kubeflow / trainer / 18529281041

51%

Build Type

push

github

Committed by web-flow

Commit Message

feat(runtimes): Add LoRA/QLoRA/DoRA support in LLM Trainer V2 (#2832)

* feat(runtimes): Add lora recipe and config in torch.plugin.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* chore: add --trainer-use-lora flag and related logics.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* feat(lora): Add support for QLoRA.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(lora): fix extra rdzv backend in single device mode.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* test(lora): Add UTs for lora/qlora in trainingruntime.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(lora): remove extra quote symbol in lora attn module.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(lint): fix lint error.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(runtime): check default GPU resources allocated in CTRs.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* chore(runtime): add getNumProcPerNode.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(runtime): remove specific vendor for GPU.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* test(runtime): add UTs for lora validation.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(runtimes): remove extra flags for torchtune.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* Update pkg/runtime/framework/plugins/torch/torch.go

Co-authored-by: Andrey Velichkevich <andrey.velichkevich@gmail.com>
Signed-off-by: Shao Wang <2690692950@qq.com>

* chore(runtime): add extractGPUCountFromRuntime.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(runtimes): remove filtering args logics.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(runtimes): use fallbackNumProcPerNode if GPU is set in runtimes.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* refactor(runtimes): refactor getNumProcPerNode logics.

Signed-off-by: Electronic-Waste <2690692950@qq.com>

* fix(lint): fix lint error.

Signed-off-by: E... (continued)

Run Details

107 of 152 new or added lines in 2 files covered. (70.39%)

2 existing lines in 1 file now uncovered.

1250 of 2397 relevant lines covered (52.15%)

0.62 hits per line