eqcctpro 0.6.2__py3-none-any.whl → 0.7.0__py3-none-any.whl

eqcctpro/functionality.py

@@ -0,0 +1,1424 @@
+"""
+functionality.py controls all the functionality of EQCCTPro, specifically how we access mseed_predictor() and parallel_predict. 
+It is a level of abstraction so we can make the code more concise and cleaner
+"""
+import os
+import gc
+import ray
+import sys
+import ast
+import math
+import queue 
+import psutil
+import random
+import numbers
+import logging
+import resource
+import threading
+from .tools import *
+from pathlib import Path
+from .parallelization import *
+from obspy import UTCDateTime
+from ray.util.queue import Queue
+from datetime import datetime, timedelta
+from .tools import _parse_gpus_field
+from logging.handlers import QueueHandler, QueueListener
+
+
+class RunEQCCTPro():  
+    """RunEQCCTPro class for running the RunEQCCTPro functions for multiple instances of the class"""
+    def __init__(self, # self is 'this instance' of the class 
+                use_gpu: bool, 
+                input_dir: str, 
+                output_dir: str, 
+                log_filepath: str, 
+                p_model_filepath: str = None, 
+                s_model_filepath: str = None, 
+                number_of_concurrent_station_predictions: int = None,
+                number_of_concurrent_timechunk_predictions: int = 1, 
+                intra_threads: int = 1, 
+                inter_threads: int = 1, 
+                P_threshold: float = 0.001, 
+                S_threshold: float = 0.02,
+                specific_stations: str = None,
+                csv_dir: str = None,
+                best_usecase_config: bool = False,
+                vram_mb: float = None,
+                selected_gpus: list = None,
+                cpu_id_list: list = [1],
+                start_time:str = None, 
+                end_time:str = None, 
+                timechunk_dt:int = 1,
+                waveform_overlap:int = 0,
+                tmp_dir:str = None,
+                # SeisBench model parameters
+                model_type: str = 'eqcct',  # 'eqcct' or 'seisbench'
+                seisbench_parent_model: str = None,  # e.g., 'PhaseNet', 'EQTransformer'
+                seisbench_child_model: str = None,  # e.g., 'original', 'stead'
+                Detection_threshold: float = 0.3):  # Detection threshold for SeisBench models 
+         
+        self.use_gpu = use_gpu  # 'this instance' of the classes object, use_gpu = use_gpu 
+        self.input_dir = input_dir
+        self.output_dir = output_dir
+        self.log_filepath = log_filepath
+        self.p_model_filepath = p_model_filepath
+        self.s_model_filepath = s_model_filepath
+        self.number_of_concurrent_station_predictions = number_of_concurrent_station_predictions
+        self.number_of_concurrent_timechunk_predictions = number_of_concurrent_timechunk_predictions
+        self.intra_threads = intra_threads
+        self.inter_threads = inter_threads
+        self.P_threshold = P_threshold
+        self.S_threshold = S_threshold
+        self.specific_stations = specific_stations
+        self.csv_dir = csv_dir
+        self.best_usecase_config = best_usecase_config
+        self.vram_mb = vram_mb
+        self.selected_gpus = selected_gpus if selected_gpus is not None else list_gpu_ids() # a list of the GPU IDs. If not provided, we use all available GPUs
+        self.cpu_id_list = cpu_id_list 
+        self.cpu_count = len(cpu_id_list)
+        self.start_time = start_time
+        self.end_time = end_time
+        self.timechunk_dt = timechunk_dt
+        self.waveform_overlap = waveform_overlap
+        self.home_tmp_dir = tmp_dir  
+        
+        # SeisBench model parameters
+        self.model_type = model_type.lower()
+        self.seisbench_parent_model = seisbench_parent_model
+        self.seisbench_child_model = seisbench_child_model
+        self.Detection_threshold = Detection_threshold
+
+        # Validate model type and parameters
+        if self.model_type not in ['eqcct', 'seisbench']:
+            raise ValueError(f"model_type must be 'eqcct' or 'seisbench', got '{model_type}'")
+        
+        if self.model_type == 'eqcct':
+            if p_model_filepath is None or s_model_filepath is None:
+                raise ValueError("For EQCCT model_type, p_model_filepath and s_model_filepath are required")
+            if number_of_concurrent_station_predictions is None:
+                raise ValueError("number_of_concurrent_station_predictions is required for EQCCT")
+        elif self.model_type == 'seisbench':
+            if seisbench_parent_model is None or seisbench_child_model is None:
+                raise ValueError("For SeisBench model_type, seisbench_parent_model and seisbench_child_model are required")
+            if number_of_concurrent_station_predictions is None:
+                raise ValueError("number_of_concurrent_station_predictions is required for SeisBench")
+
+        # Ensures that the output_dir exists. If it doesn't, we create it 
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        # Set up main logger and logger queue to retrive queued logs from Raylets to be passed to the main logger
+        self.logger = logging.getLogger("eqcctpro") # We named the logger eqcctpro (can be any name)
+        self.logger.setLevel(logging.INFO)
+        self.logger.propagate = False # if true, events logged to this logger will be passed to the handlers of higher level (ancestor) loggers, in addition to any handlers attached to this logger
+        if not self.logger.handlers: # avoid duplicating inits 
+            fmt = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+            file_h = logging.FileHandler(self.log_filepath) # Writes logs to file 
+            stream_h = logging.StreamHandler() # Sends logs to console
+            file_h.setFormatter(fmt)
+            stream_h.setFormatter(fmt)
+            self.logger.addHandler(file_h)
+            self.logger.addHandler(stream_h)
+
+        self.logger.info("")
+        self.logger.info(f"------- Welcome to EQCCTPro -------")
+        self.logger.info("")
+
+        # If the user passed a GPU but no valid VRAM, need to exit 
+        if self.use_gpu and not (isinstance(self.vram_mb, numbers.Real) and math.isfinite(self.vram_mb) and self.vram_mb > 0): 
+            self.logger.error(f"No numerical VRAM passed. Please provide vram_mb (MB per Raylet per GPU) as a positive real number. Exiting...")
+            sys.exit(1)
+
+        # We need to ensure that the vram specified does not exceed the capabilities of the system, if not, we need to exit safely before it happens
+        if self.use_gpu:
+            # Determine model VRAM requirement based on model type
+            if self.model_type == 'seisbench':
+                from .parallelization import get_seisbench_model_vram_mb
+                model_vram_mb = get_seisbench_model_vram_mb(
+                    self.seisbench_parent_model, 
+                    self.seisbench_child_model,
+                    default_mb=2000.0  # Default VRAM for SeisBench models
+                )
+                self.logger.info(f"Using VRAM requirement: {model_vram_mb:.0f} MB for SeisBench model {self.seisbench_parent_model}/{self.seisbench_child_model}")
+            else:
+                model_vram_mb = 1500.0  # Safety reserve for EQCCT
+            
+            check_vram_per_gpu_style(
+                selected_gpus=self.selected_gpus,
+                get_gpu_vram_fn=lambda gid: get_gpu_vram(gpu_index=gid),
+                intended_workers=self.number_of_concurrent_station_predictions * self.number_of_concurrent_timechunk_predictions,
+                vram_mb=self.vram_mb,
+                model_vram_mb=model_vram_mb,
+                safety_cap=0.95,
+                eqcct_overhead_gb=0.0,
+                logger=self.logger)
+    
+    # To-Do: merge dt_task_generator and chunk_time into one function and concatenate the objects so we dont have so much stuff running around
+    # Generates the dt tasks list 
+    def dt_task_generator(self): 
+        # Modifies the times_list values (see chunk_time()) so it can be in a format the mseed_predictor can use 
+        tasks = [[f"({i+1}/{len(self.times_list)})", f"{self.times_list[i][0].strftime(format='%Y%m%dT%H%M%SZ')}_{self.times_list[i][1].strftime(format='%Y%m%dT%H%M%SZ')}"] for i in range((len(self.times_list)))]
+        self.tasks_picker = tasks
+    
+    def chunk_time(self):
+        # Creates the timechunks, EI. from X specific time to Y specific time to generate the dt tasks (timechunk tasks that are run in parallel first at the top level)
+        # EX. [[UTCDateTime(2024, 12, 15, 11, 58), UTCDateTime(2024, 12, 15, 13, 0)], [UTCDateTime(2024, 12, 15, 12, 58), UTCDateTime(2024, 12, 15, 14, 0)]]
+        starttime = UTCDateTime(self.start_time) - (self.waveform_overlap * 60)
+        endtime = UTCDateTime(self.end_time)
+
+        times_list = []
+        start = starttime
+        end = start + (self.waveform_overlap * 60) + (self.timechunk_dt * 60)
+        while start <= endtime:
+            if end >= endtime:
+                end = endtime
+                times_list.append([start, end])
+                break
+            times_list.append([start, end])
+            start = end - (self.waveform_overlap * 60)
+            end = start + (self.waveform_overlap * 60) + (self.timechunk_dt * 60)
+
+        self.times_list = times_list
+    
+    def _drain_worker_logs(self):
+            while True:
+                rec = self.log_queue.get()  # blocks until a record arrives
+                if rec is None: break       # sentinel to stop thread
+                try:
+                    self.logger.handle(rec) # routes to file+console handlers
+                except Exception:
+                    # never crash on logging
+                    self.logger.exception("Failed to handle worker log record")
+
+    def configure_cpu(self): 
+        # We need to configure the tf_environ for the CPU configuration that is being inputted
+        self.logger.info(f"Running EQCCT over Requested MSeed Files using CPU(s)...")
