code/autoslm/serve/deploy.py

"""Serve a trained LoRA adapter for OpenAI-style chat via a managed RunPod Flash GPU.

Two deployment modes (the cost/latency trade-off is explicit and user-chosen):

- ``dev`` (default): scale-to-zero. ``workers=(0,1)`` with a configurable idle timeout
  and FlashBoot — you accept a cold start after inactivity, and pay $0 while idle.
- ``always-on``: ``workers=(1,1)`` — one worker stays warm 24/7 (no cold starts,
  continuous billing). ``slm deployments`` shows the projected $/day so the cost is
  never a surprise.

Each run gets its OWN uniquely-named serve endpoint (``autoslm-serve-<gpu>-<run>``), so
deployments never fight over a shared endpoint config and ``slm undeploy <run_id>``
can tear down exactly one deployment (via the REST API, from any process).

The handler boots vLLM with the base model + the LoRA adapter (pulled from the HF
dataset repo the trainer streamed it to) and returns an OpenAI-shaped chat-completion.
"""

from __future__ import annotations

import os
from dataclasses import asdict, dataclass

from autoslm._logging import get_logger
from autoslm.providers.base import canonical_gpu, gpu_short
from autoslm.providers.runpod.gpus import flash_gpu

logger = get_logger(__name__)


def _invoke_handler(handler, payload: dict) -> dict:
    """Call a Flash serve handler, awaiting it if the live path returns a coroutine."""
    import asyncio
    import inspect

    async def _call():
        res = handler(payload)
        if inspect.isawaitable(res):
            res = await res
        return res

    return asyncio.run(_call())


# Serving deps mirror the worker stack minus the trainer bits.
SERVE_DEPS = [
    "torch==2.10.0",
    "vllm==0.19.1",
    "transformers>=5.6,<5.11",
    "huggingface_hub>=0.25",
    "peft>=0.19",
    "accelerate>=1.4",
]
SERVE_SYSTEM_DEPS = ["build-essential"]

_ENDPOINT_CACHE: dict[str, object] = {}

# Serving cold start (image pull + vLLM/PEFT install + ~8 GB base model + adapter) can
# exceed 10 min on a fresh host; default the serve execution cap to 25 min
# (env-overridable) so the first `slm chat` on a cold worker doesn't fail with
# "executionTimeout exceeded".
_DEFAULT_SERVE_TIMEOUT_MS = 25 * 60 * 1000

MODES = ("dev", "always-on")
DEFAULT_IDLE_TIMEOUT_S = 300

# Projected always-on cost uses live RunPod rates (static fallback):
# providers/runpod/pricing.py (hourly_rate).


def serve_execution_timeout_ms() -> int:
    return _DEFAULT_SERVE_TIMEOUT_MS


def resolve_serve_deps() -> list[str]:
    explicit = os.environ.get("AUTOSLM_SERVE_DEPS")
    if explicit:
        # JSON list (use this for specs containing commas, e.g.
        # "transformers>=5.6,<5.11") or a whitespace-separated string. NOT comma-split:
        # a comma is part of a PEP 440 range and must not become two pip arguments
        # (mirrors providers/runpod/train.resolve_worker_deps).
        if explicit.strip().startswith("["):
            import json as _json

            deps = [str(d).strip() for d in _json.loads(explicit) if str(d).strip()]
        else:
            import shlex

            deps = [d for d in shlex.split(explicit) if d.strip()]
        if deps:
            return deps
    return SERVE_DEPS


