix typos * [misc] remove redundant code * [misc] mv module replacement into if branch * [misc] add some warning message and modify some code in unit test * [misc] fix typos --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Hoxfix] Fix CUDA_DEVICE_MAX_CONNECTIONS for comm overlap Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Feat] Diffusion Model(PixArtAlpha/StableDiffusion3) Support (#5838) * Diffusion Model Inference support * Stable Diffusion 3 Support * pixartalpha support * [HotFix] CI,import,requirements-test for #5838 (#5892) * [Hot Fix] CI,import,requirements-test --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Feature] Enable PP + SP for llama (#5868) * fix cross-PP-stage position id length diff bug * fix typo * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * use a one cross entropy func for all shardformer models --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [ShardFormer] Add Ulysses Sequence Parallelism support for Command-R, Qwen2 and ChatGLM (#5897) * add benchmark for sft, dpo, simpo, orpo. Add benchmarking result. Support lora with gradient checkpoint * fix style * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix eval * hotfix citation * [zero] support all-gather overlap (#5898) * [zero] support all-gather overlap * [zero] add overlap all-gather flag * [misc] fix typo * [zero] update api * fix orpo cross entropy loss * [Auto Parallel]: Speed up intra-op plan generation by 44% (#5446) * Remove unnecessary calls to deepcopy * Build DimSpec's difference dict only once This change considerably speeds up construction speed of DimSpec objects. The difference_dict is the same for each DimSpec object, so a single copy of it is enough. * Fix documentation of DimSpec's difference method * [ShardFormer] fix qwen2 sp (#5903) * [compatibility] support torch 2.2 (#5875) * Support Pytorch 2.2.2 * keep build_on_pr file and update .compatibility * fix object_to_tensor usage when torch>=2.3.0 (#5820) * [misc] support torch2.3 (#5893) * [misc] support torch2.3 * [devops] update compatibility ci * [devops] update compatibility ci * [devops] add debug * [devops] add debug * [devops] add debug * [devops] add debug * [devops] remove debug * [devops] remove debug * [release] update version (#5912) * [plugin] support all-gather overlap for hybrid parallel (#5919) * [plugin] fixed all-gather overlap support for hybrid parallel * add kto * fix style, add kto data sample * [Examples] Add lazy init to OPT and GPT examples (#5924) Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [ColossalChat] Hotfix for ColossalChat (#5910) * add ignore and tiny llama * fix path issue * run style * fix issue * update bash * add ignore and tiny llama * fix path issue * run style * fix issue * update bash * fix ddp issue * add Qwen 1.5 32B * refactor tokenization * [FIX BUG] UnboundLocalError: cannot access local variable 'default_conversation' where it is not associated with a value (#5931) * cannot access local variable 'default_conversation' where it is not associated with a value set default value for 'default_conversation' * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix test data * refactor evaluation * remove real data path * remove real data path * Add n_fused as an input from native_module (#5894) * [FIX BUG] convert env param to int in (#5934) * [Hotfix] Fix ZeRO typo #5936 Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Feature] Add a switch to control whether the model checkpoint needs to be saved after each epoch ends (#5941) * Add a switch to control whether the model checkpoint needs to be saved after each epoch ends * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix style * fix style * fix style * [shardformer] hotfix attn mask (#5945) * [shardformer] hotfix attn mask (#5947) * [Feat] Distrifusion Acceleration Support for