unsloth/studio/backend/core/inference/llama_keepwarm.py

287 lines
11 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Opt-in idle auto-unload (TTL keep-warm) for the local llama.cpp model.
Off by default (idle seconds = 0). When enabled, a background loop unloads the
loaded GGUF once it has been idle for the configured TTL, freeing VRAM. A
pure-ASGI middleware tracks in-flight inference requests so a long stream that
outlives the TTL is never unloaded mid-response.
"""
from __future__ import annotations
import asyncio
import contextlib
import threading
import time
from loggers import get_logger
logger = get_logger(__name__)
_lock = threading.Lock()
_inflight = 0
# Requests blocked on the unload gate but not yet in _inflight: the idle loop must not unload
# while one is waiting (it would unload out from under it).
_pending = 0
_last_active = time.monotonic()
# The (id, quant) idle-unload last freed, so an alias/unknown request that would 503 against an
# empty backend can reload the exact freed variant (set on unload, cleared on reload).
_last_unloaded_model = None
# Guards inflight bumps against the idle-check-then-unload race and blocks new inference mid-swap.
# Process-wide, not per-loop: the backend slot is shared across every loop, so a per-loop gate
# would let a request on loop B start while a swap on loop A tears the model down.
_lifecycle_lock = threading.Lock()
@contextlib.asynccontextmanager
async def _unload_gate():
# Acquire off the loop: non-blocking first, else poll off a short sleep. Polling keeps the wait
# off the loop AND cancellation-safe (a cancel lands during the sleep, gate not held).
while not _lifecycle_lock.acquire(blocking = False):
await asyncio.sleep(0.02)
try:
yield
finally:
_lifecycle_lock.release()
_INFERENCE_PREFIXES = ("/v1/", "/api/inference/")
_INFERENCE_SUFFIXES = (
"/chat/completions",
"/completions",
"/messages",
"/messages/count_tokens", # uses the loaded tokenizer; protect like /messages
"/embeddings",
"/responses",
"/generate/stream", # Studio's own streaming route on the same llama-server
"/audio/generate", # direct GGUF TTS; can outlive the idle TTL
# Image generation holds a multi-GB pipeline for the whole request; tracking it lets
# other_inference_request_count() see an in-flight generation so an API-key training start is
# refused (409) before its unload cancels it. endswith avoids matching *-progress / */cancel.
"/images/generate", # /api/inference/images/generate
"/images/generations", # /v1/images/generations (+ /api/inference/images/generations)
# Video runs as a background job (the POST returns at once), so this only covers the brief
# accept; the training-start guards additionally probe generate-progress for an in-flight clip.
"/video/generate", # /api/inference/video/generate
)
def _is_inference_path(path: str) -> bool:
if path.startswith(_INFERENCE_PREFIXES) and path.endswith(_INFERENCE_SUFFIXES):
return True
# Public checkpoint preview (/p/{run}/v1/chat/completions) streams from the same backend.
return path.startswith("/p/") and path.endswith("/v1/chat/completions")
def _note_pending() -> None:
global _pending
with _lock:
_pending += 1
def _note_unpending() -> None:
global _pending
with _lock:
_pending = max(0, _pending - 1)
def _note_start() -> None:
# Do not stamp _last_active here: while _inflight > 0 the model is already protected, and
# stamping on start would let a later-untracked external request reset the idle timer.
global _inflight, _pending
with _lock:
_pending = max(0, _pending - 1)
_inflight += 1
def _note_end() -> None:
global _inflight, _last_active
with _lock:
_inflight = max(0, _inflight - 1)
_last_active = time.monotonic()
def _note_untracked_end() -> None:
# Drop a request that never used the local GGUF without stamping activity, so periodic
# external-provider traffic can't keep the model warm.
global _inflight
with _lock:
_inflight = max(0, _inflight - 1)
def _is_idle(ttl_seconds: float) -> bool:
with _lock:
return _inflight == 0 and _pending == 0 and (time.monotonic() - _last_active) >= ttl_seconds
def _note_activity() -> None:
"""Stamp activity, e.g. on a (re)load, so the model survives at least one TTL."""
global _last_active
with _lock:
_last_active = time.monotonic()
def other_inference_request_count(
current_request_counted: bool = True, *, include_pending: bool = True
) -> int:
"""Tracked inference requests other than the current route call.
The middleware counts before route code runs, so the caller is excluded by default.
Idle-unload counts pending waiters too. The swap guard passes include_pending=False: a pending
request is blocked in the middleware and hasn't started inference, so a swap can't interrupt it.