@dataclass
class Deployment:
    run_id: str
    model: str
    adapter_hf_prefix: str
    gpu: str
    openai_model: str
    endpoint_name: str
    mode: str = "dev"
    idle_timeout_s: int = DEFAULT_IDLE_TIMEOUT_S
    est_idle_cost_usd_per_day: float = 0.0
    state: str = "ready"

    def to_dict(self) -> dict:
        return asdict(self)


def _language_model_only(model: str) -> bool:
    """Natively-multimodal checkpoints are served text-only. Approximated here by
    family name (the client can't load the HF config); the worker does the precise
    config-based check (engine.worker.is_vl_checkpoint, which covers Qwen3.5/3.6).
    Both families must match here or a Qwen3.6 checkpoint served via model_policy="allow"
    loses the text-only guard and re-hits the vision-tower VRAM/flash-attn issues."""
    return "Qwen3.5" in model or "Qwen3.6" in model


def serve_endpoint_name(friendly_gpu: str, run_id: str) -> str:
    tail = (run_id or "").split("-")[-1][:24]
    base = f"autoslm-serve-{gpu_short(canonical_gpu(friendly_gpu))}"
    return f"{base}-{tail}" if tail else base


def servable_gpu(gpu_name: str, model: str) -> str:
    """Serving runs on RunPod Flash only: a run trained on a class that is not
    RunPod-validated (a Vast-only class like L40S/RTX Pro 4000, OR a class that has a
    RunPod enum member but was validated only on Vast, e.g. RTX 3090) is served from the
    cheapest RunPod-VALIDATED class with at least the trained class's VRAM — NOT directly
    on the unvalidated RunPod substrate (which can fail on first chat) and NOT the
    catalog default (32 GB for open models, too small for the >32 GB class the allocator
    proved was needed)."""
    from autoslm.providers.base import GPU_INFO, UnsupportedGpuError, cheapest_gpu

    friendly = canonical_gpu(gpu_name)
    info = GPU_INFO[friendly]
    # Enum presence is not enough: serve directly only on a RunPod-VALIDATED class.
    if "runpod" in info.validated_on:
        return friendly
    try:
        # Prefer a RunPod-validated class big enough; only if none exists fall back
        # to an unvalidated one (better than refusing to serve at all).
        fallback = cheapest_gpu(info.vram_gb)
    except UnsupportedGpuError:
        fallback = cheapest_gpu(info.vram_gb, include_unvalidated=True)
    logger.warning(
        "%s is not RunPod-validated; serving %s on %s (>= %d GB)",
        friendly,
        model,
        fallback,
        info.vram_gb,
    )
    return fallback


# Self-contained adapter-merge script run on the serving worker (python -c _MERGE_SCRIPT
# <model> <adapter_dir> <out_dir>). It must NOT import autoslm: the serve worker only has the
# SERVE_DEPS stack (torch/transformers/peft), the autoslm package isn't installed there, and
# torch can't be imported in the Flash handler process (dynamo assertion) so this runs as a
# subprocess. ONE merge for BOTH adapter kinds — load the base in bf16 and merge_and_unload:
# exact for bf16 LoRA; the standard merge-to-bf16 recipe for 4-bit QLoRA (a LoRA can't be folded
# into a 4-bit base in place, so bf16 is the correct merge target either way).
_MERGE_SCRIPT = """
import sys
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id, adapter_dir, out_dir = sys.argv[1], sys.argv[2], sys.argv[3]
base = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, trust_remote_code=True)
merged = PeftModel.from_pretrained(base, adapter_dir).merge_and_unload()
merged.save_pretrained(out_dir, safe_serialization=True)
AutoTokenizer.from_pretrained(model_id, trust_remote_code=True).save_pretrained(out_dir)
print("[merge] %s + %s -> %s" % (model_id, adapter_dir, out_dir))
"""


def _serve_body(input_data: dict) -> dict:
    """Runs ON the GPU worker: forward a chat request to a persistent local vLLM server.

    vLLM runs as a LONG-LIVED SUBPROCESS (the OpenAI api_server on localhost), not in
    the handler process: importing torch inside the Flash handler process crashes
    (torch dynamo config-module assertion against the runpod runtime's module state -
    observed live), and a subprocess keeps the engine warm across requests anyway.
    The handler boots it on first request (the cold start) and proxies afterwards.

