onlycoding135/constrained-refactor-gauntlet

Upload UI files and Docker build config

Enable GPU PyTorch and mock local UI

docs: update project architecture, reward logic, and documentation, and add anti-cheat validation to the server environment

docs: Update README and root endpoint with Track C endpoints for HF Spaces

feat: Add Green-Code Optimizer (Track C)

fix: add HF Spaces sdk:docker metadata to README.md

fix: graceful CPU fallback for HF Spaces deployment

feat: Dockerize for HF Spaces deployment + adapter upload script

fix: zero loss — more diverse generations + granular reward function

fix: disable vLLM entirely — v0.19.1 has BitsAndBytes graph bug on ALL GPUs

feat: increase verify_pipeline to 5 steps, 4 generations for meaningful signal

fix: use fp16 on T4 (bf16 requires Ampere+), conditionally pass use_vllm

fix: auto-detect GPU capability, disable vLLM on T4 (compute < 8.0)

fix: make GPU imports conditional so reward function works on CPU-only machines

feat: integrate Unsloth for 2x faster GRPO training + 60% VRAM reduction

feat: add inference script for model evaluation with LoRA adapter support

feat: add verify_pipeline.py + reward timing logs + faster generation (512 tokens)

perf: add reward timing logs + reduce completion length to 512 for faster iterations

Stabilize GRPO training for H100 and add verification tools

fix: absolute paths for output_dir and adapter save

fix: add wandb import for TRL checkpoint save + sdpa attention

fix: use sdpa attention (flash-attn not installed)

chore: upgrade to 7B model, remove 4-bit quantization for native bf16 training, and tune GRPO parameters for H100 GPU

fix: use dtype for transformers 5.0 compatibility

refactor: implement deterministic rule resolution and replace subprocess-based code quality evaluation with faster AST-based analysis

chore: disable automatic wandb logging in GRPO training configuration

feat: implement character-based code context truncation and remove max_prompt_length constraint in GRPO trainer

feat: update GRPO trainer configuration and downgrade base model to Qwen2.5-Coder-3B

fix: correct invalid GRPOConfig arguments added in previous commit

fix: add memory allocation optimization and reduce completion length to prevent OOM

fix: reduce num_generations to 2 to prevent CUDA OutOfMemory on Colab T4

fix: change generation_batch_size to 4 to satisfy TRL divisible requirement

chore: reduce max_steps to 25 for testing on Colab

fix: remove invalid max_prompt_length argument from GRPOConfig

feat: optimize training script for Colab T4/L4 GPUs with 4-bit quantization

chore: update training pipeline, rule engine, and dependencies for RL training

Update README.md

feat: reduce num_generations from 8 to 4 in GRPO training configuration

Fix OOM by adding 4-bit quantization

Initial commit of Constrained Refactor Gauntlet