+        skip_tf = (self.model_type != 'eqcct')
+        if self.best_usecase_config:
+            # We use the best usecase configuration that was found using EvaluateSystem
+            result = find_optimal_configuration_cpu(best_overall_usecase=True, eval_sys_results_dir=self.csv_dir)
+            if result is None: 
+                self.logger.info("")
+                self.logger.info(f"Error: Could not retrieve an optimal CPU configuration. Please check that the CSV file exists and try again. Exiting...")
+                exit()  # Exit gracefully
+            cpus_to_use, num_concurrent_predictions, intra, inter, station_count = result
+            self.logger.info("")
+            self.logger.info(f"Using {cpus_to_use} CPUs, {num_concurrent_predictions} Conc. Predictions, {intra} Intra Threads, and {inter} Inter Threads...")
+            tf_environ(gpu_id=-1, intra_threads=intra, inter_threads=inter, logger=self.logger, skip_tf=skip_tf)
+        else:
+            # We pass the requested parameters to the tf_environ 
+            tf_environ(gpu_id=-1, intra_threads=self.intra_threads, inter_threads=self.inter_threads, logger=self.logger, skip_tf=skip_tf) 
+            
+    def configure_gpu(self):
+        # We need to configure the tf_environ for the GPU configuration that is being inputted
+        self.logger.info(f"Running EQCCT over Requested MSeed Files using GPU(s)...")
+        # In the main process (driver), we only set environment variables. 
+        # We ALWAYS skip TensorFlow initialization here because the main process doesn't run models.
+        # This avoids confusing "No GPUs visible" messages if the driver's environment differs from workers.
+        skip_tf = True 
+        if self.best_usecase_config: 
+            result = find_optimal_configuration_gpu(True, self.csv_dir)
+            if result is None:
+                self.logger.info("")
+                self.logger.error(f"Error: Could not retrieve an optimal GPU configuration. Please check that the CSV file exists and try again. Exiting...")
+                exit()  # Exit gracefully
+
+            self.logger.info("")
+            cpus_to_use, num_concurrent_predictions, intra, inter, gpus, vram_mb, station_count = result # Unpack values only if result is valid
+            self.logger.info(f"Using {cpus_to_use} CPUs, {num_concurrent_predictions} Conc. Predictions, {intra} Intra Threads, {inter} Inter Threads, {gpus} GPU IDs, and {vram_mb} MB VRAM per Task...")
+            tf_environ(gpu_id=1, vram_limit_mb=vram_mb, gpus_to_use=gpus, intra_threads=intra, inter_threads=inter, logger=self.logger, skip_tf=skip_tf)
+        
+        else: 
+            self.logger.info("")
+            self.logger.info(f"User requested to use GPU(s): {self.selected_gpus} with {self.vram_mb} MB of VRAM per Raylet (intra-op threads = {self.intra_threads}, inter-op threads = {self.inter_threads})") # Use the selected GPUs 
+            tf_environ(gpu_id=1, vram_limit_mb=self.vram_mb, gpus_to_use=self.selected_gpus, intra_threads=self.intra_threads, inter_threads=self.inter_threads, logger=self.logger, skip_tf=skip_tf)
+    
+    def eqcctpro_parallelization(self):
+        if self.specific_stations is None: # We check if the station dirs are consistent, if not, exit
+            statement, specific_stations_list, do_i_exit = check_station_dirs(input_dir=self.input_dir)
+            self.logger.info(f"{statement}")
+            if do_i_exit: exit()
+
+        # We want to use a specified amount of stations
+        else: specific_stations_list = [station.strip() for station in self.specific_stations.split(',')]
+        statement = f"Using {len(specific_stations_list)} selected station(s)."
+        self.logger.info(f"{statement}")
+        self.logger.info("")           
+
+        # Submit timechunk tasks to mseed_predictor
+        tasks_queue = []
+        log_queue = queue.Queue()  # Create a queue for log entries
+        
+        # Compute total analyis timeframe 
+        total_analysis_time = datetime.strptime(self.end_time, "%Y-%m-%d %H:%M:%S") - datetime.strptime(self.start_time, "%Y-%m-%d %H:%M:%S")
+        
+        max_pending_tasks = self.number_of_concurrent_timechunk_predictions 
+        self.logger.info(f"------- Starting EQCCTPro... -------")
+        self.logger.info(f"Detailed subprocess information can be found in the log file.")
+        self.logger.info("")
+        for i in range(len(self.tasks_picker)):
+            mseed_timechunk_dir_name = self.tasks_picker[i][1]
+            timechunk_dir_path = os.path.join(self.input_dir, mseed_timechunk_dir_name) 
+        
+            # Concurrent Timechunks 
+            while True: 
+                if len(tasks_queue) < max_pending_tasks: 
+                    tasks_queue.append(mseed_predictor.options(num_gpus=0, num_cpus=1).remote(input_dir=timechunk_dir_path, output_dir=self.output_dir, log_queue=self.log_queue, 
+                                        P_threshold=self.P_threshold, S_threshold=self.S_threshold, p_model=self.p_model_filepath, s_model=self.s_model_filepath, 
+                                        number_of_concurrent_station_predictions=self.number_of_concurrent_station_predictions, ray_cpus=self.cpu_id_list, use_gpu=self.use_gpu, 
+                                        gpu_id=self.selected_gpus, gpu_memory_limit_mb=self.vram_mb, specific_stations=specific_stations_list, 
+                                        timechunk_id=mseed_timechunk_dir_name, waveform_overlap=self.waveform_overlap, total_timechunks=len(self.tasks_picker), 
+                                        number_of_concurrent_timechunk_predictions=self.number_of_concurrent_timechunk_predictions, total_analysis_time=total_analysis_time,
+                                        intra_threads=self.intra_threads, inter_threads=self.inter_threads,
+                                        model_type=self.model_type, seisbench_parent_model=self.seisbench_parent_model, 
+                                        seisbench_child_model=self.seisbench_child_model, Detection_threshold=self.Detection_threshold))
+                    break
+                
+                else: # If there are more tasks than maximum, just process them
+                    tasks_finished, tasks_queue = ray.wait(tasks_queue, num_returns=1, timeout=None)
+                    for finished_task in tasks_finished:
+                        log_entry = ray.get(finished_task)
+                        log_queue.put(log_entry)  # Add log entry to the queue
+
+        # After adding all the tasks to queue, process what's left
+        while tasks_queue:
+            tasks_finished, tasks_queue = ray.wait(tasks_queue, num_returns=1, timeout=None)
+            for finished_task in tasks_finished:
+                log_entry = ray.get(finished_task)
+                self.logger.info(log_entry)
+
+        # stop log forwarder
+        self.log_queue.put(None) # remember, log_queue is a Ray Queue actor, and will only exist while Ray is still active (cannot be after the .shutdown())
+        self._log_thread.join(timeout=2)
+
+        ray.shutdown()
+        self.logger.info(f"Ray Successfully Shutdown.")
+        self.logger.info("------- Successfully Picked All Waveform(s) from all Timechunk(s) -------")
+        # self.logger.info("------- END OF FILE -------")
+        
+    def run_eqcctpro(self):
+        # Set CPU affinity
+        process = psutil.Process(os.getpid())
+        process.cpu_affinity(self.cpu_id_list)  # Limit process to the given CPU IDs
+        
+        self.chunk_time() # Generates the UTC times for each of the timesets in the given time range 
+        self.dt_task_generator() # Generates the task list so can know how many total tasks there are for our given time range 
+        
+        if self.use_gpu: # GPU
+            self.configure_gpu()
+            ray.init(ignore_reinit_error=True, num_gpus=len(self.selected_gpus), num_cpus=len(self.cpu_id_list), logging_level=logging.ERROR, log_to_driver=False, _temp_dir=self.home_tmp_dir) # Ray initalization using GPUs 
+            self.log_queue = Queue() # Create a Ray-safe queue to recieve LogRecord objects from workers so we can write them to file 
+            self._log_thread = threading.Thread(target=self._drain_worker_logs, daemon=True) # Creates background thread whose only job is to get() records from self.log_queue and hand them over to the actual logger
+            self._log_thread.start() # Starts the thread
+            # Log some import info to user 
+            statement = f"Ray Successfully Initialized with {self.selected_gpus} GPU(s) and {len(self.cpu_id_list)} CPU(s) ({list(self.cpu_id_list)} CPU Affinity Binding)."
+            self.logger.info(f"{statement}")
+            self.logger.info(f"Analyzing {len(self.times_list)} time chunk(s) from {self.start_time} to {self.end_time} (dt={self.timechunk_dt}min, overlap={self.waveform_overlap}min).")
+            
+            # Running parllelization
+            self.eqcctpro_parallelization()
+
+        else: # CPU
+            self.configure_cpu()
+            ray.init(ignore_reinit_error=True, num_cpus=len(self.cpu_id_list), logging_level=logging.ERROR, log_to_driver=False, _temp_dir=self.home_tmp_dir) # Ray initalization using CPUs
+            self.log_queue = Queue() # Create a Ray-safe queue to recieve LogRecord objects from workers so we can write them to file 
+            self._log_thread = threading.Thread(target=self._drain_worker_logs, daemon=True) # Creates background thread whose only job is to get() records from self.log_queue and hand them over to the actual logger
+            self._log_thread.start() # Starts the thread
+            # Log some import info to user
+            statement = f"Ray Successfully Initialized with {len(self.cpu_id_list)} CPU(s) ({list(self.cpu_id_list)} CPU Affinity Binding)."
+            self.logger.info(f"{statement}")
+            self.logger.info(f"Analyzing {len(self.times_list)} time chunk(s) from {self.start_time} to {self.end_time} (dt={self.timechunk_dt}min, overlap={self.waveform_overlap}min).")
