code/Abforge_Training/scripts/smoke_repro_env.py

#!/usr/bin/env python
"""Smoke-test the rebuilt ABForge env. Verifies the critical imports + versions match the
frozen spec, loads the base tokenizer/config, and (if --gpu) runs a tiny vLLM generate.
Exit 0 only if all critical checks pass.

  python scripts/smoke_repro_env.py            # CPU import + version check
  python scripts/smoke_repro_env.py --gpu      # + vLLM load/generate (needs a GPU)

NOTE: everything runs under `if __name__ == "__main__"` — vLLM spawns worker processes that
re-import this module, and without the guard they'd re-execute the whole script.
"""
import sys, importlib

EXPECT = {  # critical pins from docs/env_abforge_vllm_frozen.txt
    "torch": "2.6.0", "vllm": "0.8.5.post1", "transformers": "4.52.1",
    "flash_attn": "2.8.3", "ray": "2.43.0", "xformers": "0.0.29.post2",
}
# Use the HF hub model — the scratch copy under models/Qwen3-8B is being evicted (config.json gone).
BASE = "Qwen/Qwen3-8B"


def main():
    ok = True
    print("== imports / versions ==")
    for mod, want in EXPECT.items():
        try:
            m = importlib.import_module(mod)
            got = getattr(m, "__version__", "?")
            mark = "OK " if str(got).startswith(want) else "DIFF"
            if mark == "DIFF": ok = False
            print(f"  [{mark}] {mod:14s} got={got} want~={want}")
        except Exception as e:
            ok = False
            print(f"  [ERR] {mod:14s} {type(e).__name__}: {str(e)[:90]}")

    print("== verl (bundled) ==")
    try:
        import verl
        print(f"  [OK ] verl {getattr(verl,'__version__','?')}")
    except Exception as e:
        ok = False; print(f"  [ERR] verl {type(e).__name__}: {str(e)[:90]}")

    print("== torch CUDA ==")
    try:
        import torch
        print(f"  cuda_available={torch.cuda.is_available()} built_cuda={torch.version.cuda} "
              f"devices={torch.cuda.device_count() if torch.cuda.is_available() else 0}")
    except Exception as e:
        print(f"  [ERR] {e}")

    print("== tokenizer/config load (base Qwen3-8B) ==")
    try:
        import os
        from transformers import AutoConfig, AutoTokenizer
        src = BASE if os.path.isdir(BASE) else "Qwen/Qwen3-8B"
        AutoConfig.from_pretrained(src)
        AutoTokenizer.from_pretrained(src)
        print(f"  [OK ] loaded from {src}")
    except Exception as e:
        ok = False; print(f"  [ERR] {type(e).__name__}: {str(e)[:120]}")

    if "--gpu" in sys.argv:
        try:
            import torch
            if torch.cuda.is_available():
                print("== vLLM tiny generate ==")
                from vllm import LLM, SamplingParams
                llm = LLM(model="Qwen/Qwen3-8B", gpu_memory_utilization=0.85,
                          max_model_len=2048, enforce_eager=True)
                out = llm.generate(["Hello, the ablation study"], SamplingParams(max_tokens=8))
                print("  [OK ] generated:", repr(out[0].outputs[0].text[:60]))
        except Exception as e:
            ok = False; print(f"  [ERR] vLLM generate {type(e).__name__}: {str(e)[:160]}")

    print("\nSMOKE_OK" if ok else "\nSMOKE_FAILED")
    sys.exit(0 if ok else 1)


if __name__ == "__main__":
    main()