Commit History
-
Add regression tests for the wrapped-value-module gradient bug
Verified signatureCovers both the training-time gradient loss (confirmed against a real PEFT LoRA adapter) and the separate inference-time correctness issue (fused path reading the unadapted base weight for an unmerged adapter under no_grad single-token decode).cf6a52e
-
Upload tests/test_sparse_cmix.py with huggingface_hub
Verified signaturee8907ef