    NOTE: Flash serializes this handler and runs it standalone - all imports live
    inside the body, and state is cached in module globals while the worker is warm.
    """
    import json as _json
    import os
    import subprocess
    import sys
    import time
    import urllib.error
    import urllib.request

    g = globals()
    base = "http://127.0.0.1:8199"

    def _tail_serve_log(limit: int = 3000) -> str:
        # Defined IN the body: Flash serializes _serve_body and runs it standalone, so
        # the module-level helper of the same name is out of scope on the worker — a
        # bare reference would NameError on the boot-failure path and hide the vLLM log.
        try:
            with open("/tmp/vllm_serve.log") as f:
                return f.read()[-limit:]
        except OSError:
            return "(no serve log)"

    _thinking = input_data.get("thinking", False)
    # Robust bool: input_data may arrive loosely typed (serialized handler payload), and
    # bool("false") is True — treat the usual falsey strings as False.
    thinking = (
        _thinking.strip().lower() not in ("", "0", "false", "no", "off", "none")
        if isinstance(_thinking, str)
        else bool(_thinking)
    )
    # QLoRA tiers are served as a merged full model (no --enable-lora); bf16 tiers serve
    # base + LoRA. Derived from the catalog quant by the caller.
    merge = bool(input_data.get("merge_adapter"))
    # thinking is part of the engine key: it changes --max-model-len, and the vLLM
    # subprocess must be rebooted to change that.
    key = (input_data["model"], input_data["adapter_prefix"], thinking, merge)

    def server_alive() -> bool:
        proc = g.get("_AUTOSLM_PROC")
        if proc is None or proc.poll() is not None:
            return False
        try:
            urllib.request.urlopen(f"{base}/health", timeout=3)
            return True
        except Exception:
            return False

    if g.get("_AUTOSLM_KEY") != key or not server_alive():
        from huggingface_hub import snapshot_download

        prefix = input_data["adapter_prefix"]
        snapshot_download(
            repo_id=input_data["hf_repo"],
            repo_type="dataset",
            allow_patterns=[f"{prefix}/adapter/*"],
            local_dir="/adapter",
            token=input_data.get("token"),
        )
        adapter_dir = f"/adapter/{prefix}/adapter"
        old = g.get("_AUTOSLM_PROC")
        if old is not None and old.poll() is None:
            old.kill()
        if merge:
            # QLoRA: there is no quantized-base vLLM LoRA path, so merge the adapter into a full
            # bf16 model and serve that. Run the self-contained merge script in its own
            # interpreter (torch can't import in the Flash handler; autoslm isn't on the worker).
            menv = dict(os.environ)
            menv.setdefault("HF_TOKEN", input_data.get("token") or "")
            with open("/tmp/vllm_merge.log", "w") as mlog:
                merged = subprocess.run(
                    [sys.executable, "-c", _MERGE_SCRIPT, input_data["model"], adapter_dir, "/merged"],
                    stdout=mlog,
                    stderr=subprocess.STDOUT,
                    env=menv,
                )
            if merged.returncode != 0:
                try:
                    with open("/tmp/vllm_merge.log") as f:
                        mtail = f.read()[-3000:]
                except OSError:
                    mtail = "(no merge log)"
                raise RuntimeError(f"QLoRA adapter merge failed:\n{mtail}")
        # Merged model is served standalone (no LoRA); bf16 base is served with the LoRA attached.
        model_path = "/merged" if merge else input_data["model"]
        lora_args = (
            []
            if merge
            else [
                "--enable-lora",
                "--max-lora-rank",
                str(int(input_data.get("max_lora_rank") or 64)),
                "--lora-modules",
                f"adapter={adapter_dir}",
            ]
        )
        cmd = [
            sys.executable,
            "-m",
            "vllm.entrypoints.openai.api_server",
            "--host",
            "127.0.0.1",
            "--port",
            "8199",
            "--model",
            model_path,
            "--served-model-name",
            "base",
            "--dtype",
            "bfloat16",
            "--max-model-len",
            "4096" if thinking else "2048",  # <think> blocks need completion headroom
            "--gpu-memory-utilization",
            "0.85",
            *lora_args,
            "--trust-remote-code",
        ]
        if input_data.get("language_model_only"):
            cmd.append("--language-model-only")
        env = dict(os.environ)
        env.setdefault("HF_TOKEN", input_data.get("token") or "")
        # Popen dups the fd into the child, so the parent handle can close
        # immediately while vLLM keeps writing to the log.
        with open("/tmp/vllm_serve.log", "w") as log:
            g["_AUTOSLM_PROC"] = subprocess.Popen(
                cmd, stdout=log, stderr=subprocess.STDOUT, env=env
            )
        # Align the vLLM boot budget with the endpoint execution cap: a large model can take
        # several minutes to load on a cold host, and a hard-coded 900 s would 502 a first chat /
        # fail an always-on warmup while the RunPod request still has time budget. Leave ~60 s of
        # the window for the first generation.
        # Read the env DIRECTLY (with the same default as serve_execution_timeout_ms):
        # Flash serializes _serve_body and runs it standalone, so the module-level
        # helper is out of scope on the worker (see _tail_serve_log) and a bare call
        # would NameError before vLLM gets a chance to boot.
        serve_timeout_ms = 25 * 60 * 1000
        default_boot = max(900, serve_timeout_ms // 1000 - 60)
        deadline = time.time() + float(input_data.get("boot_timeout_s") or default_boot)
        while time.time() < deadline:
            if g["_AUTOSLM_PROC"].poll() is not None:
                tail = _tail_serve_log()
                raise RuntimeError(f"vLLM server exited during boot:\n{tail}")
            try:
                urllib.request.urlopen(f"{base}/health", timeout=3)
                break
            except Exception:
                time.sleep(2)
        else:
            tail = _tail_serve_log()
            raise RuntimeError(f"vLLM server did not become healthy in time:\n{tail}")
        g["_AUTOSLM_KEY"] = key