Diffusion Inference (#5895) * Distrifusion Support source * comp comm overlap optimization * sd3 benchmark * pixart distrifusion bug fix * sd3 bug fix and benchmark * generation bug fix * naming fix * add docstring, fix counter and shape error * add reference * readme and requirement * [zero] hotfix update master params (#5951) * [release] update version (#5952) * [Chat] Fix lora (#5946) * fix merging * remove filepath * fix style * Update README.md (#5958) * [hotfix] Remove unused plan section (#5957) * remove readme * fix readme * update * [test] add mixtral for sequence classification * [test] add mixtral transformer test * [moe] fix plugin * [test] mixtra pp shard test * [chore] handle non member group * [zero] solve hang * [test] pass mixtral shardformer test * [moe] implement transit between non moe tp and ep * [zero] solve hang * [misc] solve booster hang by rename the variable * solve hang when parallel mode = pp + dp * [moe] implement submesh initialization * [moe] add mixtral dp grad scaling when not all experts are activated * [chore] manually revert unintended commit * [chore] trivial fix * [chore] arg pass & remove drop token * [test] add mixtral modelling test * [moe] implement tp * [moe] test deepseek * [moe] clean legacy code * [Feature] MoE Ulysses Support (#5918) * moe sp support * moe sp bug solve * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [chore] minor fix * [moe] init moe plugin comm setting with sp * moe sp + ep bug fix * [moe] finalize test (no pp) * [moe] full test for deepseek and mixtral (pp + sp to fix) * [chore] minor fix after rebase * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * [chore] solve moe ckpt test failure and some other arg pass failure * [moe] remove ops * [test] fix test: test_zero1_2 * [bug] fix: somehow logger hangs the program * [moe] deepseek moe sp support * [test] add check * [deepseek] replace attn (a workaround for bug in transformers) * [misc] skip redunant test * [misc] remove debug/print code * [moe] refactor mesh assignment * Revert "[moe] implement submesh initialization" This reverts commit2f9bce6686
. * [chore] change moe_pg_mesh to private * [misc] remove incompatible test config * [misc] fix ci failure: change default value to false in moe plugin * [misc] remove useless condition * [chore] docstring * [moe] remove force_overlap_comm flag and add warning instead * [doc] add MoeHybridParallelPlugin docstring * [moe] solve dp axis issue * [chore] remove redundant test case, print string & reduce test tokens * [feat] Dist Loader for Eval (#5950) * support auto distributed data loader * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * support auto distributed data loader * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix tp error * remove unused parameters * remove unused * update inference * update docs * update inference --------- Co-authored-by: Michelle <qianranma8@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [lora] lora support hybrid parallel plugin (#5956) * lora support hybrid plugin * fix * fix * fix * fix * Support overall loss, update KTO logging * [Docs] clarify launch port Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Hotfix] README link (#5966) * update ignore * update readme * run style * update readme * [Hotfix] Avoid fused RMSnorm import error without apex (#5985) Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [Chat] fix readme (#5989) * fix readme * fix readme, tokenization fully tested * fix readme, tokenization fully tested * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: root <root@notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9-0.notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9.colossal-ai.svc.cluster.local> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix sync condition (#6000) * [plugin] add cast inputs option for