+            
+            # Running parllelization
+            self.eqcctpro_parallelization()
+
+class EvaluateSystem(): 
+    """Evaluate System class for running the evaluation system functions for multiple instances of the class"""
+    def __init__(self,
+                 eval_mode: str,
+                 input_dir: str,
+                 output_dir: str,
+                 log_filepath: str,
+                 csv_dir: str, 
+                 p_model_filepath: str = None, 
+                 s_model_filepath: str = None, 
+                 P_threshold: float = 0.001, 
+                 S_threshold: float = 0.02, 
+                 intra_threads: int = 1,
+                 inter_threads: int = 1,
+                 stations2use:int = None, 
+                 cpu_id_list:list = [1],
+                 cpu_test_step_size:int = 1, 
+                 starting_amount_of_stations: int = 1, 
+                 station_list_step_size: int = 1, 
+                 min_cpu_amount: int = 1,
+                 min_conc_stations: int = 1, 
+                 conc_station_tasks_step_size: int = 1,
+                 max_vram_mb:float = None,
+                 gpu_vram_safety_cap:float = 0.90, 
+                 selected_gpus:list = None,
+                 start_time:str = None, 
+                 end_time:str = None, 
+                 conc_timechunk_tasks_step_size: int = 1, 
+                 timechunk_dt:int = 1,
+                 waveform_overlap:int = 0,
+                 tmp_dir:str = None,
+                 # SeisBench model parameters
+                 model_type: str = 'eqcct',  # 'eqcct' or 'seisbench'
+                 seisbench_parent_model: str = None,
+                 seisbench_child_model: str = None,
+                 Detection_threshold: float = 0.3): 
+        
+        valid_modes = {"cpu", "gpu"}
+        if eval_mode not in valid_modes: 
+            raise ValueError(f"Invalid mode '{eval_mode}'. Choose either 'cpu' or 'gpu'.")
+        
+        self.eval_mode = eval_mode.lower()
+        self.intra_threads = intra_threads
+        self.inter_threads = inter_threads
+        self.input_dir = input_dir  
+        self.output_dir = output_dir
+        self.log_filepath = log_filepath
+        self.csv_dir = csv_dir
+        self.P_threshold = P_threshold
+        self.S_threshold = S_threshold
+        self.p_model_filepath = p_model_filepath
+        self.s_model_filepath = s_model_filepath
+        self.stations2use = stations2use
+        self.cpu_id_list = cpu_id_list
+        self.vram_mb = max_vram_mb
+        self.gpu_vram_safety_cap = gpu_vram_safety_cap
+        self.selected_gpus = selected_gpus
+        self.use_gpu = True if self.eval_mode == 'gpu' else False
+        self.cpu_count = len(cpu_id_list)
+        self.cpu_test_step_size = cpu_test_step_size
+        self.starting_amount_of_stations = starting_amount_of_stations
+        self.station_list_step_size = station_list_step_size
+        self.min_cpu_amount = min_cpu_amount
+        self.min_conc_stations = min_conc_stations # default is = 1 
+        self.conc_station_tasks_step_size = conc_station_tasks_step_size # default is = 1 
+        self.stations2use_list = list(range(1, 11)) + list(range(15, 50, 5)) if stations2use is None else generate_station_list(self.starting_amount_of_stations, stations2use, self.station_list_step_size,)
+        self.start_time = start_time
+        self.end_time = end_time
+        self.conc_timechunk_tasks_step_size = conc_timechunk_tasks_step_size
+        self.timechunk_dt = timechunk_dt
+        self.waveform_overlap = waveform_overlap
+        self.home_tmp_dir = tmp_dir 
+
+        # SeisBench model parameters
+        self.model_type = model_type.lower()
+        self.seisbench_parent_model = seisbench_parent_model
+        self.seisbench_child_model = seisbench_child_model
+        self.Detection_threshold = Detection_threshold
+
+        # Validate model type and parameters
+        if self.model_type not in ['eqcct', 'seisbench']:
+            raise ValueError(f"model_type must be 'eqcct' or 'seisbench', got '{model_type}'")
+        
+        if self.model_type == 'eqcct':
+            if p_model_filepath is None or s_model_filepath is None:
+                raise ValueError("For EQCCT model_type, p_model_filepath and s_model_filepath are required")
+        elif self.model_type == 'seisbench':
+            if seisbench_parent_model is None or seisbench_child_model is None:
+                raise ValueError("For SeisBench model_type, seisbench_parent_model and seisbench_child_model are required")
+        
+        # Ensures that the output_dir exists. If it doesn't, we create it 
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        # Set up main logger and logger queue to retrive queued logs from Raylets to be passed to the main logger
+        self.logger = logging.getLogger("eqcctpro") # We named the logger eqcctpro (can be any name)
+        self.logger.setLevel(logging.INFO)
+        self.logger.propagate = False # if true, events logged to this logger will be passed to the handlers of higher level (ancestor) loggers, in addition to any handlers attached to this logger
+        if not self.logger.handlers: # avoid duplicating inits 
+            fmt = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+            file_h = logging.FileHandler(self.log_filepath) # Writes logs to file 
+            stream_h = logging.StreamHandler() # Sends logs to console
+            file_h.setFormatter(fmt)
+            stream_h.setFormatter(fmt)
+            self.logger.addHandler(file_h)
+            self.logger.addHandler(stream_h)
+        
+        self.logger.info("")
+        self.logger.info(f"------- Welcome to EQCCTPro's EvaluateSystem Functionality -------")
+        self.logger.info("")
+        # Set up temp dir 
+        import tempfile
+        tempfile.tempfile = self.home_tmp_dir
+
+        os.environ['TMPDIR'] = self.home_tmp_dir
+        os.environ['TEMP'] = self.home_tmp_dir
+        os.environ['TMP'] = self.home_tmp_dir
+        self.logger.info(f"Successfully set up temp files to be stored at {self.home_tmp_dir}")
+
+        # We need to ensure that the vram specified does not exceed the capabilities of t  he system, if not, we need to exit safely before it happens
+        self.chunk_time()
+        intended_workers = int(len(self.stations2use_list)) * int(len(self.times_list) // 2)
+        if self.eval_mode == 'gpu':
+            if not self.selected_gpus:
+                raise ValueError("selected_gpus must be set in GPU mode.")
+            self.chunk_time()
+            intended_workers = int(len(self.stations2use_list)) * int(len(self.times_list) // 2)
+
+            # Determine model VRAM requirement based on model type
+            if self.model_type == 'seisbench':
+                from .parallelization import get_seisbench_model_vram_mb
+                model_vram_mb = get_seisbench_model_vram_mb(
+                    self.seisbench_parent_model, 
+                    self.seisbench_child_model,
+                    default_mb=2000.0
+                )
+            else:
+                model_vram_mb = 3000.0  # Default for EQCCT
+
+            per_gpu_free_mb = [get_gpu_vram(gpu_index=g)[1] * 1024.0 for g in self.selected_gpus]  # free_gb -> MB
+            plan = evaluate_vram_capacity(
+                intended_workers=intended_workers,
+                vram_per_worker_mb=float(self.vram_mb),
+                per_gpu_free_mb=per_gpu_free_mb,
+                model_vram_mb=model_vram_mb,
+                safety_cap=self.gpu_vram_safety_cap,
+                eqcct_overhead_gb=1.1,
+            )
+            if not plan.ok_aggregate:
+                unit = plan.per_worker_mb + plan.overhead_mb
+                raise RuntimeError(
+                    f"Insufficient aggregate VRAM. Cap={plan.aggregate_cap_mb:.0f} MB, "
+                    f"Need={plan.aggregate_need_mb:.0f} MB (= {plan.model_vram_mb:.0f}×{len(self.selected_gpus)} + "
+                    f"{plan.intended_workers}×{unit:.0f})."
+                )
+            self.logger.info(
+                f"VRAM budget OK. Need {plan.aggregate_need_mb:.0f} MB ≤ Cap {plan.aggregate_cap_mb:.0f} MB "
+                f"across {len(self.selected_gpus)} GPU(s)."
