platform/aiml/mlops/agent_tools/tools_model.py

from __future__ import annotations

import json
import os
import signal
import subprocess
from typing import Any, Dict

import requests

from .registry import ToolRegistry


VLLM_BASE_URL = os.getenv("VLLM_BASE_URL", "http://localhost:8000/v1")


def t_serve_status(_: Dict[str, Any]) -> str:
    try:
        r = requests.get(f"{VLLM_BASE_URL.rstrip('/')}/health", timeout=5)
        return json.dumps({"status": r.status_code, "body": r.text})
    except Exception as e:
        return json.dumps({"error": str(e)})


def t_vllm_reload(args: Dict[str, Any]) -> str:
    """Attempt to restart the vLLM server via optional script or PID.
    Args: script (path) OR pid (int)
    """
    script = args.get("script")
    pid = args.get("pid")
    if script:
        try:
            proc = subprocess.run(["bash", script], capture_output=True, text=True, timeout=180)
            return json.dumps({"returncode": proc.returncode, "stdout": proc.stdout, "stderr": proc.stderr})
        except Exception as e:
            return json.dumps({"error": str(e)})
    if pid:
        try:
            os.kill(int(pid), signal.SIGHUP)
            return json.dumps({"status": "signaled", "pid": int(pid)})
        except Exception as e:
            return json.dumps({"error": str(e)})
    return json.dumps({"error": "script or pid required"})


def t_hf_pull_model(args: Dict[str, Any]) -> str:
    """Pull/refresh a HF repo into MODEL_PATH using hf CLI.
    Args: repo (org/name), dest (MODEL_PATH)
    """
    repo = args.get("repo")
    dest = args.get("dest") or os.getenv("MODEL_PATH", "/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft")
    token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_API_KEY")
    if not repo:
        return json.dumps({"error": "repo required"})
    if not token:
        return json.dumps({"error": "HF_TOKEN not set"})
    try:
        proc = subprocess.run([
            "hf", "download", str(repo), "--repo-type", "model", "--include", "**", "--local-dir", str(dest)
        ], capture_output=True, text=True, timeout=3600)
        return json.dumps({"returncode": proc.returncode, "stdout": proc.stdout[-4000:], "stderr": proc.stderr[-4000:], "dest": dest})
    except Exception as e:
        return json.dumps({"error": str(e)})


def t_promote_checkpoint(args: Dict[str, Any]) -> str:
    """Promote a trained checkpoint to MODEL_PATH (rsync copy).
    Args: src (path), dest (optional overrides MODEL_PATH)
    """
    src = args.get("src")
    dest = args.get("dest") or os.getenv("MODEL_PATH", "/data/adaptai/platform/aiml/checkpoints/qwen3-8b-elizabeth-sft")
    if not src:
        return json.dumps({"error": "src required"})
    try:
        proc = subprocess.run(["rsync", "-aH", f"{src}/", f"{dest}/"], capture_output=True, text=True, timeout=3600)
        return json.dumps({"returncode": proc.returncode, "stdout": proc.stdout[-4000:], "stderr": proc.stderr[-4000:], "dest": dest})
    except Exception as e:
        return json.dumps({"error": str(e)})


def t_self_train(args: Dict[str, Any]) -> str:
    """Launch a training process (unconstrained). Provide 'script' and 'args' list.
    Example: {"script": "./train_elizabeth.sh", "args": ["--lr", "2e-5"]}
    """
    script = args.get("script")
    sargs = args.get("args") or []
    if not script:
        return json.dumps({"error": "script required"})
    try:
        cmd = ["bash", script] + list(map(str, sargs))
        proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
        return json.dumps({"status": "started", "pid": proc.pid, "cmd": cmd})
    except Exception as e:
        return json.dumps({"error": str(e)})


def register_tools(reg: ToolRegistry) -> None:
    reg.register(
        name="serve_status",
        description="Check vLLM /health upstream.",
        parameters={"type": "object", "properties": {}},
        handler=t_serve_status,
    )
    reg.register(
        name="vllm_reload",
        description="Reload/restart vLLM via script or send SIGHUP to a PID.",
        parameters={"type": "object", "properties": {"script": {"type": "string"}, "pid": {"type": "integer"}}},
        handler=t_vllm_reload,
    )
    reg.register(
        name="hf_pull_model",
        description="Pull/refresh a Hugging Face model into MODEL_PATH.",
        parameters={"type": "object", "properties": {"repo": {"type": "string"}, "dest": {"type": "string"}}, "required": ["repo"]},
        handler=t_hf_pull_model,
    )
    reg.register(
        name="promote_checkpoint",
        description="Promote a trained checkpoint into serving MODEL_PATH using rsync.",
        parameters={"type": "object", "properties": {"src": {"type": "string"}, "dest": {"type": "string"}}, "required": ["src"]},
        handler=t_promote_checkpoint,
    )
    reg.register(
        name="self_train",
        description="Launch an unconstrained training job via provided script and args.",
        parameters={"type": "object", "properties": {"script": {"type": "string"}, "args": {"type": "array", "items": {"type": "string"}}}, "required": ["script"]},
        handler=t_self_train,
    )