AMOOOMA commented on code in PR #37113: URL: https://github.com/apache/beam/pull/37113#discussion_r2730080557
########## sdks/python/apache_beam/ml/inference/model_manager.py: ########## @@ -0,0 +1,704 @@ +# +# Licensed to the Apache Software Foundation (ASF) under one or more +# contributor license agreements. See the NOTICE file distributed with +# this work for additional information regarding copyright ownership. +# The ASF licenses this file to You under the Apache License, Version 2.0 +# (the "License"); you may not use this file except in compliance with +# the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +"""Module for managing ML models in Apache Beam pipelines. + +This module provides classes and functions to efficiently manage multiple +machine learning models within Apache Beam pipelines. It includes functionality +for loading, caching, and updating models using multi-process shared memory, +ensuring that models are reused across different workers to optimize resource +usage and performance. +""" + +import uuid +import time +import threading +import subprocess +import logging +import gc +import numpy as np +from scipy.optimize import nnls +import torch +import heapq +import itertools +from collections import defaultdict, deque, Counter, OrderedDict +from typing import Dict, Any, Tuple, Optional, Callable + +logger = logging.getLogger(__name__) + + +class GPUMonitor: + """Monitors GPU memory usage in a separate thread using nvidia-smi. + + This class continuously polls GPU memory statistics to track current usage + and peak usage over a sliding time window. It serves as the source of truth + for the ModelManager's resource decisions. + + Attributes: + fallback_memory_mb: Default total memory if hardware detection fails. + poll_interval: Seconds between memory checks. + peak_window_seconds: Duration to track peak memory usage. + """ + def __init__( + self, + fallback_memory_mb: float = 16000.0, + poll_interval: float = 0.5, + peak_window_seconds: float = 30.0): + self._current_usage = 0.0 + self._peak_usage = 0.0 + self._total_memory = fallback_memory_mb + self._poll_interval = poll_interval + self._peak_window_seconds = peak_window_seconds + self._memory_history = deque() + self._running = False + self._thread = None + self._lock = threading.Lock() + + def _detect_hardware(self): + try: + cmd = [ + "nvidia-smi", + "--query-gpu=memory.total", + "--format=csv,noheader,nounits" + ] + output = subprocess.check_output(cmd, text=True).strip() + self._total_memory = float(output) + return True + except (FileNotFoundError, subprocess.CalledProcessError): + logger.warning( + "nvidia-smi not found or failed. Defaulting total memory to %s MB", + self._total_memory) + return False + except Exception as e: + logger.warning( + "Error parsing nvidia-smi output: %s. " + "Defaulting total memory to %s MB", + e, + self._total_memory) + return False + + def start(self): + self._gpu_available = self._detect_hardware() + if self._running or not self._gpu_available: + return + self._running = True + self._thread = threading.Thread(target=self._poll_loop, daemon=True) + self._thread.start() + + def stop(self): + self._running = False + if self._thread: + self._thread.join() + + def reset_peak(self): + with self._lock: + now = time.time() + self._memory_history.clear() + self._memory_history.append((now, self._current_usage)) + self._peak_usage = self._current_usage + + def get_stats(self) -> Tuple[float, float, float]: + with self._lock: + return self._current_usage, self._peak_usage, self._total_memory + + def refresh(self): + """Forces an immediate poll of the GPU.""" + usage = self._get_nvidia_smi_used() + now = time.time() + with self._lock: + self._current_usage = usage + self._memory_history.append((now, usage)) + # Recalculate peak immediately + while self._memory_history and (now - self._memory_history[0][0] + > self._peak_window_seconds): + self._memory_history.popleft() + self._peak_usage = ( + max(m for _, m in self._memory_history) + if self._memory_history else usage) + + def _get_nvidia_smi_used(self) -> float: + try: + cmd = "nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits" + output = subprocess.check_output(cmd, shell=True).decode("utf-8").strip() + free_memory = float(output) + return self._total_memory - free_memory + except Exception: + return 0.0 + + def _poll_loop(self): + while self._running: + usage = self._get_nvidia_smi_used() + now = time.time() + with self._lock: + self._current_usage = usage + self._memory_history.append((now, usage)) + while self._memory_history and (now - self._memory_history[0][0] + > self._peak_window_seconds): + self._memory_history.popleft() + self._peak_usage = ( + max(m for _, m in self._memory_history) + if self._memory_history else usage) + time.sleep(self._poll_interval) + + +class ResourceEstimator: + """Estimates individual model memory usage using statistical observation. + + Uses Non-Negative Least Squares (NNLS) to deduce the memory footprint of + individual models based on aggregate system memory readings and the + configuration of active models at that time. + """ + def __init__(self, smoothing_factor: float = 0.2, min_data_points: int = 5): + self.smoothing_factor = smoothing_factor + self.min_data_points = min_data_points + self.estimates: Dict[str, float] = {} + self.history = defaultdict(lambda: deque(maxlen=20)) + self.known_models = set() + self._lock = threading.Lock() + + def is_unknown(self, model_tag: str) -> bool: + with self._lock: + return model_tag not in self.estimates + + def get_estimate(self, model_tag: str, default_mb: float = 4000.0) -> float: + with self._lock: + return self.estimates.get(model_tag, default_mb) + + def