+            )
+        
+    def _generate_stations_list(self):
+        """Generates station list"""
+        if self.station2use is None: 
+            return list(range(1, 11)) + list(range(15, 50, 5))
+        return generate_station_list(self.stations2use, self.starting_amount_of_stations, self.station_list_step_size)
+    
+    # def _prepare_environment(self):
+    #     """Removed 'output_dir' so that there is no conflicts in the save for a clean output return"""
+    #     remove_directory(self.output_dir)
+        
+    def chunk_time(self):
+        starttime = UTCDateTime(self.start_time) - (self.waveform_overlap * 60)
+        endtime = UTCDateTime(self.end_time)
+
+        times_list = []
+        start = starttime
+        end = start + (self.waveform_overlap * 60) + (self.timechunk_dt * 60)
+        while start <= endtime:
+            if end >= endtime:
+                end = endtime
+                times_list.append([start, end])
+                break
+            times_list.append([start, end])
+            start = end - (self.waveform_overlap * 60)
+            end = start + (self.waveform_overlap * 60) + (self.timechunk_dt * 60)
+
+        self.times_list = times_list
+
+    def _drain_worker_logs(self):
+            while True:
+                rec = self.log_queue.get()  # blocks until a record arrives
+                if rec is None: break       # sentinel to stop thread
+                try:
+                    self.logger.handle(rec) # routes to file+console handlers
+                except Exception:
+                    # never crash on logging
+                    self.logger.exception("Failed to handle worker log record")
+    
+    def dt_task_generator(self): 
+        tasks = [[f"({i+1}/{len(self.times_list)})", f"{self.times_list[i][0].strftime(format='%Y%m%dT%H%M%SZ')}_{self.times_list[i][1].strftime(format='%Y%m%dT%H%M%SZ')}"] for i in range((len(self.times_list)))]
+        self.tasks_picker = tasks
+
+    def _trial_key(self, *, num_cpus: int, stations: int, predictions: int, gpu_memory_limit_mb, timechunks: int, model: str, gpus: list = None) -> str:
+        # Use provided gpus parameter if available, otherwise fall back to self.selected_gpus
+        if gpus is not None:
+            gpus_to_use = gpus
+        else:
+            gpus_to_use = self.selected_gpus if self.selected_gpus is not None else []
+        gpus_norm = tuple(sorted(int(x) for x in gpus_to_use))
+        vram_norm = int(round(float(gpu_memory_limit_mb))) if gpu_memory_limit_mb not in ("", None) else ""
+        return f"cpus={int(num_cpus)}|gpus={gpus_norm}|stations={int(stations)}|pred={int(predictions)}|timechunks={int(timechunks)}|vram={vram_norm}|model={model}"
+        
+    def _load_existing_trial_keys(self, csv_path: str) -> set[str]:
+        if not os.path.exists(csv_path):
+            return set()
+        try:
+            df = pd.read_csv(csv_path, keep_default_na=False)
+        except Exception:
+            return set()
+
+        keys = set()
+        for _, row in df.iterrows():
+            try:
+                num_cpus = int(float(row.get("Number of CPUs Allocated for Ray to Use", 0) or 0))
+                stations = int(float(row.get("Number of Stations Used", 0) or 0))
+                predictions = int(float(row.get("Number of Concurrent Station Tasks", 0) or 0))
+                timechunks = int(float(row.get("Concurrent Timechunks Used", 0) or 0))
+                vram = row.get("Inference Actor Memory Limit (MB)", "")
+                vram_mb = float(vram) if vram not in ("", None) else ""
+                # Parse GPUs from this specific row (don't overwrite self.selected_gpus)
+                row_gpus = _parse_gpus_field(row.get("GPUs Used")) or []
+                # Extract model info
+                model = row.get("Model Used", "eqcct") # Default to eqcct for legacy rows
+                # Pass the row's info directly to _trial_key
+                keys.add(self._trial_key(num_cpus=num_cpus, stations=stations, predictions=predictions, 
+                                       gpu_memory_limit_mb=vram_mb, timechunks=timechunks, model=model, gpus=row_gpus))
+            except Exception:
+                continue
+        return keys
+
+    def evaluate_cpu(self): 
+        """Evaluate system parallelization using CPUs"""
+        statement = "Evaluating System Parallelization Capability using CPU"
+        self.logger.info(f"{statement}")
+        
+        os.makedirs(self.csv_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+        
+        # Create test results csv 
+        csv_filepath = f"{self.csv_dir}/cpu_test_results.csv"
+        prepare_csv(csv_file_path=csv_filepath, logger=self.logger)
+        planned_keys = self._load_existing_trial_keys(csv_filepath)
+        
+        self.chunk_time()
+        self.dt_task_generator()
+        
+        trial_num = 1
+        log_queue = queue.Queue()  # Create a queue for log entries
+        total_analysis_time = datetime.strptime(self.end_time, "%Y-%m-%d %H:%M:%S") - datetime.strptime(self.start_time, "%Y-%m-%d %H:%M:%S")
+
+        if self.min_cpu_amount > len(self.cpu_id_list): 
+            # Code won't execute because the minimum CPU amount of > the len(cpu id list)
+            # In which the rest of the code is dependent on the len for generating cpu_count 
+            print(f"CPU ID List provided has less CPUs than the minimum requested ({len(self.cpu_id_list)} vs. {self.min_cpu_amount}). Exiting...")
+            quit()
+        
+        with open(self.log_filepath, mode="a+", buffering=1) as log: 
+            for i in range(self.min_cpu_amount, self.cpu_count+1, self.cpu_test_step_size):
+                # Set CPU affinity and initialize Ray
+                cpus_to_use = self.cpu_id_list[:i]
+                process = psutil.Process(os.getpid())
+                process.cpu_affinity(cpus_to_use)  # Limit process to the given CPU IDs
+                
+                ray.init(ignore_reinit_error=True, num_cpus=len(cpus_to_use), logging_level=logging.FATAL, log_to_driver=False, _temp_dir=self.home_tmp_dir)
+                self.log_queue = Queue() # Create a Ray-safe queue to recieve LogRecord objects from workers so we can write them to file 
+                self._log_thread = threading.Thread(target=self._drain_worker_logs, daemon=True) # Creates background thread whose only job is to get() records from self.log_queue and hand them over to the actual logger
+                self._log_thread.start() # Starts the thread
+                self.logger.info(f"Ray Successfully Initialized with {len(cpus_to_use)} CPU(s) ({list(cpus_to_use)} CPU Affinity Binding).")
+                
+                timechunks_list = []
+                timechunk = 1
+                step = self.conc_timechunk_tasks_step_size # Use the class attribute
+                while timechunk <= len(self.tasks_picker):
+                    timechunks_list.append(timechunk)
+                    if timechunk == 1:
+                        timechunk += 1
+                    else:
+                        timechunk += step
+
+                if len(self.tasks_picker) not in timechunks_list:
+                    timechunks_list.append(len(self.tasks_picker))
+                # sets are a set of multiple items stored in a single variable 
+                # unchangable after being set, cannot have duplicates and is unordered
+                timechunks_list = sorted(list(set(timechunks_list))) 
+                # Determine model name for trial key
+                if self.model_type == 'seisbench':
+                    trial_model = f"{self.seisbench_parent_model}/{self.seisbench_child_model}"
+                else:
+                    trial_model = "eqcct"
+
+                for timechunks in timechunks_list:
+                    tested_concurrency = set() # Reset for each cpu / timechunk configuration
+                    for num_stations in self.stations2use_list: 
+                        # Use a 20% step size for concurrency testing as requested
+                        # This tests 20%, 40%, 60%, 80%, and 100% of the current total stations
+                        step = max(1, int(num_stations * 0.2))
+                        concurrent_predictions_list = sorted(list(set(range(step, num_stations + 1, step))))
+                        
+                        # Efficiency optimization: only test concurrency values we haven't seen yet 
+                        # for this CPU/Timechunk combo to save compute time. 
+                        new_concurrent_values = [x for x in concurrent_predictions_list if x not in tested_concurrency]
+                        if not new_concurrent_values:
+                            continue  # All concurrency values already tested
+                        for num_concurrent_predictions in new_concurrent_values:
+                            tested_concurrency.add(num_concurrent_predictions)
+                            key = self._trial_key(
+                                num_cpus=len(cpus_to_use),
+                                stations=num_stations,
+                                predictions=num_concurrent_predictions,
+                                gpu_memory_limit_mb="",   # CPU eval has no per-task VRAM cap
+                                timechunks=timechunks,
+                                model=trial_model
+                            )
+                            if key in planned_keys:
+                                self.logger.info(f"[SKIP] Already tested: {key}")
+                                continue
+                            planned_keys.add(key)
+                                    
+                            mseed_timechunk_dir_name = self.tasks_picker[timechunks-1][1]
+                            timechunk_dir_path = os.path.join(self.input_dir, mseed_timechunk_dir_name) 
+                            max_pending_tasks = timechunks
+                            
+                            self.logger.info("")
+                            self.logger.info(f"------- Trial Number: {trial_num} -------")
+                            self.logger.info(f"CPU(s): {len(cpus_to_use)}")
+                            self.logger.info(f"Conc. Timechunks Being Analyzed: {timechunks} / Total Timechunks to be Analyzed: {len(self.tasks_picker)}")
+                            self.logger.info(f"Total Amount of Stations to be Processed in Current Trial: {num_stations} / Number of Stations Being Processed Concurrently: {num_concurrent_predictions} / Total Overall Trial Station Count: {max(self.stations2use_list)}") 
+                            
+                            # Concurrent Timechunks
+                            tasks_queue = []
+                            log_queue = queue.Queue()  # Create a queue for log entries
+
+
+                            # ===== RAM Baseline (before launching worker) =====
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_before_total_mb = _rss / 1e6
+
+                            # peak before (platform-aware)
+                            if resource is not None:  # Linux/macOS
+                                _ru = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+                                if sys.platform.startswith("linux"):
+                                    peak_before_mb = _ru / 1024.0            # ru_maxrss in KB on Linux
+                                elif sys.platform == "darwin":
+                                    peak_before_mb = _ru / (1024.0 * 1024.0) # ru_maxrss in bytes on macOS
+                                else:
+                                    peak_before_mb = mem_before_total_mb     # safe fallback
+                            else:  # Windows: no 'resource'
+                                try:
+                                    peak_before_mb = process.memory_full_info().peak_wset / 1e6
+                                except Exception:
+                                    peak_before_mb = mem_before_total_mb
+
+                            try: 
+                                while True: 
+                                    if len(tasks_queue) < max_pending_tasks: 
+                                        tasks_queue.append(mseed_predictor.options(num_gpus=0, num_cpus=1).remote(input_dir=timechunk_dir_path, output_dir=self.output_dir, log_queue=self.log_queue, 
