Refactor QKV Fusion Utilities #14047
dg845
wants to merge
19 commits into
main
from
refactor/lora-aware-qkv-fusion
Refactor QKV fusion: add inplace option, model-level helper to restor…
4e188842
Add tests for refactored QKV fusion API in AttentionMixin/AttentionMo…
658f0719
Fix Claude review comments (typo, get_added_qkv docstring)
37c0f70c
make style and make quality
b761f4a3
Merge branch 'main' into refactor/lora-aware-qkv-fusion
a7512dbd
Add inplace argument to model-overriden methods for compatibility (W…
e41dee65
Use inplace=False in ChronoEdit for consistency
799b8cf7
dg845
changed the title
Merge branch 'main' into refactor/lora-aware-qkv-fusion
0b0e3e77
Merge branch 'main' into refactor/lora-aware-qkv-fusion
41577e9e
Guard QKV fusion on LoRA adapter state, not module names
712483bb
Implement inplace QKV fusion for the Wan-family transformers
7c2bbfa0
Fix get_qkv ignoring encoder_hidden_states when fused
c3b34869
Drop restore_checkpoint_fusion_state
d8300f9e
Build the unfused projections on the meta device
84eb5d53
Merge branch 'main' into refactor/lora-aware-qkv-fusion
26d39f08
Move model-level QKV fusion tests into AttentionTesterMixin
bf553c4f
Check LoRA-guarded unfusing and split-projection restoration per model
2b9edfbc
Use plain truthiness for fused_projections assertions
7235091b
Document the inplace fusion argument
9910574d
Reviewers
Assignees
No one assigned
Labels
models
tests
size/L
utils
Milestone
No milestone
Login to write a write a comment.
Login via GitHub