Yuanheng Zhao
537a3cbc4d
[kernel] Support New KCache Layout - Triton Kernel ( #5677 )
...
* kvmemcpy triton for new kcache layout
* revise tests for new kcache layout
* naive triton flash decoding - new kcache layout
* rotary triton kernel - new kcache layout
* remove redundancy - triton decoding
* remove redundancy - triton kvcache copy
* [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>
2024-05-03 17:20:45 +08:00
Steve Luo
5cd75ce4c7
[Inference/Kernel] refactor kvcache manager and rotary_embedding and kvcache_memcpy oper… ( #5663 )
...
* refactor kvcache manager and rotary_embedding and kvcache_memcpy operator
* refactor decode_kv_cache_memcpy
* enable alibi in pagedattention
2024-04-30 15:52:23 +08:00
yuehuayingxueluo
5f00002e43
[Inference] Adapt Baichuan2-13B TP ( #5659 )
...
* adapt to baichuan2 13B
* add baichuan2 13B TP
* update baichuan tp logic
* rm unused code
* Fix TP logic
* fix alibi slopes tp logic
* rm nn.Module
* Polished the code.
* change BAICHUAN_MODEL_NAME_OR_PATH
* Modified the logic for loading Baichuan weights.
* fix typos
2024-04-30 15:47:07 +08:00
Yuanheng Zhao
5be590b99e
[kernel] Support new KCache Layout - Context Attention Triton Kernel ( #5658 )
...
* add context attn triton kernel - new kcache layout
* add benchmark triton
* tiny revise
* trivial - code style, comment
2024-04-26 17:51:49 +08:00
yuehuayingxueluo
3c91e3f176
[Inference]Adapt to baichuan2 13B ( #5614 )
...
* adapt to baichuan2 13B
* adapt to baichuan2 13B
* change BAICHUAN_MODEL_NAME_OR_PATH
* fix test_decoding_attn.py
* Modifications based on review comments.
* change BAICHUAN_MODEL_NAME_OR_PATH
* mv attn mask processes to test flash decoding
* mv get_alibi_slopes baichuan modeling
* fix bugs in test_baichuan.py
2024-04-25 23:11:30 +08:00
Yuanheng Zhao
f342a93871
[Fix] Remove obsolete files - inference ( #5650 )
2024-04-25 22:04:59 +08:00
Steve Luo
a8fd3b0342
[Inference/Kernel] Optimize paged attention: Refactor key cache layout ( #5643 )
...
* optimize flashdecodingattention: refactor code with different key cache layout(from [num_blocks, num_kv_heads, block_size, head_size] to [num_blocks, num_kv_heads, head_size/x, block_size, x])
* [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>
2024-04-25 14:24:02 +08:00
yuehuayingxueluo
12f10d5b0b
[Fix/Inference]Fix CUDA Rotary Rmbedding GQA ( #5623 )
...
* fix rotary embedding GQA
* change test_rotary_embdding_unpad.py KH
2024-04-23 13:44:49 +08:00
Yuanheng Zhao
5d4c1fe8f5
[Fix/Inference] Fix GQA Triton and Support Llama3 ( #5624 )
...
* [fix] GQA calling of flash decoding triton
* fix kv cache alloc shape
* fix rotary triton - GQA
* fix sequence max length assigning
* Sequence max length logic
* fix scheduling and spec-dec
* skip without import error
* fix pytest - skip without ImportError
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2024-04-23 13:09:55 +08:00
Steve Luo
ccf72797e3
feat baichuan2 rmsnorm whose hidden size equals to 5120 ( #5611 )
2024-04-19 15:34:53 +08:00
Runyu Lu
e37ee2fb65
[Feat]Tensor Model Parallel Support For Inference ( #5563 )
...
* tensor parallel support naive source
* [fix]precision, model load and refactor the framework
* add tp unit test
* docstring
* fix do_sample
2024-04-18 16:56:46 +08:00
Steve Luo
be396ad6cc
[Inference/Kernel] Add Paged Decoding kernel, sequence split within the same thread block ( #5531 )
...
* feat flash decoding for paged attention
* refactor flashdecodingattention
* [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>
2024-04-18 16:45:07 +08:00
yuehuayingxueluo
56b222eff8
[inference/model]Adapted to the baichuan2-7B model ( #5591 )
...
* Adapted to the baichuan2-7B model
* modified according to the review comments.
* Modified the method of obtaining random weights.
* modified according to the review comments.