+                                                            P_threshold=self.P_threshold, S_threshold=self.S_threshold, p_model=self.p_model_filepath, s_model=self.s_model_filepath, 
+                                                            number_of_concurrent_station_predictions=num_concurrent_predictions, ray_cpus=cpus_to_use, use_gpu=self.use_gpu, 
+                                                            gpu_id=self.selected_gpus, gpu_memory_limit_mb=self.vram_mb, stations2use=num_stations, 
+                                                            timechunk_id=mseed_timechunk_dir_name, waveform_overlap=self.waveform_overlap, total_timechunks=len(self.tasks_picker), 
+                                                            number_of_concurrent_timechunk_predictions=max_pending_tasks, total_analysis_time=total_analysis_time, testing_gpu=False, 
+                                                            test_csv_filepath=csv_filepath, intra_threads=self.intra_threads, inter_threads=self.inter_threads, timechunk_dt=self.timechunk_dt,
+                                                            model_type=self.model_type, seisbench_parent_model=self.seisbench_parent_model, 
+                                                            seisbench_child_model=self.seisbench_child_model, Detection_threshold=self.Detection_threshold))
+                                    
+                                        break
+                                
+                                    else: 
+                                        tasks_finished, tasks_queue = ray.wait(tasks_queue, num_returns=1, timeout=None)
+                                        for finished_task in tasks_finished:
+                                            log_entry = ray.get(finished_task)
+                                            log_queue.put(log_entry)  # Add log entry to the queue
+                                
+                                # After adding all the tasks to queue, process what's left
+                                while tasks_queue:
+                                    tasks_finished, tasks_queue = ray.wait(tasks_queue, num_returns=1, timeout=None)
+                                    for finished_task in tasks_finished:
+                                        log_entry = ray.get(finished_task)
+                                        log_queue.put(log_entry)  # Add log entry to the queue
+
+                                    update_csv(csv_filepath, success=1, error_message="")
+                            except Exception as e:
+                                # Failure occured, need to add to log 
+                                error_msg = f"{type(e).__name__}: {str(e)}"
+                                update_csv(csv_filepath, success=0, error_message=error_msg)
+                                self.logger.error(f"Trial {trial_num} FAILED: {error_msg}")
+                                
+                            # Write log entries from the queue to the file
+                            while not log_queue.empty():
+                                log_entry = log_queue.get()
+                                
+                            remove_output_subdirs(self.output_dir, logger=self.logger)
+                            trial_num += 1  
+                            
+                            # RAM cleanup
+                            # ===== AFTER RUN (before cleanup) =====
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_after_run_total_mb = _rss / 1e6
+                            delta_run_mb = mem_after_run_total_mb - mem_before_total_mb
+
+                            # updated peak (platform-aware)
+                            if resource is not None:
+                                _ru = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+                                if sys.platform.startswith("linux"):
+                                    peak_after_mb = _ru / 1024.0
+                                elif sys.platform == "darwin":
+                                    peak_after_mb = _ru / (1024.0 * 1024.0)
+                                else:
+                                    peak_after_mb = mem_after_run_total_mb
+                            else:
+                                try:
+                                    peak_after_mb = process.memory_full_info().peak_wset / 1e6
+                                except Exception:
+                                    peak_after_mb = mem_after_run_total_mb
+
+                            self.logger.info("")
+                            self.logger.info(
+                                f"[MEM] Baseline: {mem_before_total_mb:.2f} MB | After run: {mem_after_run_total_mb:.2f} MB "
+                                f"| Δrun: {delta_run_mb:.2f} MB | Peak≈{max(peak_before_mb, peak_after_mb):.2f} MB"
+                            )
+
+                            # ===== CLEANUP =====
+                            # drop strong refs so GC matters
+                            try: del ref
+                            except NameError: pass
+                            try: del log_entry
+                            except NameError: pass
+
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_before_clean_mb = _rss / 1e6
+
+                            gc.collect()
+                            time.sleep(0.1)
+
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_after_clean_mb = _rss / 1e6
+
+                            freed_mb = mem_before_clean_mb - mem_after_clean_mb
+                            self.logger.info(f"[MEM] Freed ~{max(freed_mb, 0):.2f} MB; Post-clean total: {mem_after_clean_mb:.2f} MB") # To-Do: Fix the Freed so its beeter (for cpu and gpu)
+                            self.logger.info("")
+                            
+                        # tested_concurrency.update([x for x in concurrent_predictions_list if x <= num_stations])
+                        tested_concurrency.update(new_concurrent_values)
+
+                    # stop log forwarder
+                    self.log_queue.put(None) # remember, log_queue is a Ray Queue actor, and will only exist while Ray is still active (cannot be after the .shutdown())
+                    self._log_thread.join(timeout=2)
+
+                    ray.shutdown() # Shutdown Ray after processing all timechunks for this CPU count 
+                    self.logger.info(f"Ray Successfully Shutdown.")
+                                
+     
+        self.logger.info(f"Testing complete.")
+        self.logger.info(f"")
+        self.logger.info(f"Finding Optimal Configurations...")
+        # Compute optimal configurations (CPU)
+        df = pd.read_csv(csv_filepath)
+        optimal_configuration_df, best_overall_usecase_df = find_optimal_configurations_cpu(df)
+        optimal_configuration_df.to_csv(f"{self.csv_dir}/optimal_configurations_cpu.csv", index=False)
+        best_overall_usecase_df.to_csv(f"{self.csv_dir}/best_overall_usecase_cpu.csv", index=False)
+        self.logger.info(f"Optimal Configurations Found. Findings saved to:") 
+        self.logger.info(f" 1) Optimal CPU/Station/Concurrent Prediction Configurations: {self.csv_dir}/optimal_configurations_cpu.csv") 
+        self.logger.info(f" 2) Best Overall Usecase Configuration: {self.csv_dir}/best_overall_usecase_cpu.csv")
+
+    def evaluate_gpu(self): 
+        """Evaluate system parallelization using GPUs"""
+        statement = "Evaluating System Parallelization Capability using GPUs"
+        self.logger.info(f"{statement}")
+        
+        os.makedirs(self.csv_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        # Create test results csv 
+        csv_filepath = f"{self.csv_dir}/gpu_test_results.csv"
+        prepare_csv(csv_file_path=csv_filepath, logger=self.logger)
+        
+        # Normalize existing CSV to ensure consistent "GPUs Used" formatting and quoting
+        if os.path.exists(csv_filepath):
+            from .tools import normalize_gpu_csv_quoting
+            normalize_gpu_csv_quoting(csv_filepath)
+            self.logger.info("Normalized existing CSV entries for consistent 'GPUs Used' formatting.")
+        
+        planned_keys = self._load_existing_trial_keys(csv_filepath)
+        
+        # Log summary of existing trials
+        if planned_keys:
+            self.logger.info(f"Loaded {len(planned_keys)} existing trial(s) from CSV. These will be skipped.")
+            # Count trials by GPU configuration
+            gpu_counts = {}
+            for key in planned_keys:
+                # Extract GPU info from key (format: cpus=X|gpus=(...)|...)
+                if "gpus=" in key:
+                    gpu_part = key.split("gpus=")[1].split("|")[0]
+                    gpu_counts[gpu_part] = gpu_counts.get(gpu_part, 0) + 1
+            if gpu_counts:
+                self.logger.info("Existing trials by GPU configuration:")
+                for gpu_config, count in sorted(gpu_counts.items()):
+                    self.logger.info(f"  {gpu_config}: {count} trial(s)")
+        else:
+            self.logger.info("No existing trials found in CSV. Starting fresh evaluation.")
+
+        # Calculate these at the start
+        self.chunk_time()
+        self.dt_task_generator()
+
+        trial_num = 1
+        log_queue = queue.Queue()  # Create a queue for log entries
+        total_analysis_time = datetime.strptime(self.end_time, "%Y-%m-%d %H:%M:%S") - datetime.strptime(self.start_time, "%Y-%m-%d %H:%M:%S")
+        
+        # Track statistics
+        trials_skipped = 0
+        trials_run = 0
+        
+        if self.min_cpu_amount > len(self.cpu_id_list): 
+            # Code won't execute because the minimum CPU amount of > the len(cpu id list)
+            # In which the rest of the code is dependent on the len for generating cpu_count 
+            print(f"CPU ID List provided has less CPUs than the minimum requested ({len(self.cpu_id_list)} vs. {self.min_cpu_amount}). Exiting...")
+            quit()
+
+        for gpu in range(len(self.selected_gpus)):
+            for cpu in range(self.min_cpu_amount, len(self.cpu_id_list)+1, self.cpu_test_step_size):
+                # Set CPU affinity and initialize Ray
+                cpus_to_use = self.cpu_id_list[:cpu] # 'cpu' is a count (e.g., 1, 2, 3). Slicing [:cpu] gets that many CPU IDs, we want :cpu bc we are using 0 index counting so 0-20 exclusive = 0-19 IDs = 20 CPUs to use explicitely. ([:n is exclusive])
+                gpus_to_use = self.selected_gpus[:gpu+1] # 'gpu' is an index (0, 1, 2). We need +1 to include the current GPU.
+                # Set CPU affinity
+                process = psutil.Process(os.getpid())
+                process.cpu_affinity(cpus_to_use)  # Limit process to the given CPU IDs
+                
+                # VRAM budget per GPU (MB). If vram_mb is provided, treat it as an explicit per-GPU budget override.
+                free_vram_mb = float(self.vram_mb) if self.vram_mb else self.calculate_vram()
+                self.logger.info("")
+                self.logger.info("=" * 80)
+                self.logger.info(f"Testing Using {len(gpus_to_use)} GPU(s) with IDs {gpus_to_use} and {len(cpus_to_use)} CPU(s)")
+                self.logger.info("=" * 80)
+                
+                os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+                os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(map(str, gpus_to_use))
+                # Initialize Ray with GPUs
+                ray.init(ignore_reinit_error=True, num_gpus=len(gpus_to_use), num_cpus=len(cpus_to_use), 
+                        logging_level=logging.FATAL, log_to_driver=False, _temp_dir=self.home_tmp_dir)
+                self.log_queue = Queue() # Create a Ray-safe queue to recieve LogRecord objects from workers so we can write them to file 
+                self._log_thread = threading.Thread(target=self._drain_worker_logs, daemon=True) # Creates background thread whose only job is to get() records from self.log_queue and hand them over to the actual logger
+                self._log_thread.start() # Starts the thread
+                self.logger.info(f"Ray Successfully Initialized with {len(gpus_to_use)} GPU(s) and {len(cpus_to_use)} CPU(s) ({list(cpus_to_use)} CPU Affinity Binding).")
