fix: reduce grpo training runtime

fix: parse chat completions in rewards

fix: emit prompt column for grpo trainer

fix: add grpo warnings compatibility attr

fix: avoid build-time root-owned cache env dirs

fix: select writable cache root at startup

fix: set writable cache env vars for training space

fix: write grpo checkpoints to tmp in space

fix: use writable runtime dirs before training imports

chore: update comment on torchao shim strategy

fix: comprehensive torchao dtype + register_constant shims

revert: back to stable cu121 torch — patch applied at runtime

fix: shim register_constant before importing unsloth (no-op safe for non-torchao training)

fix: use PyTorch nightly — torchao needs register_constant not in stable 2.6

fix: switch to cu124 wheels for torch==2.6.0 (cu121 tops out at 2.5.1)

fix: upgrade torch to 2.6.0 — torchao requires torch.int1 added in 2.6

fix: verbose step-by-step logs, catch BaseException with full traceback

fix: capture transformers/TRL logging into UI via logging.Handler

fix: install unsloth_zoo before unsloth

fix: remove runtime unsloth install, verify import at startup

fix: install unsloth in Dockerfile at build time (not at click-time)

fix: read pip output in chunks splitting on and

feat: stream pip output live to UI logs during unsloth install

fix: install unsloth to /app/pkgs (writable), detect torch version for git extra

fix: capture unsloth install errors, try PyPI then git fallback

fix: replace demo.load(every=) with gr.Timer for Gradio 4.x compat

chore: upload pre-generated facts.json to avoid network dependency at build

fix: space_runner installs unsloth at click-time on live GPU

fix: use python:3.11-slim base, install unsloth at runtime

feat: training space with manual start UI

initial commit