* change mlp layewr 'NOTE'
2024-04-15 16:53:02 +08:00
Yuanheng Zhao
e60d430cf5
[Fix] resolve conflicts of rebasing feat/speculative-decoding ( #5557 )
...
- resolve conflicts of rebasing feat/speculative-decoding
2024-04-10 18:13:49 +08:00
Yuanheng Zhao
d85d91435a
[Inference/SpecDec] Support GLIDE Drafter Model ( #5455 )
...
* add glide-llama policy and modeling
* update glide modeling, compitable with transformers 4.36.2
* revise glide llama modeling/usage
* fix issues of glimpsing large kv
* revise the way re-loading params for glide drafter
* fix drafter and engine tests
* enable convert to glide strict=False
* revise glide llama modeling
* revise vicuna prompt template
* revise drafter and tests
* apply usage of glide model in engine
2024-04-10 11:07:52 +08:00
Yuanheng Zhao
a37f82629d
[Inference/SpecDec] Add Speculative Decoding Implementation ( #5423 )
...
* fix flash decoding mask during verification
* add spec-dec
* add test for spec-dec
* revise drafter init
* remove drafter sampling
* retire past kv in drafter
* (trivial) rename attrs
* (trivial) rename arg
* revise how we enable/disable spec-dec
2024-04-10 11:07:52 +08:00
Yuanheng Zhao
5a9b05f7b2
[Inference/SpecDec] Add Basic Drafter Model Container ( #5405 )
...
* [Infer/Fix] Fix Dependency in test - RMSNorm kernel (#5399 )
fix dependency in pytest
* add drafter model container (basic ver)
2024-04-10 11:07:51 +08:00
Yuanheng Zhao
d63c469f45
[Infer] Revise and Adapt Triton Kernels for Spec-Dec ( #5401 )
...
* [Infer/Fix] Fix Dependency in test - RMSNorm kernel (#5399 )
fix dependency in pytest
* resolve conflicts for revising flash-attn
* adapt kv cache copy kernel for spec-dec
* fix seqlen-n kvcache copy kernel/tests
* test kvcache copy - use torch.equal
* add assertions
* (trivial) comment out
2024-04-10 11:07:51 +08:00
Yuanheng
ed5ebd1735
[Fix] resolve conflicts of merging main
2024-04-08 16:21:47 +08:00
Hongxin Liu
641b1ee71a
[devops] remove post commit ci ( #5566 )
...
* [devops] remove post commit ci
* [misc] run pre-commit on all files
* [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>
2024-04-08 15:09:40 +08:00
Zhongkai Zhao
8e412a548e
[shardformer] Sequence Parallelism Optimization ( #5533 )
...
* sequence parallel optimization
* validate sequence parallel in llama (code to be polished)
* shardformer api writing
* integrate sequence parallel in ShardFormer
* fix pp bugs and sp bugs for LlaMa model
* integrating ring-based sequence parallelism into ShardFormer
* [sequence parallelism]: Add fused megatron function
* integrating ring-based sequence parallelism into ShardFormer
---------
Co-authored-by: linsj20 <linsj20@mails.tsinghua.edu.cn >
* fix bugs when useing sp and flashattention together
* fix operation function name
* support flash attention for ulysses-style sp
* clarify sp process group
* fix compatibility bugs in moe plugin
* fix fused linear bugs
* fix linear layer test
* support gpt model all-to-all sp
* modify shard data dimension (meant to be dim=-1)
* support megtron-style sp and distributed attn for llama model
* [shardformer] add megatron sp to llama
* support llama7B 128k with distributed attention
* [shardformer] robustness enhancement
* add block attn
* sp mode 1: keep input as a complete sequence
* fix sp compatability
* finish sp mode 3 support for gpt
* using all_to_all_single when batch size is 1
* support mode 2 sp in gpt2 (#5 )
* [shardformer] add megatron sp to llama
* support llama7B 128k with distributed attention
* [shardformer] robustness enhancement
* add block attn
* sp mode 1: keep input as a complete sequence
* fix sp compatability
* refactor ring implementation
* support mode 2 sp in gpt2
* polish code
* enable distributed attn mask when using sp mode 2 and 3 in llama
* automatically enable flash attn when using sp mode 2 and 3 in llama
* inplace attn mask
* add zero2 support for sequence parallel
* polish code
* fix bugs
* fix gemini checkpoint io
* loose tensor checking atol and rtol
* add comment
* fix llama layernorm grad
* fix zero grad
* fix zero grad
* fix conflict
* update split and gather auto grad func
* sequence parallel: inside text split (#6 )
* polish code (part 1)
* polish code (part 2)
* polish code (part 2.5)
* polish code (part 3)
* sequence parallel: inside text split
* miscellaneous minor fixes
* polish code
* fix ulysses style ZeRO
* sequence parallel: inside text split
* miscellaneous minor fixes
* disaggregate sp group and dp group for sp
* fix llama and gpt sp
* polish code
* move ulysses grad sync to ddp (#9 )
* remove zero_stage and unbind the grad sync for alltoall sp
* add 2d group creation test
* move ulysses grad sync to ddp
* add 2d group creation test
* remove useless code
* change shard config not to enable sp when enable_all_optimizations
* add sp warnings for several model
* remove useless code
---------
Co-authored-by: linsj20 <linsj20@mails.tsinghua.edu.cn >
2024-04-03 17:15:47 +08:00
yuehuayingxueluo
04aca9e55b
[Inference/Kernel]Add get_cos_and_sin Kernel ( #5528 )
...