zero (#6003) * [pre-commit.ci] pre-commit autoupdate (#5995) updates: - [github.com/psf/black-pre-commit-mirror: 24.4.2 → 24.8.0](https://github.com/psf/black-pre-commit-mirror/compare/24.4.2...24.8.0) Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [misc] Bypass the huggingface bug to solve the mask mismatch problem (#5991) * [Feature] Zigzag Ring attention (#5905) * halfway * fix cross-PP-stage position id length diff bug * fix typo * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * unified cross entropy func for all shardformer models * remove redundant lines * add basic ring attn; debug cross entropy * fwd bwd logic complete * fwd bwd logic complete; add experimental triton rescale * precision tests passed * precision tests passed * fix typos and remove misc files * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add sp_mode to benchmark; fix varlen interface * update softmax_lse shape by new interface * change tester name * remove buffer clone; support packed seq layout * add varlen tests * fix typo * all tests passed * add dkv_group; fix mask * remove debug statements --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [misc] update compatibility (#6008) * [misc] update compatibility * [misc] update requirements * [devops] disable requirements cache * [test] fix torch ddp test * [test] fix rerun on address in use * [test] fix lazy init * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix the merge * fix the merge * overlap kv comm with output rescale (#6017) Co-authored-by: Edenzzzz <wtan45@wisc.edu> * fix the merge * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix the merge * fix * fix * fix the merge * fix * [misc] Use dist logger in plugins (#6011) * use dist logger in plugins * remove trash * print on rank 0 --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> * fix * fix * fix * fix * fix the merge * fix * fix * fix * fix --------- Co-authored-by: YeAnbang <anbangy2@outlook.com> Co-authored-by: Haze188 <haze188@qq.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Edenzzzz <wenxuan.tan@wisc.edu> Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: Runyu Lu <77330637+LRY89757@users.noreply.github.com> Co-authored-by: Guangyao Zhang <xjtu521@qq.com> Co-authored-by: YeAnbang <44796419+YeAnbang@users.noreply.github.com> Co-authored-by: Hongxin Liu <lhx0217@gmail.com> Co-authored-by: Stephan Kö <stephankoe@users.noreply.github.com> Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com> Co-authored-by: Tong Li <tong.li352711588@gmail.com> Co-authored-by: zhurunhua <1281592874@qq.com> Co-authored-by: Insu Jang <insujang@umich.edu> Co-authored-by: Gao, Ruiyuan <905370712@qq.com> Co-authored-by: hxwang <wang1570@e.ntu.edu.sg> Co-authored-by: Michelle <qianranma8@gmail.com> Co-authored-by: root <root@notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9-0.notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9.colossal-ai.svc.cluster.local> * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update train_dpo.py * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update low_level_zero_plugin.py * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * [CI] Remove triton version for compatibility bug; update req torch >=2.2 (#6018) * remove triton version * remove torch 2.2 * remove torch 2.1 * debug * remove 2.1 build tests * require torch >=2.2 --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> * [plugin] hotfix zero plugin (#6036) * [plugin] hotfix zero plugin * [plugin] hotfix zero plugin * [Colossal-LLaMA] Refactor latest APIs (#6030) * refactor latest code * update api * add dummy dataset * update Readme * add setup * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update files * add PP support * update arguments * update argument * reorg folder * update version * remove IB infor * update utils * update readme * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update save for zero * update save * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add apex * update --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * add fused norm (#6038) * [FP8] unsqueeze scale to make it compatible with torch.compile (#6040) * [colossalai/checkpoint_io/...] fix bug in load_state_dict_into_model; format error msg (#6020) * fix bug in load_state_dict_into_model; format error msg * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update utils.py to support checking missing_keys * Update general_checkpoint_io.py fix bug in missing_keys error message * retrigger tests --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Hotfix] Remove deprecated install (#6042) * remove deprecated install * remove unused folder * [fp8] optimize all-gather (#6043) * [fp8] optimize all-gather * [fp8] fix all gather fp8 ring * [fp8] enable compile * [fp8] fix all gather fp8 ring * [fp8] fix linear hook (#6046) * [fp8] disable all_to_all_fp8 in intranode (#6045) * enhance all_to_all_fp8 with internode comm control * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * disable some fp8 ops due to performance issue * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [release] update version (#6041) * [release] update version * [devops] update comp test * [devops] update comp test debug * [devops] debug comp test * [devops] debug comp test * [devops] debug comp test * [devops] debug comp test * [devops] debug comp test * [Feature] Split cross-entropy computation in SP (#5959) * halfway * fix cross-PP-stage position id length diff bug * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * unified cross entropy func for all shardformer models * remove redundant lines * add basic ring attn; debug cross entropy * fwd bwd logic complete * fwd bwd logic complete; add experimental triton rescale * precision tests passed * precision tests passed * fix typos and remove misc files * update softmax_lse shape by new interface * change tester name * remove buffer clone; support packed seq layout * add varlen tests * fix typo * all tests passed * add dkv_group; fix mask * remove debug statements * adapt chatglm, command-R, qwen * debug * halfway * fix cross-PP-stage position id length diff bug * fix typo * fix typo * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * unified cross entropy func for all shardformer models * remove redundant lines * add basic ring attn; debug cross entropy * fwd bwd logic complete * fwd bwd logic complete; add experimental triton rescale * precision tests passed * precision tests passed * fix typos and remove misc files * add sp_mode to benchmark; fix varlen interface * update softmax_lse shape by new interface * add varlen tests * fix typo * all tests passed * add dkv_group; fix mask * remove debug statements * add comments * q1 index only once * remove events to simplify stream sync * simplify forward/backward logic * 2d ring forward passed * 2d ring backward passed * fixes * fix ring attn loss * 2D ring backward + llama passed * merge * update logger * fix typo * rebase * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix typo * remove typos * fixes * support GPT --------- Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [hotfix] moe hybrid parallelism benchmark & follow-up fix (#6048) * [example] pass use_fp8_comm flag to all plugins * [example] add mixtral benchmark * [moe] refine assertion and check * [moe] fix mixtral & add more tests * [moe] consider checking dp * sp group and moe_dp_group * [mixtral] remove gate tp & add more tests * [deepseek] fix tp & sp for deepseek * [mixtral] minor fix * [deepseek] add deepseek benchmark * [fp8] hotfix backward hook (#6053) * [fp8] hotfix backward hook * [fp8] hotfix pipeline loss accumulation * [doc] update sp doc (#6055) * update