    # always-on warmup: pay the cold start (download + vLLM boot) at deploy time
    # so the user's first real chat is warm, then return without a completion.
    if input_data.get("warmup"):
        return {"ok": True, "warmed": True}

    body = {
        # merged QLoRA model is served under its base name; bf16 tiers serve the LoRA ("adapter").
        "model": "base" if merge else "adapter",
        "messages": input_data.get("messages") or [],
        "temperature": float(input_data.get("temperature", 0.0)),
        "top_p": float(input_data.get("top_p", 1.0)),
        "max_tokens": int(input_data.get("max_tokens", 512)),
        # Serve with the run's training-time thinking flag (decoding parity). Thinking
        # responses carry the raw <think>...</think> block in message.content.
        "chat_template_kwargs": {"enable_thinking": thinking},
    }

    def post(payload: dict) -> dict:
        req = urllib.request.Request(
            f"{base}/v1/chat/completions",
            data=_json.dumps(payload).encode(),
            headers={"Content-Type": "application/json"},
        )
        with urllib.request.urlopen(req, timeout=600) as resp:
            return _json.loads(resp.read())

    try:
        out = post(body)
    except urllib.error.HTTPError as e:
        if e.code == 400:
            # A vLLM build that rejects the kwarg falls back to the chat template's own
            # default (thinking ON for hybrid Qwen3 — correct for thinking runs, a logged
            # degradation for non-thinking ones).
            body.pop("chat_template_kwargs", None)
            print("vLLM rejected chat_template_kwargs; retrying without enable_thinking")
            out = post(body)
        else:
            raise
    out["model"] = input_data.get("served_model", "autoslm-adapter")
    return out


def _get_serve_endpoint(
    friendly_gpu: str,
    run_id: str,
    mode: str = "dev",
    idle_timeout_s: int = DEFAULT_IDLE_TIMEOUT_S,
):
    os.environ["FLASH_IS_LIVE_PROVISIONING"] = "true"
    from runpod_flash import Endpoint

    from autoslm.providers.runpod.auth import ensure_auth
    from autoslm.providers.runpod.train import FLASH_SDK_LOCK, isolate_flash_state, min_cuda_for

    ensure_auth()
    friendly = canonical_gpu(friendly_gpu)
    name = serve_endpoint_name(friendly, run_id)
    cache_key = f"{name}:{mode}:{idle_timeout_s}"

