Add shared outer lora - #3958
Conversation
|
/claude review |
|
Light Code Review - Critical: AssertionError typo (utils.py:1676) - except AssertionError will never match, Python has no built-in AssertionError, at runtime this raises NameError instead of gracefully returning, must be AssertionError, see inline comment. Minor: Missing trailing newline at end of utils.py, see inline comment. Test coverage: This PR adds three new classes but no tests are included. The existing GroupedExpertLinearAdapter has extensive unit tests, the new adapter should have comparable coverage. Suggested test cases: SharedOuterGroupedExpertAdapter forward pass fc1 and fc2 paths, sharded_state_dict for both, PackedPerExpertLinear forward and sharded_state_dict, _make_cross_ep_replicated broadcast and gradient hook, LoRA.transform with experts_shared_outer_loras=True, stream_adapter_weights_megatron_to_hf shared-outer branch for fc1 and fc2, shared-outer fused fc1 linear_out slicing. No perf tests impacted. |
83ab19e to
3587d82
Compare
|
/ok to test 46663cf |
|
/ok to test eaa9592 |
|
/ok to test f789f92 |
|
/ok to test 8e3a2eb |
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
|
/ok to test 5fbd49d |
Signed-off-by: adityavavreNVDA <avavre@nvidia.com>
|
/ok to test 20136f5 |
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
What does this PR do ?
follow the shared outer lora design here sgl-project/sglang#21466.
e2e from radixark#3
Changelog
GitHub Actions CI
See the CI section in the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.
Before your PR is "Ready for review"
Pre checks:
If you haven't finished some of the above items you can still open "Draft" PR.
Additional Information