sp doc * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * fix the sp * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix the attn * fix * fix * fix * fix * [zerobubble]Support ZeroBubble Pipeline (#6034) * [feat] add zerobubble pp (just a frame now); add POC test for dx_dw; add test for zerobubble; * [feat] add dw test; * [fix] fix weight not close; * [update] update text; * [feat] add test run_fwd_bwd automatic scheduling; * [feat] split communication and calculation; fix pop empty send_bwd_buffer error; * [feat] add test for p & p grad; * [feat] add comments for ZBV func; * [fix] rm useless assign and comments; * [fix] fix ci test; add pytest; * [feat] add run_fwd_bwd_with_microbatch (replace input) & test; add p&p.grad assert close test & all pass; * [feat] add apply v_schedule graph; p & p.grad assert err exist; * [fix] update * [feat] fix ci; add assert; * [feat] fix poc format * [feat] fix func name & ci; add comments; * [fix] fix poc test; add comments in poc; * [feat] add optim backward_b_by_grad * [feat] fix optimizer bwd b & w; support return accum loss & output * [feat] add fwd_bwd_step, run_fwd_only; * [fix] fix optim bwd; add license for v_schedule; remove redundant attributes; fix schedule loop "while"--> "for"; add communication dict; * [fix] fix communication_map; * [feat] update test; rm comments; * [fix] rm zbv in hybridplugin * [fix] fix optim bwd; * [fix] fix optim bwd; * [fix] rm output.data after send fwd; * [fix] fix bwd step if condition; remove useless comments and format info; * [fix] fix detach output & release output; * [fix] rm requir_grad for output; * [fix] fix requir grad position and detach position and input&output local buffer append position; * [feat] add memory assertation; * [fix] fix mem check; * [fix] mem assertation' * [fix] fix mem assertation * [fix] fix mem; use a new model shape; only assert mem less and equal than theo; * [fix] fix model zoo import; * [fix] fix redundant detach & clone; add buffer assertation in the end; * [fix] add output_obj_grad assert None at bwd b step; replace input_obj.require_grad_ with treemap; * [fix] update optim state dict assert (include param group & state); fix mem assert after add optim; * [fix] add testcase with microbatch 4; * [fp8] fix missing fp8_comm flag in mixtral (#6057) * fix * fix * fix * [fp8] Disable all_gather intranode. Disable Redundant all_gather fp8 (#6059) * all_gather only internode, fix pytest * fix cuda arch <89 compile pytest error * fix pytest failure * disable all_gather_into_tensor_flat_fp8 * fix fp8 format * fix pytest * fix conversations * fix chunk tuple to list * [doc] FP8 training and communication document (#6050) * Add FP8 training and communication document * add fp8 docstring for plugins * fix typo * fix typo * fix * fix * [moe] add parallel strategy for shared_expert && fix test for deepseek (#6063) * [ColossalEval] support for vllm (#6056) * support vllm * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * modify vllm and update readme * run pre-commit * remove dupilicated lines and refine code * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * update param name * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * refine code * update readme * refine code * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [release] update version (#6062) * [feat] add zerobubble pp (just a frame now); add POC test for dx_dw; add test for zerobubble; * [update] update text; * [feat] add test run_fwd_bwd automatic scheduling; * [feat] fix poc format * [fix] fix poc test; add comments in poc; * [feat] add optim backward_b_by_grad * [feat] fix optimizer bwd b & w; support return accum loss & output * [fix] fix optim bwd; add license for v_schedule; remove redundant attributes; fix schedule loop "while"--> "for"; add communication dict; * [feat] update