Commit Graph

6 Commits (a37f82629d7b9e3c3a0f430b8dd3ff6f38ddf1d4)

Author SHA1 Message Date
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
FoolPlayer d3bc530849 [shardformer] Refactor shardformer api (#4001)
* fix an error in readme

* simplify code

* refactor shardformer

* add todo

* remove slicer

* resolve code review
2023-07-04 16:05:01 +08:00
FoolPlayer f1cb5ac6bf [shardformer] Align bert value (#3907)
* add bert align test, fix dist loss bug

* forward and backward align

* add ignore index

* add shardformer CI

* add gather_output optional for user in shardconfig

* update readme with optional gather_ouput

* add dist crossentropy loss test, remove unused files

* remove unused file

* remove unused file

* rename the file

* polish code
2023-07-04 16:05:01 +08:00
FoolPlayer 8d68de767d [shardformer] init shardformer code structure (#3731)
* init shardformer code structure

* add implement of sharder (inject and replace)

* add implement of replace layer to colossal layer

* separate different layer policy, add some notion

* implement 1d and 2d slicer, can tell col or row

* fix bug when slicing and inject model

* fix some bug; add inference test example
2023-07-04 16:05:01 +08:00
Frank Lee ddcf58cacf
Revert "[sync] sync feature/shardformer with develop" 2023-06-09 09:41:27 +08:00
FoolPlayer 6a69b44dfc [shardformer] init shardformer code structure (#3731)
* init shardformer code structure

* add implement of sharder (inject and replace)

* add implement of replace layer to colossal layer

* separate different layer policy, add some notion

* implement 1d and 2d slicer, can tell col or row

* fix bug when slicing and inject model

* fix some bug; add inference test example
2023-06-08 15:01:34 +08:00