fix: call model.forward() directly instead of @torch .no_grad encode()

custom training loop — bypass SentenceTransformerTrainer

disable gradient checkpointing (NVEmbedModel doesn't support it)

eval + LoRA training + eval in one pipeline

eval-only: use AutoModel directly, no peft/trainer deps

faiss-gpu -> faiss-cpu (search is trivial, avoids CUDA build issues)

add faiss-gpu for evaluation

Add eval+train pipeline with benchmark data

pin transformers==4.42.4 for NV-Embed-v2 compat

fix: total_mem -> total_memory

NV-Embed-v2 ESCO fine-tuning via LoRA

initial commit