# Optional: copy to sql_compare_ui_qwen/.env for UI-only overrides. # Repo-root .env is also loaded (HF_TOKEN, QWEN_MODEL_ID, QWEN_COMPARE_*). # Hub base model QWEN_COMPARE_HUB_MODEL_ID=Qwen/Qwen3.5-0.8B # Fine-tuned model is hard-coded in the app: Tuana/qwen35-08b-text2sql. # Optional token for gated Hub models (else HF_TOKEN from repo .env) # QWEN_COMPARE_HF_TOKEN= # Max new tokens for both columns # QWEN_COMPARE_MAX_NEW_TOKENS=512 # Unload fine-tuned model before Hub base, then unload Hub base (peak ~one large model) # QWEN_COMPARE_SEQUENTIAL_UNLOAD=true # Single-column mode # QWEN_COMPARE_SKIP_HUB=true # QWEN_COMPARE_SKIP_FINETUNED=true # Read-only SQLite for extracted SQL (default: repo data/spider_eval_synthetic/synthetic.db) # QWEN_COMPARE_DB_PATH= # Override default DDL in the prompt (multiline); if unset, the built-in synthetic schema is used. # QWEN_COMPARE_DB_SCHEMA= # Model load: unset = CUDA auto if available, else Apple MPS if available, else CPU. # Explicit: auto | mps | cpu | none | cuda:0. # QWEN_COMPARE_DEVICE_MAP= # QWEN_COMPARE_HUB_DEVICE_MAP= # MPS weight dtype: fp16 | bf16 | float32 (default float32 — set fp16 for speed if stable) # QWEN_COMPARE_MPS_DTYPE= # Gradio bind # QWEN_COMPARE_GRADIO_HOST=127.0.0.1 # QWEN_COMPARE_GRADIO_PORT=7861 # Set to 1 to show multiprocessing resource_tracker semaphore warnings # QWEN_COMPARE_SHOW_RESOURCE_TRACKER_WARNINGS=false