set_initial_estimate(self, model_tag: str, cost: float): + with self._lock: + self.estimates[model_tag] = cost + self.known_models.add(model_tag) + logger.info("Initial Profile for %s: %s MB", model_tag, cost) + + def add_observation( + self, active_snapshot: Dict[str, int], peak_memory: float): + logger.info( + "Adding Observation: Snapshot=%s, PeakMemory=%.1f MB", + active_snapshot, + peak_memory) + if not active_snapshot: + return + with self._lock: + config_key = tuple(sorted(active_snapshot.items())) + self.history[config_key].append(peak_memory) + for tag in active_snapshot: + self.known_models.add(tag) + self._solve() + + def _solve(self): + """ + Solves Ax=b using raw readings (no pre-averaging) and NNLS. + This creates a 'tall' matrix A where every memory reading is + a separate equation. + """ + unique = sorted(list(self.known_models)) + + # We need to build the matrix first to know if we have enough data points + A, b = [], [] + + for config_key, mem_values in self.history.items(): + if not mem_values: + continue + + # 1. Create the feature row for this configuration ONCE + # (It represents the model counts + bias) + counts = dict(config_key) + feature_row = [counts.get(model, 0) for model in unique] + feature_row.append(1) # Bias column + + # 2. Add a separate row to the matrix for EVERY individual reading + # Instead of averaging, we flatten the history into the matrix + for reading in mem_values: + A.append(feature_row) # The inputs (models) stay the same + b.append(reading) # The output (memory) varies due to noise + + # Convert to numpy for SciPy + A = np.array(A) + b = np.array(b) + + if len( + self.history.keys()) < len(unique) + 1 or len(A) < self.min_data_points: + # Not enough data to solve yet + return + + logger.info( + "Solving with %s total observations for %s models.", + len(A), + len(unique)) + + try: + # Solve using Non-Negative Least Squares + # x will be >= 0 + x, _ = nnls(A, b) + + weights = x[:-1] + bias = x[-1] + + for i, model in enumerate(unique): + calculated_cost = weights[i] + + if model in self.estimates: + old = self.estimates[model] + new = (old * (1 - self.smoothing_factor)) + ( + calculated_cost * self.smoothing_factor) + self.estimates[model] = new + else: + self.estimates[model] = calculated_cost + + logger.info( + "Updated Estimate for %s: %.1f MB", model, self.estimates[model]) + logger.info("System Bias: %s MB", bias) + + except Exception as e: + logger.error("Solver failed: %s", e) + + +class TrackedModelProxy: + """A transparent proxy for model objects that adds tracking metadata. + + Wraps the underlying model object to attach a unique ID and intercept + calls, allowing the manager to track individual instances across processes. + """ + def __init__(self, obj): + object.__setattr__(self, "_wrapped_obj", obj) + object.__setattr__(self, "_beam_tracking_id", str(uuid.uuid4())) + + def __getattr__(self, name): + return getattr(self._wrapped_obj, name) + + def __setattr__(self, name, value): + setattr(self._wrapped_obj, name, value) + + def __call__(self, *args, **kwargs): + return self._wrapped_obj(*args, **kwargs) + + def __setstate__(self, state): + self.__dict__.update(state) + + def __getstate__(self): + return self.__dict__ + + def __str__(self): + return str(self._wrapped_obj) + + def __repr__(self): + return repr(self._wrapped_obj) + + def __dir__(self): + return dir(self._wrapped_obj) + + def trackedModelProxy_unsafe_hard_delete(self): + if hasattr(self._wrapped_obj, "singletonProxy_unsafe_hard_delete"): + try: + self._wrapped_obj.singletonProxy_unsafe_hard_delete() + except Exception: + pass + + +class ModelManager: + """Manages model lifecycles, caching, and resource arbitration. + + This class acts as the central controller for acquiring model instances. + It handles: + 1. LRU Caching of idle models. + 2. Resource estimation and admission control (preventing OOM). + 3. Dynamic eviction of low-priority models when space is needed. + 4. 'Isolation Mode' for safely profiling unknown models. + """ + _lock = threading.Lock() + + def __init__( + self, + monitor: Optional['GPUMonitor'] = None, + slack_percentage: float = 0.10, + poll_interval: float = 0.5, + peak_window_seconds: float = 30.0, + min_data_points: int = 5, + smoothing_factor: float = 0.2, + eviction_cooldown_seconds: float = 10.0, + min_model_copies: int = 1): + + self._estimator = ResourceEstimator( + min_data_points=min_data_points, smoothing_factor=smoothing_factor) + self._monitor = monitor if monitor else GPUMonitor( + poll_interval=poll_interval, peak_window_seconds=peak_window_seconds) + self._slack_percentage = slack_percentage + + self._eviction_cooldown = eviction_cooldown_seconds + self._min_model_copies = min_model_copies + + # Resource State + self._models = defaultdict(list) + self._idle_lru = OrderedDict() + self._active_counts = Counter() + self._total_active_jobs = 0 + self._pending_reservations = 0.0 + + self._isolation_mode = False + self._pending_isolation_count = 0 + self._isolation_baseline = 0.0 + + self._wait_queue = [] + self._ticket_counter = itertools.count() + self._cv = threading.Condition() + self._load_lock = threading.Lock() + + self._monitor.start() + + def all_models(self, tag) -> list[Any]: + return self._models[tag] + + def acquire_model(self, tag: str, loader_func: Callable[[], Any]) -> Any: + current_priority = 0 if self._estimator.is_unknown(tag) else 1 + ticket_num = next(self._ticket_counter) + my_id = object() + + with self._cv: + # FAST PATH + if self._pending_isolation_count == 0 and not self._isolation_mode: Review Comment: Good catch! This is some old code from older iterations. Removed now. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
