feat: make version deployable

api/salience/__init__.py

@@ -1,27 +1,14 @@
-# Memory Sharing for ML Models
-# ============================
-# This app is designed to run with Gunicorn's --preload flag, which loads the
-# SentenceTransformer models once in the master process before forking workers.
-# On Linux, fork uses copy-on-write (COW) semantics, so workers share the
-# read-only model weights in memory rather than each loading their own copy.
-# This is critical for keeping memory usage reasonable with large transformer models.
-#
-# ResourceTracker errors on shutdown (Python 3.14):
-# When you Ctrl+C the Gunicorn process, you may see
-# "ChildProcessError: [Errno 10] No child processes"
-# from multiprocessing.resource_tracker.
-#
-# I think this is harmless. I think what happens is each forked worker gets a
-# copy of the ResourceTracker object, then each copy tries to deallocate the
-# same resources. The process still shuts down reasonbly quickly, so I'm not
-# concerned.
+# Salience API
+# ============
+# Uses a worker thread for model inference to avoid fork() issues with Metal/MPS.
+# The worker thread owns all model instances; HTTP handlers submit work via queue.
 
 print("Starting salience __init__.py...")
 
 from flask import Flask, request
 from flask_cors import CORS
 import numpy as np
-from .salience import extract, AVAILABLE_MODELS
+from .salience import submit_work, AVAILABLE_MODELS
 import json
 import time
 from collections import deque
@@ -117,7 +104,7 @@ def salience_view_default():
     if model_name not in AVAILABLE_MODELS:
         return json.dumps({'error': f'Invalid model: {model_name}'}), 400
 
-    sentence_ranges, adjacency = extract(default_source_text, model_name)
+    sentence_ranges, adjacency = submit_work(default_source_text, model_name)
 
     end_time = time.time()
     stats_tracker.add_processing_span(start_time, end_time)
@@ -146,7 +133,7 @@ def salience_view_custom():
     if not source_text:
         return json.dumps({'error': 'No text provided'}), 400
 
-    sentence_ranges, adjacency = extract(source_text, model_name)
+    sentence_ranges, adjacency = submit_work(source_text, model_name)
 
     end_time = time.time()
     stats_tracker.add_processing_span(start_time, end_time)