nodes-api: enforce single thread for the processor

On hyperthreaded CPUs we get two threads operating on the queue by
default on each core. This cases two threads to process queue items.
This results in pytorch errors and sometimes generates garbage.

Locking this to single thread makes sense because we are bound by the
number of GPUs in the system, not by CPU cores. And to parallelize
across GPUs we should just start multiple processors (and use async
instead of threading)

Fixes #3289

invokeai/app/services/processor.py

@@ -1,5 +1,5 @@
 import traceback
-from threading import Event, Thread
+from threading import Event, Thread, BoundedSemaphore
 
 from ..invocations.baseinvocation import InvocationContext
 from .invocation_queue import InvocationQueueItem
@@ -10,8 +10,11 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
     __invoker_thread: Thread
     __stop_event: Event
     __invoker: Invoker
+    __threadLimit: BoundedSemaphore
 
     def start(self, invoker) -> None:
+        # if we do want multithreading at some point, we could make this configurable
+        self.__threadLimit = BoundedSemaphore(1)
         self.__invoker = invoker
         self.__stop_event = Event()
         self.__invoker_thread = Thread(
@@ -20,7 +23,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
             kwargs=dict(stop_event=self.__stop_event),
         )
         self.__invoker_thread.daemon = (
-            True  # TODO: probably better to just not use threads?
+            True  # TODO: make async and do not use threads
         )
         self.__invoker_thread.start()
 
@@ -29,6 +32,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
 
     def __process(self, stop_event: Event):
         try:
+            self.__threadLimit.acquire()
             while not stop_event.is_set():
                 queue_item: InvocationQueueItem = self.__invoker.services.queue.get()
                 if not queue_item:  # Probably stopping
@@ -110,7 +114,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                     )
 
                     pass
-                
+
                 # Check queue to see if this is canceled, and skip if so
                 if self.__invoker.services.queue.is_canceled(
                     graph_execution_state.id
@@ -127,4 +131,6 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                     )
 
         except KeyboardInterrupt:
-            ...  # Log something?
+            pass  # Log something? KeyboardInterrupt is probably not going to be seen by the processor
+        finally:
+            self.__threadLimit.release()