    # Serialize against training deploy/teardown on the same process: isolate_flash_state()
    # swaps runpod_flash's process-wide registry globals and Endpoint() touches the SDK's
    # asyncio singleton, so a concurrent terminate_endpoint()/always-on warmup on another
    # thread could race the registry scope. Hold the same lock across isolation + construction.
    with FLASH_SDK_LOCK:
        isolate_flash_state(f"serve-{run_id.split('-')[-1]}")
        if cache_key in _ENDPOINT_CACHE:
            return _ENDPOINT_CACHE[cache_key]
        kwargs = {
            "name": name,
            "gpu": flash_gpu(friendly),
            "gpu_count": 1,
            "min_cuda_version": min_cuda_for(friendly),
            # dev: scale to zero after idle_timeout (cold start accepted, $0 idle).
            # always-on: one permanently warm worker (no cold start, 24/7 billing).
            "workers": (0, 1) if mode == "dev" else (1, 1),
            "idle_timeout": int(idle_timeout_s),
            "flashboot": True,
            "execution_timeout_ms": serve_execution_timeout_ms(),
        }
        image = os.environ.get("AUTOSLM_WORKER_IMAGE")
        if image:
            kwargs["image"] = image
        else:
            kwargs["dependencies"] = resolve_serve_deps()
            kwargs["system_dependencies"] = SERVE_SYSTEM_DEPS
        ep = Endpoint(**kwargs)
        handler = ep(_serve_body)
        _ENDPOINT_CACHE[cache_key] = handler
        return handler


def _needs_merge(model: str) -> bool:
    """Whether the trained adapter must be MERGED into a full bf16 model before serving.

    QLoRA tiers (catalog ``quant == "4bit-qlora"``) trained the adapter against a 4-bit NF4
    base; there is no quantized-base vLLM LoRA serving path, so we merge the adapter into a
    full bf16 model on the serving worker and serve that directly (no ``--enable-lora``).
    bf16 tiers serve base + LoRA as usual."""
    from autoslm.catalog import MODELS

    info = MODELS.get(model)
    return info is not None and info.quant == "4bit-qlora"


def deploy_adapter(
    run_id: str,
    model: str,
    hf_repo: str,
    adapter_prefix: str,
    gpu_name: str = "RTX 5090",
    mode: str = "dev",
    idle_timeout_s: int = DEFAULT_IDLE_TIMEOUT_S,
    dry_run: bool = False,
    lora_rank: int = 64,
    thinking: bool = False,
) -> Deployment:
    """Provision a serving endpoint for a trained adapter (managed, no Docker)."""
    if mode not in MODES:
        raise ValueError(f"mode must be one of {MODES}, got {mode!r}")
    friendly = servable_gpu(gpu_name, model)
    from autoslm.runner import _gpu_rate