test; rm comments; * [fix] rm zbv in hybridplugin * [fix] fix optim bwd; * [fix] fix optim bwd; * [fix] rm output.data after send fwd; * [fix] fix bwd step if condition; remove useless comments and format info; * [fix] fix mem check; * [fix] fix mem assertation * [fix] fix mem; use a new model shape; only assert mem less and equal than theo; * [fix] fix model zoo import; * [feat] moehybrid support zerobubble; * [fix] fix zerobubble pp for shardformer type input; * [fix] fix require_grad & deallocate call; * [fix] fix mem assert; * [fix] fix fwd branch, fwd pass both micro_batch & internal_inputs' * [fix] fix pipeline util func deallocate --> release_tensor_data; fix bwd_b loss bwd branch; * [fix] fix zerobubble; support shardformer model type; * [fix] fix test_pipeline_utils ci; * [plugin] hybrid support zero bubble pipeline (#6060) * hybrid support zbv * fix fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * Update zero_bubble_pp.py * fix * fix-ci * fix [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix * [zerobubble]Support ZeroBubble Pipeline (#6034) * [feat] add zerobubble pp (just a frame now); add POC test for dx_dw; add test for zerobubble; * [feat] add dw test; * [fix] fix weight not close; * [update] update text; * [feat] add test run_fwd_bwd automatic scheduling; * [feat] split communication and calculation; fix pop empty send_bwd_buffer error; * [feat] add test for p & p grad; * [feat] add comments for ZBV func; * [fix] rm useless assign and comments; * [fix] fix ci test; add pytest; * [feat] add run_fwd_bwd_with_microbatch (replace input) & test; add p&p.grad assert close test & all pass; * [feat] add apply v_schedule graph; p & p.grad assert err exist; * [fix] update * [feat] fix ci; add assert; * [feat] fix poc format * [feat] fix func name & ci; add comments; * [fix] fix poc test; add comments in poc; * [feat] add optim backward_b_by_grad * [feat] fix optimizer bwd b & w; support return accum loss & output * [feat] add fwd_bwd_step, run_fwd_only; * [fix] fix optim bwd; add license for v_schedule; remove redundant attributes; fix schedule loop "while"--> "for"; add communication dict; * [fix] fix communication_map; * [feat] update test; rm comments; * [fix] rm zbv in hybridplugin * [fix] fix optim bwd; * [fix] fix optim bwd; * [fix] rm output.data after send fwd; * [fix] fix bwd step if condition; remove useless comments and format info; * [fix] fix detach output & release output; * [fix] rm requir_grad for output; * [fix] fix requir grad position and detach position and input&output local buffer append position; * [feat] add memory assertation; * [fix] fix mem check; * [fix] mem assertation' * [fix] fix mem assertation * [fix] fix mem; use a new model shape; only assert mem less and equal than theo; * [fix] fix model zoo import; * [fix] fix redundant detach & clone; add buffer assertation in the end; * [fix] add output_obj_grad assert None at bwd b step; replace input_obj.require_grad_ with treemap; * [fix] update optim state dict assert (include param group & state); fix mem assert after add optim; * [fix] add testcase with microbatch 4; * hybrid support zbv * fix fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update zero_bubble_pp.py * fix * fix-ci * fix [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix * fix * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: duanjunwen <935724073@qq.com> * [feat] add zerobubble pp (just a frame now); add POC test for dx_dw; add test for zerobubble; * [update] update text; * [feat] add test run_fwd_bwd automatic scheduling; * [feat] fix poc format * [fix] fix poc test; add comments in poc; * [feat] add optim backward_b_by_grad * [feat] fix optimizer bwd b & w; support return accum loss & output * [fix] fix optim bwd; add license for v_schedule; remove redundant attributes; fix