https://www.boundaryml.com logo
Powered by
🟡 Trouble with BAML GPT-5.4 in thinking mode and ...
# questions
j
Hi! We're long-time BAML users. We're having trouble with the GPT-5.4 in thinking mode. The prompts are large, and the context windows are full. Importantly (we think), we have a job scheduler firing off these calls asynchronously. When there are simultaneous requests of this type, BAML hangs and only returns after the timeout. It seems like it is retrying during that time. This is v 220. Any ideas? Thanks in advance.
v
heya! let me check
are you using studio by chance? it should be faily easy to see exactly the trace thats running and full webrequest there
then debugging that / seeing what internals failed may work better
also, a streaming call to that function should make it quick to debug as well!
j
Thanks @Vaibhav (BAML 🐑). We're not using Studio. Here's this is deepest we're gone:
v
@baml-sam
b
Hmm- can you describe your python setup in more detail? Do you have a sense of how many calls to call-function-sync were happening in parallel?
Here’s another option to help investigate; you should be able to run something like this in prod if you’re not already:
Copy code
import sys
import threading
import traceback
import time

def dump_threads():
    print("\n=== THREAD DUMP ===")
    frames = sys._current_frames()
    for thread in threading.enumerate():
        print(f"\n--- Thread: {thread.name} (id={thread.ident}) ---")
        frame = frames.get(thread.ident)
        if frame:
            traceback.print_stack(frame)

def watchdog(interval_secs=10):
    while True:
        time.sleep(interval_secs)
        dump_threads()

t = threading.Thread(target=watchdog, daemon=True)
t.start()
2 Views
PreviousNext
Powered by