    rate = _gpu_rate(friendly)
    dep = Deployment(
        run_id=run_id,
        model=model,
        adapter_hf_prefix=adapter_prefix,
        gpu=friendly,
        openai_model=f"autoslm-{run_id}",
        endpoint_name=serve_endpoint_name(friendly, run_id),
        mode=mode,
        idle_timeout_s=idle_timeout_s,
        est_idle_cost_usd_per_day=0.0 if mode == "dev" else round(rate * 24, 2),
        state="dry_run" if dry_run else "ready",
    )
    if dry_run:
        return dep
    handler = _get_serve_endpoint(friendly, run_id, mode=mode, idle_timeout_s=idle_timeout_s)
    # always-on promises no cold start: warm the worker now (download + vLLM boot)
    # BEFORE returning ready, so the user's first chat is genuinely warm. dev mode
    # is scale-to-zero by design, so it warms lazily on first chat.
    if mode == "always-on":
        warmup = {
            "hf_repo": hf_repo,
            "model": model,
            "adapter_prefix": adapter_prefix,
            "token": os.environ.get("HF_TOKEN", ""),
            "max_lora_rank": max(64, int(lora_rank)),
            "language_model_only": _language_model_only(model),
            # QLoRA tiers: merge the adapter into a full bf16 model on the worker and serve that.
            "merge_adapter": _needs_merge(model),
            # warm the engine with the run's thinking flag so the first real chat (same
            # flag) reuses the warmed subprocess instead of rebooting for --max-model-len.
            "thinking": thinking,
            "warmup": True,
        }
        try:
            _invoke_handler(handler, warmup)
        except Exception:
            # The warmup invocation is what actually provisions the always-on worker;
            # if adapter download / vLLM boot fails the endpoint is registered (and may
            # bill) but no deployment is persisted. Tear it down before propagating.
            import contextlib

            with contextlib.suppress(Exception):
                undeploy_adapter(run_id, gpu_name=friendly)
            raise
    return dep


def undeploy_adapter(run_id: str, gpu_name: str = "RTX 5090") -> list[str]:
    """Tear down the run's serve endpoint via the REST API (works from any process)."""
    from autoslm.providers.runpod import api as runpod_api

    name = serve_endpoint_name(gpu_name, run_id)
    # find_endpoints_by_name is a substring filter, so guard with an exact-name
    # check (mirrors providers/runpod/train.py terminate_endpoint) — otherwise a
    # name that is a substring of another run's endpoint would over-delete and
    # mis-report the returned `deleted` list.
    deleted = [
        ep["name"]
        for ep in runpod_api.find_endpoints_by_name(name)
        if ep.get("name") == name and runpod_api.delete_endpoint(ep["id"])
    ]
    # Drop in-process handler cache entries for this endpoint (keyed name:mode:idle)
    # so a later redeploy constructs a fresh endpoint instead of reusing a handler
    # pointing at the just-deleted one.
    for key in [k for k in _ENDPOINT_CACHE if k.startswith(f"{name}:")]:
        _ENDPOINT_CACHE.pop(key, None)
    return deleted


def chat(
    run_id: str,
    messages: list[dict],
    model: str,
    hf_repo: str,
    adapter_prefix: str,
    gpu_name: str = "RTX 5090",
    temperature: float = 0.0,
    max_tokens: int = 512,
    mode: str = "dev",
    idle_timeout_s: int = DEFAULT_IDLE_TIMEOUT_S,
    lora_rank: int = 64,
    thinking: bool = False,
) -> dict:
    """Send an OpenAI-style chat request to the adapter's managed Flash GPU."""
    handler = _get_serve_endpoint(
        servable_gpu(gpu_name, model), run_id, mode=mode, idle_timeout_s=idle_timeout_s
    )
    # Natively-multimodal checkpoints are served text-only (no transformers needed
    # client-side; the family-name check mirrors the worker's config-based one).
    language_model_only = _language_model_only(model)
    payload = {
        "hf_repo": hf_repo,
        "model": model,
        "adapter_prefix": adapter_prefix,
        "token": os.environ.get("HF_TOKEN", ""),
        "served_model": f"autoslm-{run_id}",
        "messages": messages,
        "temperature": temperature,
        "max_tokens": max_tokens,
        # vLLM rejects an adapter whose rank exceeds --max-lora-rank; cover the
        # run's configured rank, not just the 64 default.
        "max_lora_rank": max(64, int(lora_rank)),
        "language_model_only": language_model_only,
        "merge_adapter": _needs_merge(model),
        "thinking": thinking,
    }
    return _invoke_handler(handler, payload)