* Add get_cos_and_sin kernel
* fix code comments
* fix code typos
* merge common codes of get_cos_and_sin kernel.
* Fixed a typo
* Changed 'asset allclose' to 'assert equal'.
2024-04-01 13:47:14 +08:00
Wenhao Chen
e614aa34f3
[shardformer, pipeline] add gradient_checkpointing_ratio and heterogenous shard policy for llama ( #5508 )
...
* feat: add `GradientCheckpointConfig` and `PipelineGradientCheckpointConfig`
* feat: apply `GradientCheckpointConfig` to policy and llama_forward
* feat: move `distribute_layer` and `get_stage_index` to PipelineStageManager
* fix: add optional args for `distribute_layer` and `get_stage_index`
* fix: fix changed API calls
* test: update llama tests
* style: polish `GradientCheckpointConfig`
* fix: fix pipeline utils tests
2024-04-01 11:34:58 +08:00
Insu Jang
00525f7772
[shardformer] fix pipeline forward error if custom layer distribution is used ( #5189 )
...
* Use self.[distribute_layers|get_stage_index] to exploit custom layer distribution
* Change static methods for t5 layer distribution to member functions
* Change static methods for whisper layer distribution to member functions
* Replace whisper policy usage with self one
* Fix test case to use non-static layer distribution methods
* fix: fix typo
---------
Co-authored-by: Wenhao Chen <cwher@outlook.com >
2024-03-27 13:57:00 +08:00
Hongxin Liu
19e1a5cf16
[shardformer] update colo attention to support custom mask ( #5510 )
...
* [feature] refactor colo attention (#5462 )
* [extension] update api
* [feature] add colo attention
* [feature] update sdpa
* [feature] update npu attention
* [feature] update flash-attn
* [test] add flash attn test
* [test] update flash attn test
* [shardformer] update modeling to fit colo attention (#5465 )
* [misc] refactor folder structure
* [shardformer] update llama flash-attn
* [shardformer] fix llama policy
* [devops] update tensornvme install
* [test] update llama test
* [shardformer] update colo attn kernel dispatch
* [shardformer] update blip2
* [shardformer] update chatglm
* [shardformer] update gpt2
* [shardformer] update gptj
* [shardformer] update opt
* [shardformer] update vit
* [shardformer] update colo attention mask prep
* [shardformer] update whisper
* [test] fix shardformer tests (#5514 )
* [test] fix shardformer tests
* [test] fix shardformer tests
2024-03-27 11:19:32 +08:00
Edenzzzz
61da3fbc52
fixed layout converter caching and updated tester
2024-03-26 17:22:27 +08:00
flybird11111
0688d92e2d
[shardformer]Fix lm parallel. ( #5480 )
...
* fix
* padding vocab_size when using pipeline parallellism
padding vocab_size when using pipeline parallellism
fix
fix
* fix
* fix
fix
fix
* fix gather output
* fix
* fix
* fix
fix resize embedding
fix resize embedding
* fix resize embedding
fix
* revert
* revert
* revert
* fix lm forward distribution
* fix
* test ci
* fix
2024-03-25 17:21:51 +08:00
Runyu Lu
68e9396bc0
[fix] merge conflicts
2024-03-25 14:48:28 +08:00
yuehuayingxueluo
87079cffe8
[Inference]Support FP16/BF16 Flash Attention 2 And Add high_precision Flag To Rotary Embedding ( #5461 )
...