+                self.logger.info(f"Trials will evalute GPU(s) performance against Iterative Total Station Tasks ({self.stations2use_list}) with Varying Concurrent Predictions.")
+                self.logger.info("")
+                
+                # Efficiency optimization: track tested configurations for this GPU/CPU combo
+                tested_gpu_configs = set() 
+
+                # Determine model name for trial key
+                if self.model_type == 'seisbench':
+                    trial_model = f"{self.seisbench_parent_model}/{self.seisbench_child_model}"
+                else:
+                    trial_model = "eqcct"
+
+                for stations in self.stations2use_list:
+                    # Use a 20% step size for concurrency testing as requested
+                    step = max(1, int(stations * 0.2))
+                    concurrent_predictions_list = sorted(list(set(range(step, stations + 1, step))))
+                    
+                    self.logger.info(f"Evaluating GPU(s) against {stations} TOTAL STATION(s) with 20% STEP CONCURRENT STATION PREDICTIONS: {concurrent_predictions_list}")
+                    for predictions in concurrent_predictions_list:
+                        vram_per_task_mb = free_vram_mb / predictions
+                        step_size = vram_per_task_mb * 0.2
+                        vram_steps = np.arange(step_size, vram_per_task_mb + step_size, step_size)
+                        
+                        # Determine minimum VRAM filter based on model type
+                        if self.model_type == 'seisbench':
+                            from .parallelization import get_seisbench_model_vram_mb
+                            min_vram = get_seisbench_model_vram_mb(
+                                self.seisbench_parent_model, 
+                                self.seisbench_child_model,
+                                default_mb=2000.0
+                            )
+                        else:
+                            min_vram = 3000.0  # EQCCT minimum
+                            
+                        vram_steps = vram_steps[vram_steps >= min_vram] 
+                        
+                        # We are defining the hard upper limit on how much VRAM (GPU memory) TensorFlow is allowed to use inside each ModelActor process
+                        # This includes 1) Loading the model weights into GPU memory during init and 2) Usage of the ModelActor (predictions, etc.) 
+                        # If the actor tries to allocate more memory than available, then a OOME will occur inside that actor only
+                        # Good for testing OOME prevention while not letting actors steal available memory from other actors 
+                        # LSS: It is purely the VRAM ceiling for each shared inference actor that handles the actual model predictions
+                        for gpu_memory_limit_mb in vram_steps: 
+                            gpu_memory_limit_mb = int(round(float(gpu_memory_limit_mb)))
+
+                            # Efficiency check: avoid redundant tests for same (concurrency, vram)
+                            config_key = (predictions, gpu_memory_limit_mb)
+                            if config_key in tested_gpu_configs:
+                                continue
+                            tested_gpu_configs.add(config_key)
+
+                            key = self._trial_key(
+                                num_cpus=len(cpus_to_use),
+                                stations=stations,
+                                predictions=predictions,
+                                gpu_memory_limit_mb=gpu_memory_limit_mb,
+                                timechunks=1,  # your GPU eval is explicitly "one timechunk at a time"
+                                model=trial_model,
+                                gpus=gpus_to_use,  # Pass the actual GPUs being used in this iteration
+                            )
+                            if key in planned_keys:
+                                self.logger.info(f"[SKIP] Already tested: {key}")
+                                trials_skipped += 1
+                                continue
+                            planned_keys.add(key)
+                            trials_run += 1
+                            self.logger.info("")
+                            self.logger.info(f"------- Trial Number: {trial_num} -------")
+                            # self.logger.info(f"VRAM Limited to {gpu_memory_limit_mb:.2f} MB per Parallel Task")
+                            
+                            # Get the first timechunk for testing
+                            mseed_timechunk_dir_name = self.tasks_picker[0][1]
+                            timechunk_dir_path = os.path.join(self.input_dir, mseed_timechunk_dir_name)
+                            
+                            self.logger.info(f"Stations: {stations}")
+                            self.logger.info(f"Concurrent Station Predictions: {predictions}")
+                            self.logger.info(f"VRAM per Parallel Task: {gpu_memory_limit_mb:.2f} MB")
+                            self.logger.info("")
+
+
+                            # ===== Baseline RAM consumption (before launching worker) =====
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_before_total_mb = _rss / 1e6
+
+                            # peak before (platform-aware)
+                            if resource is not None:  # Linux/macOS
+                                _ru = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+                                if sys.platform.startswith("linux"):
+                                    peak_before_mb = _ru / 1024.0            # ru_maxrss in KB on Linux
+                                elif sys.platform == "darwin":
+                                    peak_before_mb = _ru / (1024.0 * 1024.0) # ru_maxrss in bytes on macOS
+                                else:
+                                    peak_before_mb = mem_before_total_mb     # safe fallback
+                            else:  # Windows: no 'resource'
+                                try:
+                                    peak_before_mb = process.memory_full_info().peak_wset / 1e6
+                                except Exception:
+                                    peak_before_mb = mem_before_total_mb
+                            
+                            try: # To Do: Add Concurrent Timechunks Testing for GPU/CPU too, reference eqcctpro_parallelization()
+                                # Call mseed_predictor directly via Ray (just like evaluate_cpu does)
+                                ref = mseed_predictor.options(num_gpus=0, num_cpus=1).remote(
+                                    input_dir=timechunk_dir_path, 
+                                    output_dir=self.output_dir, 
+                                    log_queue=self.log_queue, 
+                                    P_threshold=self.P_threshold, 
+                                    S_threshold=self.S_threshold, 
+                                    p_model=self.p_model_filepath, 
+                                    s_model=self.s_model_filepath, 
+                                    number_of_concurrent_station_predictions=predictions, 
+                                    ray_cpus=cpus_to_use, 
+                                    use_gpu=self.use_gpu, 
+                                    gpu_id=gpus_to_use, 
+                                    gpu_memory_limit_mb=gpu_memory_limit_mb, 
+                                    stations2use=stations, 
+                                    timechunk_id=mseed_timechunk_dir_name, 
+                                    waveform_overlap=self.waveform_overlap, 
+                                    total_timechunks=len(self.tasks_picker), 
+                                    number_of_concurrent_timechunk_predictions=1,  # Testing one timechunk at a time
+                                    total_analysis_time=total_analysis_time, 
+                                    testing_gpu=True,  # Enable test mode
+                                    test_csv_filepath=csv_filepath, 
+                                    intra_threads=self.intra_threads, 
+                                    inter_threads=self.inter_threads, 
+                                    timechunk_dt=self.timechunk_dt,
+                                    model_type=self.model_type, seisbench_parent_model=self.seisbench_parent_model, 
+                                    seisbench_child_model=self.seisbench_child_model, Detection_threshold=self.Detection_threshold
+                                )
+                                
+                                # Wait for result
+                                log_entry = ray.get(ref)
+                                log_queue.put(log_entry)  # Add log entry to the queue
+                                
+                                # Success - update CSV
+                                update_csv(csv_filepath, success=1, error_message="")
+                                
+                            except Exception as e:
+                                # Failure occurred, need to add to log 
+                                error_msg = f"{type(e).__name__}: {str(e)}"
+                                update_csv(csv_filepath, success=0, error_message=error_msg)
+                                self.logger.info(f"Trial {trial_num} FAILED: {error_msg}")
+                            
+                            # Write log entries from the queue to the file
+                            while not log_queue.empty():
+                                log_entry = log_queue.get()
+                                self.logger.info(f"{log_entry}") # FIX ME 
+                            
+                            remove_output_subdirs(self.output_dir, logger=self.logger) 
+                            trial_num += 1
+                            
+                            # RAM cleanup
+                            # ===== AFTER RUN (before cleanup) =====
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_after_run_total_mb = _rss / 1e6
+                            delta_run_mb = mem_after_run_total_mb - mem_before_total_mb
+
+                            # updated peak (platform-aware)
+                            if resource is not None:
+                                _ru = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
+                                if sys.platform.startswith("linux"):
+                                    peak_after_mb = _ru / 1024.0
+                                elif sys.platform == "darwin":
+                                    peak_after_mb = _ru / (1024.0 * 1024.0)
+                                else:
+                                    peak_after_mb = mem_after_run_total_mb
+                            else:
+                                try:
+                                    peak_after_mb = process.memory_full_info().peak_wset / 1e6
+                                except Exception:
+                                    peak_after_mb = mem_after_run_total_mb
+
+                            self.logger.info(
+                                f"[MEM] Baseline: {mem_before_total_mb:.2f} MB | After run: {mem_after_run_total_mb:.2f} MB "
+                                f"| Δrun: {delta_run_mb:.2f} MB | Peak≈{max(peak_before_mb, peak_after_mb):.2f} MB"
+                            )
+
+                            # ===== CLEANUP =====
+                            # drop strong refs so GC matters
+                            try: del ref
+                            except NameError: pass
+                            try: del log_entry
+                            except NameError: pass
+
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_before_clean_mb = _rss / 1e6
+
+                            gc.collect()
+                            time.sleep(0.1)
+
+                            _rss = process.memory_info().rss
+                            for _ch in process.children(recursive=True):
+                                try:
+                                    _rss += _ch.memory_info().rss
+                                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                                    pass
+                            mem_after_clean_mb = _rss / 1e6
+
+                            freed_mb = mem_before_clean_mb - mem_after_clean_mb
+                            self.logger.info(f"[MEM] Freed ~{max(freed_mb, 0):.2f} MB; Post-clean total: {mem_after_clean_mb:.2f} MB\n")
+                            self.logger.info("")
+                
+                # stop log forwarder
+                self.log_queue.put(None) # remember, log_queue is a Ray Queue actor, and will only exist while Ray is still active (cannot be after the .shutdown())
+                self._log_thread.join(timeout=2)
+
+                ray.shutdown()  # Shutdown Ray after all testing
+                self.logger.info(f"Ray Successfully Shutdown.")