"""
with _lock:
active = _inflight
if current_request_counted and active > 0:
active -= 1
return max(0, active) + (_pending if include_pending else 0)
# Set on the ASGI scope by a route that proved this request won't touch llama.cpp, so the
# keep-warm count excludes it and the middleware skips its end-decrement.
_UNTRACKED_SCOPE_KEY = "_unsloth_keepwarm_untracked"
def untrack_current_request(scope) -> None:
"""Drop this request from the in-flight count once the route knows it won't use the local GGUF,
so external-provider traffic can't trip the swap busy guard. Idempotent."""
if not isinstance(scope, dict) or scope.get(_UNTRACKED_SCOPE_KEY):
return
scope[_UNTRACKED_SCOPE_KEY] = True
_note_untracked_end()
def inference_lifecycle_gate():
"""The gate a model swap holds so new inference can't start mid-load. Process-wide, so a swap
on one loop blocks inference on any other loop."""
return _unload_gate()
def note_model_loaded() -> None:
"""Record a successful GGUF load: stamp activity and drop any reload stash (a manual load
clears it synchronously, not only on the next idle poll)."""
_note_activity()
_set_last_unloaded(None)
def note_model_unloaded() -> None:
"""Record a deliberate (user/API) unload: drop the idle reload stash so the next request can't
resurrect the just-unloaded model. Unlike the idle loop it means "stay unloaded", so it must
not stamp activity."""
_set_last_unloaded(None)
def get_last_unloaded_model():
with _lock:
return _last_unloaded_model
def _set_last_unloaded(value) -> None:
global _last_unloaded_model
with _lock:
_last_unloaded_model = value
class LlamaKeepWarmMiddleware:
"""Pure ASGI: count in-flight inference requests and stamp activity on completion."""
def __init__(self, app):
self.app = app
async def __call__(self, scope, receive, send):
# Inference endpoints are all POST; skipping non-POST avoids counting CORS preflight.
if (
scope.get("type") != "http"
or scope.get("method") != "POST"
or not _is_inference_path(scope.get("path") or "")
):
await self.app(scope, receive, send)
return
# Always track in-flight, even when the feature is off, so a stream started before
# idle-unload is enabled can't be unloaded mid-response if the operator turns it on.
# Mark pending before the gate so the idle loop can't free the model while this waits.
_note_pending()
started = False
try:
async with _unload_gate():
_note_start()
started = True
finally:
if not started:
_note_unpending()
ended = {"done": False}
status = {"code": None}
def _finish() -> None:
# A route that untracked itself already decremented; don't double-count.
if ended["done"]:
return
ended["done"] = True
if scope.get(_UNTRACKED_SCOPE_KEY):
return
# This runs before FastAPI auth, so a 401/403 reaches here without touching llama.cpp.
# Decrement (to balance _note_start) but do NOT stamp activity, else repeated
# unauthenticated probes would keep the model warm forever.
if status["code"] in (401, 403):
_note_untracked_end()
else:
_note_end()
async def send_wrapper(message):
if message.get("type") == "http.response.start":
status["code"] = message.get("status")
# Final body frame marks the end of a (possibly streaming) response.
elif message.get("type") == "http.response.body" and not message.get(
"more_body", False
):
_finish()
await send(message)
try:
await self.app(scope, receive, send_wrapper)
finally:
_finish()
def _loaded_identity(backend):
if not backend.is_loaded or not backend.model_identifier:
return None
# Third slot is the advertised id (repo id) an auto-switch load sets; it's the override key, so
# an idle stash keyed by the concrete load path keeps the user's saved launch flags on reload.
advertised = getattr(backend, "_openai_advertised_id", None) or backend.model_identifier
return (backend.model_identifier, getattr(backend, "hf_variant", None), advertised)
async def idle_unload_loop(poll_seconds: float = 15.0) -> None:
"""Unload the loaded GGUF once idle past the configured TTL. Inert when off."""
from utils.openai_auto_switch_settings import get_auto_unload_idle_seconds
seen_model = None
while True:
await asyncio.sleep(poll_seconds)
try:
ttl = get_auto_unload_idle_seconds()
if ttl <= 0:
continue
from routes.inference import get_llama_cpp_backend
backend = get_llama_cpp_backend()
# Track by (id, variant): a (re)loaded model counts as activity so it survives one TTL
# before its first request (loads bypass the activity middleware).
current = _loaded_identity(backend)
if current != seen_model:
seen_model = current
if current is not None:
_note_activity()
_set_last_unloaded(None) # a model is loaded; drop stale stash
async with _unload_gate():
if backend.is_loaded and _is_idle(ttl):
freed = _loaded_identity(backend)
await asyncio.to_thread(backend.unload_model)
_set_last_unloaded(freed) # let an alias request reload it
logger.info("Idle auto-unload: freed GGUF after %ss idle", ttl)
seen_model = None
except Exception as exc:
logger.debug("idle_unload_loop iteration failed: %s", exc)