* Support FP16/BF16 Flash Attention 2
* fix bugs in test_kv_cache_memcpy.py
* add context_kv_cache_memcpy_kernel.cu
* rm typename MT
* add tail process
* add high_precision
* add high_precision to config.py
* rm unused code
* change the comment for the high_precision parameter
* update test_rotary_embdding_unpad.py
* fix vector_copy_utils.h
* add comment for self.high_precision when using float32
2024-03-25 13:40:34 +08:00
Wenhao Chen
bb0a668fee
[hotfix] set return_outputs=False in examples and polish code ( #5404 )
...
* fix: simplify merge_batch
* fix: use return_outputs=False to eliminate extra memory consumption
* feat: add return_outputs warning
* style: remove `return_outputs=False` as it is the default value
2024-03-25 12:31:09 +08:00
Runyu Lu
9fe61b4475
[fix]
2024-03-25 11:37:58 +08:00
Runyu Lu
aabc9fb6aa
[feat] add use_cuda_kernel option
2024-03-19 13:24:25 +08:00
flybird11111
5e16bf7980
[shardformer] fix gathering output when using tensor parallelism ( #5431 )
...
* fix
* padding vocab_size when using pipeline parallellism
padding vocab_size when using pipeline parallellism
fix
fix
* fix
* fix
fix
fix
* fix gather output
* fix
* fix
* fix
fix resize embedding
fix resize embedding
* fix resize embedding
fix
* revert
* revert
* revert
2024-03-18 15:55:11 +08:00
Runyu Lu
d02e257abd
Merge branch 'feature/colossal-infer' into colossal-infer-cuda-graph
2024-03-14 10:37:05 +08:00
Runyu Lu
ae24b4f025
diverse tests
2024-03-14 10:35:08 +08:00
Runyu Lu
1821a6dab0
[fix] pytest and fix dyn grid bug
2024-03-13 17:28:32 +08:00
yuehuayingxueluo
f366a5ea1f
[Inference/kernel]Add Fused Rotary Embedding and KVCache Memcopy CUDA Kernel ( #5418 )
...
* add rotary embedding kernel
* add rotary_embedding_kernel
* add fused rotary_emb and kvcache memcopy
* add fused_rotary_emb_and_cache_kernel.cu
* add fused_rotary_emb_and_memcopy
* fix bugs in fused_rotary_emb_and_cache_kernel.cu
* fix ci bugs
* use vec memcopy and opt the gloabl memory access
* fix code style
* fix test_rotary_embdding_unpad.py
* codes revised based on the review comments
* fix bugs about include path
* rm inline
2024-03-13 17:20:03 +08:00
Steve Luo
ed431de4e4
fix rmsnorm template function invocation problem(template function partial specialization is not allowed in Cpp) and luckily pass e2e precision test ( #5454 )
2024-03-13 16:00:55 +08:00
Hongxin Liu
f2e8b9ef9f
[devops] fix compatibility ( #5444 )
...
* [devops] fix compatibility
* [hotfix] update compatibility test on pr
* [devops] fix compatibility
* [devops] record duration during comp test
* [test] decrease test duration
* fix falcon
2024-03-13 15:24:13 +08:00
Steve Luo
f7aecc0c6b
feat rmsnorm cuda kernel and add unittest, benchmark script ( #5417 )
2024-03-08 16:21:12 +08:00
xs_courtesy
95c21498d4
add silu_and_mul for infer
2024-03-07 16:57:49 +08:00
flybird11111
29695cf70c
[example]add gpt2 benchmark example script. ( #5295 )
...
