Delete generate_plots.py

docs(readme): point blog links to the HF Space copy of blog.md

docs: clarify scenario count, OPENGRID_MODE flag; drop runtime/epoch info

Polish for hackathon submission: training evidence, two pipelines, UI, docs

Fix health check timeout: start UI server in background before training

Print every reward call so terminal shows continuous progress

Reduce prompt/completion length to fix silent OOM on backward pass

Replace env-simulation reward with fast pure-heuristic to fix hang

Fix batch/num_generations: 4/4 on 1 GPU, grad_accum=4

Trigger HF Space rebuild for batch_size=4 fix

Fix GRPO training: reward variance, batch/gen alignment, generation config

Update run_training.py and train_grpo.py, remove Dockerfile.training

Add pre-train gen sanity check, explicit GenerationConfig, dynamic GRPOConfig params, torch_compile/vllm off

QLoRA best practices: prepare_model_for_kbit_training, paged_adamw_8bit, cosine LR, faster iteration

Fix: enable_input_require_grads for gradient checkpointing + 4-bit

Fix OOM: reduce batch/gen/tokens, add grad checkpointing + adafactor

Fix batch size: 8 to match num_generations=8

Drop unsloth: use standard bitsandbytes 4-bit + peft LoRA + TRL GRPOTrainer

Pin transformers <4.52 and unsloth_zoo==2025.11.1 for API compat

Pin TRL <0.16 for unsloth 2025.11.1 GRPO API compatibility

Dynamic NVIDIA lib path discovery in entrypoint for bitsandbytes

Add LD_LIBRARY_PATH for pip-installed NVIDIA libs (bitsandbytes fix)

Add torchvision and hf_transfer

Remove torchao entirely - transformers handles absence gracefully

Fix: install torchao 0.8.0 separately, unsloth --no-deps to avoid torchao>=0.13 conflict

Pin compatible versions: torch 2.6.0 + torchao <0.9 + transformers <5.0

Update CUDA to 12.4.1 and unpin PyTorch version to fix torchao/int1 compatibility

Remove --no-deps to allow installing sub-dependencies like regex

Add explicit compiler location debugging in run_training.py

Explicitly install gcc, g++, python3-dev and set CC/CXX

Add build-essential for Triton compilation

GRPO training with CUDA + results in UI

fix: notebook uses compute_grpo_reward_env, updated hyperparams, no emojis

fix: remove ENV OPENGRID_MODE to avoid HF secrets collision

fix: transformers>=4.51.3 to resolve unsloth dep conflict

fix: add unsloth back with pinned versions to avoid dep backtracking

fix: lean Dockerfile + remove unsloth from training deps

fix: unified Dockerfile with entrypoint for server/training mode

fix: update .dockerignore for new repo structure

feat: curriculum training + Karnataka scenarios + repo cleanup

restore env Dockerfile

training dockerfile

fix: total_memory attribute name

Restore env Dockerfile

Use training Dockerfile for GPU Space

Add GRPO training runner for HF Spaces GPU

OpenGrid: Multi-agent POMDP power grid environment with GRPO training