Fix torch bfloat16 errors on T4 GPUs by enabling Unsloth dtype auto-detection and explicitly wrapping forward passes in autocast

Set TRITON_CACHE_DIR to /tmp/triton_cache to avoid root permission denied error

Add mandatory Unsloth inference state toggles around generation for RL pipeline

Pin pydantic, fastapi, and starlette to fix Gradio 4.x JSON schema and TemplateResponse bugs

Pin Gradio to 4.36.1 to fix TypeError during json schema parsing on startup

Fix Gradio runtime error by moving theme to gr.Blocks

Pin huggingface-hub to 0.24.7 to fix Unsloth _token import error

Capture and display exact Unsloth import exception

Switch SDK to docker to use custom Dockerfile and fix pip build

Fix Gradio sdk_version to a valid fully-specified version (4.44.0)

Add HuggingFace Space configuration reference to README

Update with existing environment

Fix GRPO group-loss training and align UI defaults.

Migrate LLM pipeline to custom GRPO with robust rewards

Multi-model benchmark pipeline: VRAM cleanup + EMA graph + detailed output

Fix truncation: 80 tokens, regex safety net, strict prompt

Hackathon speedrun: max_new_tokens=32, seq_len=512 for 4-8x faster iterations

Replace flash-attn with PyTorch built-in SDPA (no CUDA compile needed)

Fix: install torch before flash-attn (needs torch at build time)

Max GPU utilization: flash-attn2 + grad accumulation + 15 steps/ep + 1024 seq len

Cap LLM iterations at 50 to prevent timeout on 8B models

Switch to Llama 3.1 8B + fix low-timestep crash (min 5000)

Pin torch/transformers/peft versions to fix cache conflict

Fix permission: mkdir after COPY, chmod /app

Fix matplotlib permission + HF cache dirs

Add LLM RL training with Gemma 7B + LoRA

Fix Gradio 6.0 theme deprecation

Add Cloud Arena Mathematical Model RL environment

initial commit