* benchmark gpt2
* fix
fix
fix
fix
* [doc] fix typo in Colossal-LLaMA-2/README.md (#5247 )
* [workflow] fixed build CI (#5240 )
* [workflow] fixed build CI
* polish
* polish
* polish
* polish
* polish
* [ci] fixed booster test (#5251 )
* [ci] fixed booster test
* [ci] fixed booster test
* [ci] fixed booster test
* [ci] fixed ddp test (#5254 )
* [ci] fixed ddp test
* polish
* fix typo in applications/ColossalEval/README.md (#5250 )
* [ci] fix shardformer tests. (#5255 )
* fix ci
fix
* revert: revert p2p
* feat: add enable_metadata_cache option
* revert: enable t5 tests
---------
Co-authored-by: Wenhao Chen <cwher@outlook.com >
* [doc] fix doc typo (#5256 )
* [doc] fix annotation display
* [doc] fix llama2 doc
* [hotfix]: add pp sanity check and fix mbs arg (#5268 )
* fix: fix misleading mbs arg
* feat: add pp sanity check
* fix: fix 1f1b sanity check
* [workflow] fixed incomplete bash command (#5272 )
* [workflow] fixed oom tests (#5275 )
* [workflow] fixed oom tests
* polish
* polish
* polish
* [ci] fix test_hybrid_parallel_plugin_checkpoint_io.py (#5276 )
* fix ci
fix
* fix test
* revert: revert p2p
* feat: add enable_metadata_cache option
* revert: enable t5 tests
* fix
---------
Co-authored-by: Wenhao Chen <cwher@outlook.com >
* [shardformer] hybridparallelplugin support gradients accumulation. (#5246 )
* support gradients acc
fix
fix
fix
fix
fix
fix
fix
fix
fix
fix
fix
fix
fix
* fix
fix
* fix
fix
fix
* [hotfix] Fix ShardFormer test execution path when using sequence parallelism (#5230 )
* fix auto loading gpt2 tokenizer (#5279 )
* [doc] add llama2-13B disyplay (#5285 )
* Update README.md
* fix 13b typo
---------
Co-authored-by: binmakeswell <binmakeswell@gmail.com >
* fix llama pretrain (#5287 )
* fix
* fix
* fix
fix
* fix
fix
fix
* fix
fix
* benchmark gpt2
* fix
fix
fix
fix
* [workflow] fixed build CI (#5240 )
* [workflow] fixed build CI
* polish
* polish
* polish
* polish
* polish
* [ci] fixed booster test (#5251 )
* [ci] fixed booster test
* [ci] fixed booster test
* [ci] fixed booster test
* fix
fix
* fix
fix
fix
* fix
* fix
fix
fix
fix
fix
* fix
* Update shardformer.py
---------
Co-authored-by: digger yu <digger-yu@outlook.com >
Co-authored-by: Frank Lee <somerlee.9@gmail.com >
Co-authored-by: Wenhao Chen <cwher@outlook.com >
Co-authored-by: binmakeswell <binmakeswell@gmail.com >
Co-authored-by: Zhongkai Zhao <kanezz620@gmail.com >
Co-authored-by: Michelle <97082656+MichelleMa8@users.noreply.github.com >
Co-authored-by: Desperado-Jia <502205863@qq.com >
2024-03-04 16:18:13 +08:00
FrankLeeeee
0310b76e9d
Merge branch 'main' into sync/main
2024-03-04 10:09:36 +08:00
yuehuayingxueluo
0aa27f1961
[Inference]Move benchmark-related code to the example directory. ( #5408 )
...
* move benchmark-related code to the example directory.
* fix bugs in test_fused_rotary_embedding.py
2024-02-28 16:46:03 +08:00
yuehuayingxueluo
600881a8ea
[Inference]Add CUDA KVCache Kernel ( #5406 )
...
* add cuda KVCache kernel
* annotation benchmark_kvcache_copy
* add use cuda
* fix import path
* move benchmark scripts to example/
* rm benchmark codes in test_kv_cache_memcpy.py
* rm redundancy codes
* rm redundancy codes
* pr was modified according to the review
2024-02-28 14:36:50 +08:00
QinLuo
bf34c6fef6
[fsdp] impl save/load shard model/optimizer ( #5357 )
2024-02-27 13:51:14 +08:00
Yuanheng Zhao
19061188c3
[Infer/Fix] Fix Dependency in test - RMSNorm kernel ( #5399 )
...
fix dependency in pytest
2024-02-26 16:17:47 +08:00
yuehuayingxueluo
bc1da87366
[Fix/Inference] Fix format of input prompts and input model in inference engine ( #5395 )
...
* Fix bugs in inference_engine
* fix bugs in engine.py
* rm CUDA_VISIBLE_DEVICES
* add request_ids in generate
* fix bug in engine.py
* add logger.debug for BatchBucket
2024-02-23 10:51:35 +08:00
yuehuayingxueluo
2a718c8be8
Optimized the execution interval time between cuda kernels caused by view and memcopy ( #5390 )
...
* opt_view_and_memcopy
* fix bugs in ci
* fix ci bugs
* update benchmark scripts
* fix ci bugs
2024-02-21 13:23:57 +08:00
Jianghai
730103819d
[Inference]Fused kv copy into rotary calculation ( #5383 )
...
* revise rotary embedding
* remove useless print
* adapt
* fix
* add
* fix
* modeling
* fix
* fix
* fix
* fused kv copy
* fused copy
* colossalai/kernel/triton/no_pad_rotary_embedding.py
* del padding llama
* del
2024-02-21 11:31:48 +08:00
