fix: set per_device_eval_batch_size=1 to prevent OOM during evaluation loop

perf: increase batch size to 2 and enable group_by_length to massively speed up training

fix: change device_map='auto' to local_process_index to prevent cross-GPU loss calculation error in transformers 4.46

fix: set remove_unused_columns=True so Trainer drops 'messages' list before data collation

fix: change max_length to max_seq_length in SFTConfig for trl 0.12.1 compatibility

refactor: address Phase 2 review findings

fix: LoRA adapts all 7 projections via Phi-3 fused names + guard assertion; r=16, Spider data, max_length=2560

feat: Spider data pipeline pinned to minktn/spider-data (per-example CREATE TABLE schema); drop find_spider.py

Complete CPU GGUF serving + docs + minimal UI redesign

Revert "Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)"

Revert "Make torch/transformers/peft lazy imports so CPU Space boots without them"

Make torch/transformers/peft lazy imports so CPU Space boots without them

Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)

Fix bugs

Fix bugs

Update CPU warning banner with accurate load and query times

Fix inference device bug, auto-load model, remove fine-tuning console, UI cleanup

initial deployment