final commit

ghost

ghost

fix: install peft patch as staticmethod to avoid double cls-binding

fix: monkey-patch PeftModel.from_pretrained to force lora requires_grad on resume

fix: install kbit-training grad hooks at from_pretrained (resumed-LoRA path was missing them)

fix: skip real-unsloth chain; use shim deterministically (-runtime image, no nvcc)

fix: default ALLOW_UNSLOTH_SHIM_FALLBACK=1 (-runtime image has no nvcc)

fix: --no-build-isolation + layered unsloth fallbacks (flash-attn build)

fix: unsloth working

fix: enable_input_require_grads so LoRA actually learns

fix: cap baseline at 1 episode/difficulty (3 total) for time-constrained runs

fix: silence transformers max_length/max_new_tokens warning spam

feat: replace unsloth with transformers-backed shim (drop unsloth dep entirely)

fix: pin xformers==0.0.28.post3 (torch 2.5.x compat)

fix: pin unsloth_zoo<2025 to dodge torch.int1 (needs torch 2.6+)

fix: pin torch cu121 before+after unsloth; print version at build

fix: build env venv via virtualenv+pip; add torchvision to agent venv

fix: replace uv sync with virtualenv+pip for agent venv (matches Colab)

fix: pin torch cu121 + install unsloth into venv; bypass uv at runtime

fix: useradd uid 1000 + torch cache dirs to satisfy getpwuid

fix: drop sticky bit on uv cache + chown to uid 1000

fix: install uvicorn for python3.11 (was landing in python3.10)

fix: docker error

deploy: continuous SFT + baselines worker

initial commit