+
+        self.logger.info(f"Testing complete.")
+        self.logger.info(f"")
+        self.logger.info(f"Trial Summary: {trials_run} new trial(s) executed, {trials_skipped} trial(s) skipped (already in CSV)")
+        self.logger.info(f"Finding Optimal Configurations...")
+        self.logger.info(f"Recalculating optimal configurations from all trial data (including new trials)...")
+        # Compute optimal configurations (GPU)
+        df = pd.read_csv(csv_filepath)
+        optimal_configuration_df, best_overall_usecase_df = find_optimal_configurations_gpu(df)
+        # Overwrite existing files with recalculated optimal configurations
+        optimal_configuration_df.to_csv(f"{self.csv_dir}/optimal_configurations_gpu.csv", index=False)
+        best_overall_usecase_df.to_csv(f"{self.csv_dir}/best_overall_usecase_gpu.csv", index=False)
+        self.logger.info(f"Optimal Configurations Found. Findings saved to:") 
+        self.logger.info(f" 1) Optimal GPU/Station/Concurrent Prediction Configurations: {self.csv_dir}/optimal_configurations_gpu.csv") 
+        self.logger.info(f" 2) Best Overall Usecase Configuration: {self.csv_dir}/best_overall_usecase_gpu.csv")
+
+    def evaluate(self):
+        if self.eval_mode == "cpu":
+            self.evaluate_cpu()
+        elif self.eval_mode == "gpu":
+            self.evaluate_gpu()
+        else: 
+            exit()
+        
+    def calculate_vram(self):
+        cap = float(self.gpu_vram_safety_cap)
+        if not (0.0 < cap <= 0.99):
+            raise ValueError(f"gpu_vram_safety_cap must be in (0, 0.99], got {cap}.")
+
+        gpus = self.selected_gpus if self.selected_gpus else list_gpu_ids()
+        if not gpus:
+            raise RuntimeError("No GPUs detected for VRAM calculation.")
+
+        per_gpu_budget_mb = []
+        for gid in gpus:
+            total_gb, free_gb = get_gpu_vram(gpu_index=gid)
+            total_mb = float(total_gb) * 1024.0
+            free_mb  = float(free_gb)  * 1024.0
+
+            # hard cap vs physical total, plus never exceed currently free memory
+            budget_mb = min(total_mb * cap, free_mb)
+            per_gpu_budget_mb.append(budget_mb)
+
+            self.logger.info(
+                f"GPU {gid}: total={total_gb:.2f} GB, free={free_gb:.2f} GB, "
+                f"budget={budget_mb/1024.0:.2f} GB (cap={cap:.2f})"
+            )
+
+        budget_mb_min = float(min(per_gpu_budget_mb))
+        self.logger.info(f"Using per-GPU VRAM budget = {budget_mb_min:.0f} MB (min across selected GPUs).")
+        return budget_mb_min
+
+
+"""
+Finds the optimal CPU configuration based on evaluation results
+"""
+class OptimalCPUConfigurationFinder: 
+    def __init__(self, 
+                 eval_sys_results_dir: str, 
+                 log_file_path: str):
+        
+        self.eval_sys_results_dir = eval_sys_results_dir
+        if not self.eval_sys_results_dir or not os.path.isdir(self.eval_sys_results_dir): 
+            raise ValueError(f"Error: The provided directory path '{self.eval_sys_results_dir}' is invalid or does not exist.")
+        self.log_file_path = log_file_path
+
+        # Set up main logger and logger queue to retrive queued logs from Raylets to be passed to the main logger
+        self.logger = logging.getLogger("eqcctpro") # We named the logger eqcctpro (can be any name)
+        self.logger.setLevel(logging.INFO)
+        self.logger.propagate = False # if true, events logged to this logger will be passed to the handlers of higher level (ancestor) loggers, in addition to any handlers attached to this logger
+        if not self.logger.handlers: # avoid duplicating inits 
+            fmt = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+            # ensure parent dir
+            Path(self.log_file_path).parent.mkdir(parents=True, exist_ok=True)
+            file_h = logging.FileHandler(self.log_file_path) # Writes logs to file 
+            stream_h = logging.StreamHandler() # Sends logs to console
+            file_h.setFormatter(fmt)
+            stream_h.setFormatter(fmt)
+            self.logger.addHandler(file_h)
+            self.logger.addHandler(stream_h)
+
+
+    def find_best_overall_usecase(self):
+        """Finds the best overall CPU usecase configuation from eval results"""
+        file_path = f"{self.eval_sys_results_dir}/best_overall_usecase_cpu.csv"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"[{datetime.now()}] Error: The file '{file_path}' does not exist. Ensure it is in the correct directory.")
+
+        df_best_overall = pd.read_csv(file_path)
+        # best_config_dict = df_best_overall.set_index(df_best_overall.columns[0]).to_dict()[df_best_overall.columns[1]]
+        best_config_dict = df_best_overall.to_dict(orient='records')[0]
+        
+        # Extract required values
+        num_cpus = best_config_dict.get("Number of CPUs Allocated for Ray to Use")
+        waveform_timespace = best_config_dict.get("Total Waveform Analysis Timespace (min)")
+        total_num_timechunks = best_config_dict.get("Total Number of Timechunks")
+        num_concurrent_timechunks = best_config_dict.get("Concurrent Timechunks Used")
+        length_of_timechunks = best_config_dict.get("Length of Timechunk (min)")
+        num_concurrent_stations = best_config_dict.get("Number of Concurrent Station Tasks per Timechunk")
+        intra_threads = best_config_dict.get("Intra-parallelism Threads")
+        inter_threads = best_config_dict.get("Inter-parallelism Threads")
+        num_stations = best_config_dict.get("Number of Stations Used")
+        total_runtime = best_config_dict.get("Total Run time for Picker (s)")
+        model_used = best_config_dict.get("Model Used")
+        
+        self.logger.info("")
+        self.logger.info(f"------- Finding the Best Overall CPU Usecase Configuration Based on Available Trial Data in {self.eval_sys_results_dir} -------")
+        self.logger.info(f"Model Used: {model_used}")
+        self.logger.info(f"CPU(s): {num_cpus}")
+        self.logger.info(f"Intra-parallelism Threads: {intra_threads}")
+        self.logger.info(f"Inter-parallelism Threads: {inter_threads}")
+        self.logger.info(f"Waveform Timespace: {waveform_timespace}")
+        self.logger.info(f"Total Number of Stations Used: {num_stations}")
+        self.logger.info(f"Total Number of Timechunks: {total_num_timechunks}")
+        self.logger.info(f"Length of Timechunks (min): {length_of_timechunks}")
+        self.logger.info(f"Concurrent Timechunk Processes: {num_concurrent_timechunks}")
+        self.logger.info(f"Concurrent Station Processes Per Timechunk: {num_concurrent_stations}")
+        self.logger.info(f"Total Runtime (s): {total_runtime}")
+        self.logger.info("")
+
+        # return int(float(num_cpus)), int(float(intra_threads)), int(float(inter_threads)), int(float(num_concurrent_timechunks)), int(float(num_concurrent_stations)), int(float(num_stations))
+    
+    def find_optimal_for(self, cpu: int, station_count: int):
+        """Finds the optimal configuration for a given number of CPUs and stations."""
+        if cpu is None or station_count is None:
+            raise ValueError("Error: CPU and station_count must have valid values.")
+
+        file_path = f"{self.eval_sys_results_dir}/optimal_configurations_cpu.csv"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"[{datetime.now()}] Error: The file '{file_path}' does not exist. Ensure it is in the correct directory.")
+
+        df_optimal = pd.read_csv(file_path)
+
+        # Convert relevant columns to numeric
+        df_optimal["Number of Stations Used"] = pd.to_numeric(df_optimal["Number of Stations Used"], errors="coerce")
+        df_optimal["Number of CPUs Allocated for Ray to Use"] = pd.to_numeric(df_optimal["Number of CPUs Allocated for Ray to Use"], errors="coerce")
+        df_optimal["Number of Concurrent Station Tasks"] = pd.to_numeric(df_optimal["Number of Concurrent Station Tasks"], errors="coerce")
+        df_optimal["Total Run time for Picker (s)"] = pd.to_numeric(df_optimal["Total Run time for Picker (s)"], errors="coerce")
+
+        filtered_df = df_optimal[
+            (df_optimal["Number of CPUs Allocated for Ray to Use"] == cpu) &
+            (df_optimal["Number of Stations Used"] == station_count)]
+
+        if filtered_df.empty:
+            raise ValueError("No matching configuration found. Please enter a valid entry.")
