File size: 3,791 Bytes
22decf5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | """One-time machine setup: container images, HF cache, benchmark test data. Idempotent."""
import os
import shutil
from . import config, sandbox
# `claude_non_api*` scaffolds abort unless this file exists; upstream fills it with a subscription
# token. Here the passthrough authenticates the request instead (see scripts/x2p_anthropic_shim.py),
# so the file only has to be present. Matches what the cloud api-key-helper hands Claude Code.
OAUTH_PLACEHOLDER = "cloud-proxy-auth-placeholder"
def setup(containers: list[str], models: list[str], tasks: list[str], full_cache: bool) -> None:
for container in containers:
if not config.image(container).is_file():
print(f"Building {config.image(container)} (takes a while) ...")
sandbox.build_image(container)
# Upstream's trace sanitizer redacts the API keys listed in <upstream>/.env.
upstream_env = config.UPSTREAM / ".env"
if config.ENV_FILE.is_file() and not upstream_env.exists():
upstream_env.symlink_to(config.ENV_FILE)
for agent in (a for a in config.AGENTS if "non_api" in a):
token = config.UPSTREAM / "agents" / agent / "oauth_token"
if not token.exists():
token.write_text(OAUTH_PLACEHOLDER)
scratch = config.DATA_DIR / "setup_sandbox"
hf_token = os.environ.get("HF_TOKEN", "")
tool_container = containers[-1] if containers else config.EVAL_CONTAINER
if full_cache:
# Upstream's `containers/download_hf_cache/download_hf_cache.sh`: 14 models and 382 dataset
# splits, the working set the agent is expected to have without reaching the internet.
print("Downloading the full upstream HF cache (hundreds of GB, hours) ...")
# Neither container image bundles hf_transfer; install it into the writable tmpfs overlay
# before asking huggingface_hub to use it, rather than baking it into the image.
script = config.UPSTREAM / "containers/download_hf_cache/download_resources.py"
exit_code = sandbox.run([
"bash", "-c", f"uv pip install --system --quiet hf_transfer && python {script}"],
scratch=scratch, hf_cache=config.HF_CACHE,
container=tool_container, binds={config.UPSTREAM: str(config.UPSTREAM)},
env={"HF_TOKEN": hf_token, "HF_HUB_ENABLE_HF_TRANSFER": "1"})
if exit_code:
raise SystemExit("Cache download failed (gated repos such as google/gemma-3-4b-pt and "
f"Idavidrein/gpqa need HF_TOKEN in {config.ENV_FILE}).")
for model in models:
print(f"Downloading {model} ...")
exit_code = sandbox.run(["hf", "download", model], scratch=scratch, hf_cache=config.HF_CACHE,
container=tool_container, env={"HF_TOKEN": hf_token})
if exit_code:
raise SystemExit(f"Download of {model} failed (gated models need HF_TOKEN in {config.ENV_FILE}).")
# Test sets feed the agent's decontamination checker; upstream writes them into its task dirs.
missing = [t for t in tasks if not (config.UPSTREAM / "src/eval/tasks" / t / "test_data.json").is_file()]
if missing:
print(f"Downloading test data for {missing} ...")
script = config.UPSTREAM / "src/judges/test_data_download/download_test_data.py"
exit_code = sandbox.run(["python", str(script), "--tasks", *missing], scratch=scratch,
hf_cache=config.HF_CACHE, container=tool_container,
binds={config.UPSTREAM: str(config.UPSTREAM)},
env={"MY_HF_TOKEN": hf_token})
if exit_code:
raise SystemExit("Test data download failed (gpqamain is gated and needs HF_TOKEN).")
shutil.rmtree(scratch, ignore_errors=True)
|