#!/usr/bin/env python3
"""Run in your workload's Python environment. Prints a read-only Doctor report.

No packages are installed, credentials collected, or data sent over the network.
A tiny GPU operation verifies execution; localhost ports are checked for a TCP
listener only. Run from your intended data directory to check that filesystem.
"""
import csv
import importlib.metadata
import json
import platform
import shutil
import socket
import subprocess


def package(name):
    try:
        return importlib.metadata.version(name)
    except importlib.metadata.PackageNotFoundError:
        return None


def collect():
    report = {"schema_version": 1, "python_version": platform.python_version(),
              "torch_version": package("torch"), "vllm_version": package("vllm"),
              "gpus": [], "ports": {}, "probe_errors": []}
    try:
        report["disk_free_gb"] = round(shutil.disk_usage(".").free / (1024 ** 3), 2)
    except OSError:
        report["probe_errors"].append("Disk check unavailable")
    try:
        result = subprocess.run(["nvidia-smi", "--query-gpu=name,driver_version,memory.total", "--format=csv,noheader,nounits"],
                                capture_output=True, text=True, timeout=10, check=True)
        for row in list(csv.reader(result.stdout.splitlines()))[:16]:
            name, driver, memory = (value.strip() for value in row)
            report["driver_version"] = driver
            report["gpus"].append({"name": name[:120], "memory_gb": round(float(memory) / 1024, 2)})
    except (OSError, subprocess.SubprocessError, ValueError):
        report["probe_errors"].append("nvidia-smi unavailable")
    # Isolate CUDA imports/initialization so a stuck driver cannot hang Doctor.
    if report["torch_version"]:
        code = '''import json, torch
r = {"cuda_available": torch.cuda.is_available(), "torch_cuda_version": torch.version.cuda, "cuda_archs": torch.cuda.get_arch_list() if torch.cuda.is_available() else [], "kernel_test_ok": False}
gpus = []
try:
    for i in range(min(torch.cuda.device_count(), 16)):
        p = torch.cuda.get_device_properties(i)
        free, total = torch.cuda.mem_get_info(i)
        gpus.append({"name": p.name, "memory_gb": round(total / 1024**3, 2), "free_memory_gb": round(free / 1024**3, 2), "compute_capability": str(p.major)+"."+str(p.minor)})
    if r["cuda_available"]:
        x = torch.ones(1, device="cuda")
        r["kernel_test_ok"] = bool((x + 1).item() == 2)
except Exception:
    r["kernel_test_ok"] = False
if gpus: r["gpus"] = gpus
print(json.dumps(r))
'''
        try:
            import sys
            result = subprocess.run([sys.executable, "-c", code], capture_output=True, text=True, timeout=30, check=True)
            report.update(json.loads(result.stdout.splitlines()[-1]))
        except (OSError, subprocess.SubprocessError, ValueError, IndexError):
            report["probe_errors"].append("PyTorch GPU check unavailable")
    for port in (8000, 8188, 8888):
        try:
            with socket.create_connection(("127.0.0.1", port), timeout=1):
                report["ports"][str(port)] = True
        except OSError:
            report["ports"][str(port)] = False
    return report


if __name__ == "__main__":
    print(json.dumps(collect(), indent=2))