Buckets:

Mercity
/

red-teaming

Files

xet

Mercity/red-teaming / 29-10-progress /setup.sh

Rishik001

10 days ago

download

raw

2.02 kB

#!/usr/bin/env bash
# Install pinned dependencies and pre-download the attacker model. Run once on the pod (2x H100 for async GRPO).
# vLLM pins its own torch (2.13.0), so the image's PyTorch gets replaced; any CUDA 12.8+/13 image works.
set -euo pipefail

# The attacker (abliterated Qwen in safetensors, so vLLM loads it directly); keep in sync with config.yaml.
MODEL_REPO="${MODEL_REPO:-lukey03/Qwen3.5-9B-abliterated}"

# TRL main, pinned: v1.16.0's AsyncGRPOTrainer refuses Qwen3.5's linear-attention layers; main accepts them
# with transformers >= 5.15. Bump this sha deliberately, never float on main.
TRL_COMMIT=f66d9e832b92f02a1adf334a6c2a7ab7f9390809

# vLLM first: it pins torch==2.13.0 and transformers>=5.16.1,<5.20.0. TRL main allows vllm<0.33.
pip install -q "vllm==0.32.0"
# 5.19.0 sits inside vLLM's range and is the version prime-rl pairs with vLLM 0.32.0.
pip install -q "transformers==5.19.0"
pip install -q "trl @ git+https://github.com/huggingface/trl.git@${TRL_COMMIT}"
pip install -q "peft==0.21.2" "accelerate==1.15.0" "datasets==5.1.0" "wandb==0.31.0" \
    "huggingface_hub==2.3.0" "aiohttp==3.14.5" "requests==2.34.2" "python-dotenv==1.2.4" "pyyaml==6.0.3"
# Qwen3.5's Gated DeltaNet layers fall back to a ~20x slower PyTorch scan without flash-linear-attention;
# kernels fetches flash-attn3 for AsyncGRPO's tree packing on H100.
pip install -q "flash-linear-attention==0.6.0" "kernels==0.18.0"
# Data generation (datagen/) uses Simula, installed separately.

python -c "import torch, vllm, transformers, trl, peft; print('torch', torch.__version__, '| vllm', vllm.__version__, '| transformers', transformers.__version__, '| trl', trl.__version__, '| peft', peft.__version__)"

# huggingface_hub >= 1.0 ships the `hf` CLI (huggingface-cli was removed) and downloads via Xet by default.
hf download "$MODEL_REPO"            # warm the HF cache; vLLM reads from it

echo "Downloaded $MODEL_REPO into the HF cache."
echo "Now create .env with OPENROUTER_BASE_URL and OPENROUTER_API_KEY."

Xet Storage Details

Size:: 2.02 kB
Xet hash:: 5196dc2f6a4eb66d0d902653ffe437f7b4358a7c36144ea135f9fdd6489dd097

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	#!/usr/bin/env bash
	# Install pinned dependencies and pre-download the attacker model. Run once on the pod (2x H100 for async GRPO).
	# vLLM pins its own torch (2.13.0), so the image's PyTorch gets replaced; any CUDA 12.8+/13 image works.
	set -euo pipefail

	# The attacker (abliterated Qwen in safetensors, so vLLM loads it directly); keep in sync with config.yaml.
	MODEL_REPO="${MODEL_REPO:-lukey03/Qwen3.5-9B-abliterated}"

	# TRL main, pinned: v1.16.0's AsyncGRPOTrainer refuses Qwen3.5's linear-attention layers; main accepts them
	# with transformers >= 5.15. Bump this sha deliberately, never float on main.
	TRL_COMMIT=f66d9e832b92f02a1adf334a6c2a7ab7f9390809

	# vLLM first: it pins torch==2.13.0 and transformers>=5.16.1,<5.20.0. TRL main allows vllm<0.33.
	pip install -q "vllm==0.32.0"
	# 5.19.0 sits inside vLLM's range and is the version prime-rl pairs with vLLM 0.32.0.
	pip install -q "transformers==5.19.0"
	pip install -q "trl @ git+https://github.com/huggingface/trl.git@${TRL_COMMIT}"
	pip install -q "peft==0.21.2" "accelerate==1.15.0" "datasets==5.1.0" "wandb==0.31.0" \
	"huggingface_hub==2.3.0" "aiohttp==3.14.5" "requests==2.34.2" "python-dotenv==1.2.4" "pyyaml==6.0.3"
	# Qwen3.5's Gated DeltaNet layers fall back to a ~20x slower PyTorch scan without flash-linear-attention;
	# kernels fetches flash-attn3 for AsyncGRPO's tree packing on H100.
	pip install -q "flash-linear-attention==0.6.0" "kernels==0.18.0"
	# Data generation (datagen/) uses Simula, installed separately.

	python -c "import torch, vllm, transformers, trl, peft; print('torch', torch.__version__, '\| vllm', vllm.__version__, '\| transformers', transformers.__version__, '\| trl', trl.__version__, '\| peft', peft.__version__)"

	# huggingface_hub >= 1.0 ships the `hf` CLI (huggingface-cli was removed) and downloads via Xet by default.
	hf download "$MODEL_REPO" # warm the HF cache; vLLM reads from it

	echo "Downloaded $MODEL_REPO into the HF cache."
	echo "Now create .env with OPENROUTER_BASE_URL and OPENROUTER_API_KEY."