schedule loop "while"--> "for"; add communication dict; * [feat] update test; rm comments; * [fix] fix optim bwd; * [fix] fix optim bwd; * [fix] rm output.data after send fwd; * [fix] fix bwd step if condition; remove useless comments and format info; * [fix] fix mem check; * [fix] fix mem assertation * [fix] fix mem; use a new model shape; only assert mem less and equal than theo; * [fix] fix model zoo import; * [fix] fix mem assert; * [fix] fix fwd branch, fwd pass both micro_batch & internal_inputs' * [plugin] hybrid support zero bubble pipeline (#6060) * hybrid support zbv * fix fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * Update zero_bubble_pp.py * fix * fix-ci * fix [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix * [zerobubble]Support ZeroBubble Pipeline (#6034) * [feat] add zerobubble pp (just a frame now); add POC test for dx_dw; add test for zerobubble; * [feat] add dw test; * [fix] fix weight not close; * [update] update text; * [feat] add test run_fwd_bwd automatic scheduling; * [feat] split communication and calculation; fix pop empty send_bwd_buffer error; * [feat] add test for p & p grad; * [feat] add comments for ZBV func; * [fix] rm useless assign and comments; * [fix] fix ci test; add pytest; * [feat] add run_fwd_bwd_with_microbatch (replace input) & test; add p&p.grad assert close test & all pass; * [feat] add apply v_schedule graph; p & p.grad assert err exist; * [fix] update * [feat] fix ci; add assert; * [feat] fix poc format * [feat] fix func name & ci; add comments; * [fix] fix poc test; add comments in poc; * [feat] add optim backward_b_by_grad * [feat] fix optimizer bwd b & w; support return accum loss & output * [feat] add fwd_bwd_step, run_fwd_only; * [fix] fix optim bwd; add license for v_schedule; remove redundant attributes; fix schedule loop "while"--> "for"; add communication dict; * [fix] fix communication_map; * [feat] update test; rm comments; * [fix] rm zbv in hybridplugin * [fix] fix optim bwd; * [fix] fix optim bwd; * [fix] rm output.data after send fwd; * [fix] fix bwd step if condition; remove useless comments and format info; * [fix] fix detach output & release output; * [fix] rm requir_grad for output; * [fix] fix requir grad position and detach position and input&output local buffer append position; * [feat] add memory assertation; * [fix] fix mem check; * [fix] mem assertation' * [fix] fix mem assertation * [fix] fix mem; use a new model shape; only assert mem less and equal than theo; * [fix] fix model zoo import; * [fix] fix redundant detach & clone; add buffer assertation in the end; * [fix] add output_obj_grad assert None at bwd b step; replace input_obj.require_grad_ with treemap; * [fix] update optim state dict assert (include param group & state); fix mem assert after add optim; * [fix] add testcase with microbatch 4; * hybrid support zbv * fix fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update zero_bubble_pp.py * fix * fix-ci * fix [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix * fix * fix * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * fix * fix * fix --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: duanjunwen <935724073@qq.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: HangXu <hangxu0304@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: GuangyaoZhang <xjtu521@qq.com> Co-authored-by: Hongxin Liu <lhx0217@gmail.com> Co-authored-by: YeAnbang <anbangy2@outlook.com> Co-authored-by: Haze188 <haze188@qq.com> Co-authored-by: Edenzzzz <wenxuan.tan@wisc.edu> Co-authored-by: Edenzzzz <wtan45@wisc.edu> Co-authored-by: Runyu Lu <77330637+LRY89757@users.noreply.github.com> Co-authored-by: YeAnbang <44796419+YeAnbang@users.noreply.github.com> Co-authored-by: Stephan Kö <stephankoe@users.noreply.github.com> Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com> Co-authored-by: