fix: multimodal-LM (Qwen3.5_5) support for ex-LRP - #1

Open

mann1x wants to merge 13 commits into

Tusm11:feature/ex-lrpfrom

mann1x:fix/multimodal-lm-support

Commits on Apr 25, 2026

docs: Update README with eX-LRP documentation and features
Tusm11
committed
90a9d86

Commits on Apr 26, 2026

Address PR feedback: wire lrp_scores, migrate to lxt for true AttnLRP, add LRP mask test, and clean up files
Tusm11
committed
0125b2e
Fix LRP bugs: add safetensors support, implement global cache, support multiple architectures, and enforce explicit lrp_scores
Tusm11
committed
3e92daf
Fix pipeline argument parsing, enforce strict LRP scores without fallback, add lru_cache for memory safety, and clean up unused LRP parameters
Tusm11
committed
4ac9c1d
Fix test fallback logic, use dynamic mergekit-yaml path, and cleanly restore model state after LRP backward pass
Tusm11
committed
a899412
Guarantee gradient checkpointing restoration via try/finally and add lxt to pyproject extras
Tusm11
committed
d6e8dd1
Fix build_mask import, correctly trace embedding gradients in LRP, isolate LRP score output directories, and patch notebook paths
Tusm11
committed
ae32700
Fix tied embeddings in LRP, remove incompatible 4-bit loading, fix left-padded attention masks, and fix pipeline device arguments
Tusm11
committed
26571b8
fix: Address Cursor Bugbot issues - CUDA fallback, Qwen matching, subprocess interpreter
Tusm11
committed
2022d37
fix: Address 4 Cursor Bugbot issues - weight normalization, cache mutation, float16 underflow
Tusm11
committed
9c039e6
fix: Remove misleading dead code - unused total_weight validation
Tusm11
committed
a928e72
fix: Address 2 pipeline issues - fail fast on missing paths, resolve script location
Tusm11
committed
cfcfc82

Commits on Apr 30, 2026

fix: multimodal-LM (Qwen3.5_5) support for ex-LRP

mann1x
and
claude
committed
ef76d4b