+
+        # Finds for the "Total Run time for Picker (s)" the row with the smallest value and the '1' is to say I only want 
+        # only the single row where the smallest runtime is 
+        # iloc gets the selection of data from a numerical index from the df and turns that access point into a Series
+        best_config = filtered_df.nsmallest(1, "Total Run time for Picker (s)").iloc[0]
+
+        self.logger.info(f"------- Best CPU-EQCCTPro Configuration for Requested Input Parameters Based on the available Trial Data in {self.eval_sys_results_dir} -------")
+        self.logger.info(f"Model Used: {best_config.get('Model Used')}")
+        self.logger.info(f"CPU(s): {cpu}")
+        self.logger.info(f"Intra-parallelism Threads: {best_config['Intra-parallelism Threads']}")
+        self.logger.info(f"Inter-parallelism Threads: {best_config['Inter-parallelism Threads']}")
+        self.logger.info(f"Waveform Timespace: {best_config['Total Waveform Analysis Timespace (min)']}")
+        self.logger.info(f"Total Number of Stations Used: {station_count}")
+        self.logger.info(f"Total Number of Timechunks: {best_config['Total Number of Timechunks']}")
+        self.logger.info(f"Length of Timechunks (min): {best_config['Length of Timechunk (min)']}")
+        self.logger.info(f"Concurrent Timechunk Processes: {best_config['Concurrent Timechunks Used']}")
+        self.logger.info(f"Concurrent Station Processes Per Timechunk: {best_config['Number of Concurrent Station Tasks']}")
+        self.logger.info(f"Total Runtime (s): {best_config['Total Run time for Picker (s)']}")
+        self.logger.info("")
+
+        # return int(float(cpu)), int(float(best_config["Intra-parallelism Threads"])), int(float(best_config["Inter-parallelism Threads"])), int(float(best_config["Concurrent Timechunks Used"])), int(float(best_config["Number of Concurrent Station Tasks"])), int(float(station_count))
+
+
+class OptimalGPUConfigurationFinder:
+    """Finds the optimal GPU configuration based on evaluation system results."""
+
+    def __init__(self, 
+                 eval_sys_results_dir: str, 
+                 log_file_path: str):
+        
+        self.eval_sys_results_dir = eval_sys_results_dir
+        if not self.eval_sys_results_dir or not os.path.isdir(self.eval_sys_results_dir): 
+            raise ValueError(f"Error: The provided directory path '{self.eval_sys_results_dir}' is invalid or does not exist.")
+        self.log_file_path = log_file_path
+
+        # Set up main logger and logger queue to retrive queued logs from Raylets to be passed to the main logger
+        self.logger = logging.getLogger("eqcctpro") # We named the logger eqcctpro (can be any name)
+        self.logger.setLevel(logging.INFO)
+        self.logger.propagate = False # if true, events logged to this logger will be passed to the handlers of higher level (ancestor) loggers, in addition to any handlers attached to this logger
+        if not self.logger.handlers: # avoid duplicating inits 
+            fmt = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
+            # ensure parent dir
+            Path(self.log_file_path).parent.mkdir(parents=True, exist_ok=True)
+            file_h = logging.FileHandler(self.log_file_path) # Writes logs to file 
+            stream_h = logging.StreamHandler() # Sends logs to console
+            file_h.setFormatter(fmt)
+            stream_h.setFormatter(fmt)
+            self.logger.addHandler(file_h)
+            self.logger.addHandler(stream_h)
+
+    def find_best_overall_usecase(self):
+        """Finds the best overall GPU configuration from evaluation results."""
+        file_path = f"{self.eval_sys_results_dir}/best_overall_usecase_gpu.csv"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"[{datetime.now()}] Error: The file '{file_path}' does not exist. Ensure it is in the correct directory.")
+
+        df = pd.read_csv(file_path)
+        if df.empty:
+            raise ValueError(f"[{datetime.now()}] Error: '{file_path}' is empty.")
+
+        row = df.iloc[0]  # the best row you wrote out
+
+        # Some codepaths use two different column names for concurrency; support both
+        conc_col = "Number of Concurrent Station Tasks per Timechunk" \
+            if "Number of Concurrent Station Tasks per Timechunk" in df.columns \
+            else "Number of Concurrent Station Tasks"
+
+        # Robust GPU parse: accepts [0], (0,), "0", 0, "", None
+        num_gpus_list = _parse_gpus_field(row.get("GPUs Used"))
+        # Keep as tuple for display/consistency
+        num_gpus = tuple(num_gpus_list)
+
+        # Pull/normalize scalars
+        num_cpus = row.get("Number of CPUs Allocated for Ray to Use")
+        num_concurrent = row.get(conc_col)
+        intra_threads = row.get("Intra-parallelism Threads")
+        inter_threads = row.get("Inter-parallelism Threads")
+        num_stations = row.get("Number of Stations Used")
+        total_runtime = row.get("Total Run time for Picker (s)")
+        vram_used = row.get("Inference Actor Memory Limit (MB)")
+        model_used = row.get("Model Used")
+
+        self.logger.info("")
+        self.logger.info(f"------- Finding the Best Overall GPU Usecase Configuration Based on Available Trial Data in {self.eval_sys_results_dir} -------")
+        self.logger.info("")
+        self.logger.info(f"Model Used: {model_used}")
+        self.logger.info(f"CPU(s): {num_cpus}")
+        self.logger.info(f"GPU ID(s): {num_gpus_list}")
+        self.logger.info(f"Concurrent Predictions: {num_concurrent}")
+        self.logger.info(f"Intra-parallelism Threads: {intra_threads}")
+        self.logger.info(f"Inter-parallelism Threads: {inter_threads}")
+        self.logger.info(f"Stations: {num_stations}")
+        self.logger.info(f"Inference Actor Memory Limit (MB): {vram_used}")
+        self.logger.info(f"Total Runtime (s): {total_runtime}")
+        self.logger.info("")
+        # return int(float(num_cpus)), int(float(num_concurrent_predictions)), int(float(intra_threads)), int(float(inter_threads)), num_gpus, int(float(vram_used)), int(float(num_stations))
+
+    def find_optimal_for(self, num_cpus: int, gpu_list: list, station_count: int):
+        """Finds the optimal configuration for a given number of CPUs, GPUs, and stations."""
+        if num_cpus is None or station_count is None or gpu_list is None:
+            raise ValueError("Error: num_cpus, station_count, and gpu_list must have valid values.")
+
+        file_path = f"{self.eval_sys_results_dir}/optimal_configurations_gpu.csv"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"[{datetime.now()}] Error: The file '{file_path}' does not exist. Ensure it is in the correct directory.")
+
+        df_optimal = pd.read_csv(file_path)
+
+        # Convert relevant columns to numeric, handling NaNs
+        df_optimal["Number of Stations Used"] = pd.to_numeric(df_optimal["Number of Stations Used"], errors="coerce")
+        df_optimal["Number of CPUs Allocated for Ray to Use"] = pd.to_numeric(df_optimal["Number of CPUs Allocated for Ray to Use"], errors="coerce")
+        df_optimal["Number of Concurrent Station Tasks"] = pd.to_numeric(df_optimal["Number of Concurrent Station Tasks"], errors="coerce")
+        df_optimal["Total Run time for Picker (s)"] = pd.to_numeric(df_optimal["Total Run time for Picker (s)"], errors="coerce")
+        df_optimal["Inference Actor Memory Limit (MB)"] = pd.to_numeric(df_optimal["Inference Actor Memory Limit (MB)"], errors="coerce")
+
+        # Convert "GPUs Used" from string representation to list
+        df_optimal["GPUs Used"] = df_optimal["GPUs Used"].apply(lambda x: ast.literal_eval(x) if isinstance(x, str) else x)
+
+        # Convert GPU lists to tuples for comparison
+        df_optimal["GPUs Used"] = df_optimal["GPUs Used"].apply(lambda x: tuple(x) if isinstance(x, list) else (x,))
+
+        # Ensure gpu_list is in tuple format for comparison
+        gpu_list_tuple = tuple(gpu_list) if isinstance(gpu_list, list) else (gpu_list,)
+
+        filtered_df = df_optimal[
+            (df_optimal["Number of CPUs Allocated for Ray to Use"] == num_cpus) &
+            (df_optimal["GPUs Used"] == gpu_list_tuple) &
+            (df_optimal["Number of Stations Used"] == station_count)
+        ]
+
+        if filtered_df.empty:
+            raise ValueError("No matching configuration found. Please enter a valid entry.")
+
+        best_config = filtered_df.nsmallest(1, "Total Run time for Picker (s)").iloc[0]
+
+        self.logger.info(f"------- Best GPU-EQCCTPro Configuration for Requested Input Parameters Based on the Available Trial Data in {self.eval_sys_results_dir} -------")
+        self.logger.info(f"CPU(s): {num_cpus}")
+        self.logger.info(f"GPU(s): {gpu_list}")
+        self.logger.info(f"Concurrent Predictions: {best_config['Number of Concurrent Station Tasks']}")
+        self.logger.info(f"Intra-parallelism Threads: {best_config['Intra-parallelism Threads']}")
+        self.logger.info(f"Inter-parallelism Threads: {best_config['Inter-parallelism Threads']}")
+        self.logger.info(f"Stations: {station_count}")
+        self.logger.info(f"Inference Actor Memory Limit (MB): {best_config['Inference Actor Memory Limit (MB)']}")
+        self.logger.info(f"Total Runtime (s): {best_config['Total Run time for Picker (s)']}")
+
+        # return int(float(best_config["Number of CPUs Allocated for Ray to Use"])), \
+        #        int(float(best_config["Number of Concurrent Station Tasks"])), \
+        #        int(float(best_config["Intra-parallelism Threads"])), \
+        #        int(float(best_config["Inter-parallelism Threads"])), \
+        #        gpu_list, \
+        #        int(float(best_config["Inference Actor Memory Limit (MB)"])), \
+        #        int(float(station_count))