Tong Li <tong.li352711588@gmail.com> Co-authored-by: zhurunhua <1281592874@qq.com> Co-authored-by: Insu Jang <insujang@umich.edu> Co-authored-by: Gao, Ruiyuan <905370712@qq.com> Co-authored-by: hxwang <wang1570@e.ntu.edu.sg> Co-authored-by: Michelle <qianranma8@gmail.com> Co-authored-by: Wang Binluo <32676639+wangbluo@users.noreply.github.com> Co-authored-by: wangbluo <2538539015@qq.com> Co-authored-by: root <root@notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9-0.notebook-8f919155-6035-47b4-9c6f-1be133b9e2c9.colossal-ai.svc.cluster.local> Co-authored-by: duanjunwen <935724073@qq.com> Co-authored-by: Camille Zhong <44392324+Camille7777@users.noreply.github.com>
8.9 KiB
Executable File
并行技术
作者: Shenggui Li, Siqi Mai
简介
随着深度学习的发展,对并行训练的需求越来越大。这是因为模型和数据集越来越大,如果我们坚持使用单 GPU 训练,训练过程的等待将会成为一场噩梦。在本节中,我们将对现有的并行训练方法进行简要介绍。如果您想对这篇文章进行补充,欢迎在GitHub论坛上进行讨论。
数据并行
数据并行是最常见的并行形式,因为它很简单。在数据并行训练中,数据集被分割成几个碎片,每个碎片被分配到一个设备上。这相当于沿批次维度对训练过程进行并行化。每个设备将持有一个完整的模型副本,并在分配的数据集碎片上进行训练。在反向传播之后,模型的梯度将被全部减少,以便在不同设备上的模型参数能够保持同步。

模型并行
在数据并行训练中,一个明显的特点是每个 GPU 持有整个模型权重的副本。这就带来了冗余问题。另一种并行模式是模型并行,即模型被分割并分布在一个设备阵列上。通常有两种类型的并行:张量并行和流水线并行。张量并行是在一个操作中进行并行计算,如矩阵-矩阵乘法。流水线并行是在各层之间进行并行计算。因此,从另一个角度来看,张量并行可以被看作是层内并行,流水线并行可以被看作是层间并行。
张量并行
张量并行训练是将一个张量沿特定维度分成 N
块,每个设备只持有整个张量的 1/N
,同时不影响计算图的正确性。这需要额外的通信来确保结果的正确性。
以一般的矩阵乘法为例,假设我们有 C = AB
。我们可以将B沿着列分割成 [B0 B1 B2 ... Bn]
,每个设备持有一列。然后我们将 A
与每个设备上 B
中的每一列相乘,我们将得到 [AB0 AB1 AB2 ... ABn]
。此刻,每个设备仍然持有一部分的结果,例如,设备(rank=0)持有 AB0
。为了确保结果的正确性,我们需要收集全部的结果,并沿列维串联张量。通过这种方式,我们能够将张量分布在设备上,同时确保计算流程保持正确。

在 Colossal-AI 中,我们提供了一系列的张量并行方法,即 1D、2D、2.5D 和 3D 张量并行。我们将在高级教程
中详细讨论它们。
相关文章:
- GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
- An Efficient 2D Method for Training Super-Large Deep Learning Models
- 2.5-dimensional distributed model training
- Maximizing Parallelism in Distributed Training for Huge Neural Networks
流水线并行
流水线并行一般来说很容易理解。请您回忆一下您的计算机结构课程,这确实存在于 CPU 设计中。

流水线并行的核心思想是,模型按层分割成若干块,每块都交给一个设备。在前向传递过程中,每个设备将中间的激活传递给下一个阶段。在后向传递过程中,每个设备将输入张量的梯度传回给前一个流水线阶段。这允许设备同时进行计算,并增加了训练的吞吐量。流水线并行训练的一个缺点是,会有一些设备参与计算的冒泡时间,导致计算资源的浪费。

相关文章:
- PipeDream: Fast and Efficient Pipeline Parallel DNN Training
- GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
- Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines
序列并行
序列并行是一种对于序列维度进行切分的并行策略,它是训练长文本序列的有效方法。现成熟的序列并行方法包括megatron提出的序列并行,DeepSpeed-Ulysses序列并行和ring-attention序列并行等。
megatron sp:
该序列并行方法是在张量并行的基础上实现的序列并行,模型并行的每个gpu上,样本独立且重复的,对于非线性运算的部分如layernorm等无法使用张量并行的模块,可以在序列维度将样本数据切分为多个部分,每个gpu计算部分数据,然后在计算attention及mlp等线性部分使用张量并行策略,需要将activation汇总,这样可以在模型进行切分的情况下进一步减少activation的内存占用,需要注意的是该序列并行方法只能与张量并行一起使用。
DeepSpeed-Ulysses:
序列并行通过在序列维度上分割样本并利用all-to-all通信操作,使每个GPU接收完整序列但仅计算注意力头的非重叠子集,从而实现序列并行。该并行方法具有完全通用的attention,可支持密集和稀疏的注意力。 alltoall是一个全交换操作,相当于分布式转置的操作,在attention计算之前,将样本沿序列维度进行切分,每个设备只有N/P的序列长度,然而使用alltoall后,qkv的子部分shape变为[N, d/p],在计算attention时仍考虑了整体的序列。
ring attention:
ring attention思路类似于flash attention,每个GPU只计算一个局部的attention,最后将所有的attention块结果进行归约计算出总的attention。在Ring Attention中,输入序列被沿着序列维度切分为多个块,每个块由不同的GPU或处理器负责处理,Ring Attention采用了一种称为“环形通信”的策略,通过跨卡的p2p通信相互传递kv子块来实现迭代计算,可以实现多卡的超长文本。在这种策略下,每个处理器只与它的前一个和后一个处理器交换信息,形成一个环形网络。通过这种方式,中间结果可以在处理器之间高效传递,而无需全局同步,减少了通信开销。
相关论文: Reducing Activation Recomputation in Large Transformer Models DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models Ring Attention with Blockwise Transformers for Near-Infinite Context
优化器相关的并行
另一种并行方法和优化器相关,目前这种并行最流行的方法是 ZeRO
,即零冗余优化器。 ZeRO 在三个层面上工作,以消除内存冗余(ZeRO需要进行fp16训练)。
- Level 1: 优化器状态在各进程中被划分。
- Level 2: 用于更新模型权重的32位梯度也被划分,因此每个进程只存储与其优化器状态划分相对应的梯度。
- Level 3: 16位模型参数在各进程中被划分。
相关文章:
异构系统的并行
上述方法通常需要大量的 GPU 来训练一个大型模型。然而,人们常常忽略的是,与 GPU 相比,CPU 的内存要大得多。在一个典型的服务器上,CPU 可以轻松拥有几百GB的内存,而每个 GPU 通常只有16或32GB的内存。这促使人们思考为什么 CPU 内存没有被用于分布式训练。
最近的进展是依靠 CPU 甚至是 NVMe 磁盘来训练大型模型。主要的想法是,在不使用张量时,将其卸载回 CPU 内存或 NVMe 磁盘。通过使用异构系统架构,有可能在一台机器上容纳一个巨大的模型。

相关文章: