Closed Pipboyguy closed 10 months ago
Thanks for reporting the issue. Could you provide some more details, including:
Note: we are currently investigating an issue with the local variant when running with Docker on an ARM Macs (it seems to be very slow) – if that is your configuration you may want to try running directly using poetry
.
I ran this directly on poetry (for GPU support). This is what raised the issue initially.
I have to add that I was getting segfaults prior to it working, but then not responding to post requests.
Could you share some details about the segfaults? Also, could you share the log messages?
The segfaults were only "core dumped" by the pathway process. No more logs or indication of what went wrong,
For some reason it started working after a couple of more tries, but when I run curl -i --data '{"user": "user", "query": "How to use LLMs in Pathway?"}' http://localhost:8080/
LOGS
[08/07/23 14:04:21] INFO Preparing Pathway computation
INFO FilesystemReader-0: 0 entries (1 minibatch(es)) have been sent to the engine
INFO PythonReader-1: 0 entries (1 minibatch(es)) have been sent to the engine
and nothing else. So I'd label it a timeout because I've been leaving it for quite a while, and I have no small system at 12 cores 64GB RAM. If the logs were more verbose it be better.
Could you share what the monitoring display above the logs shows? Do the latencies keep increasing?
Could you try running it on kuba/async-cap
branch?
PATHWAY PROGRESS DASHBOARD
no. messages operator latency to wall clock [ms] lag to input [ms]
in the last in the last ───────────────────────────────────────────────────────────
connector minibatch minute since start input 30
──────────────────────────────────────────────────────────── output 30 0
FilesystemRe… 0 0 6
PythonReader… 0 0 1 Above you can see the latency of input and output operators.
The latency is measured as the difference between the time
when the operator processed the data and the time when
pathway acquired the data.
LOGS
[08/07/23 15:49:40] INFO Preparing Pathway computation
INFO FilesystemReader-0: 0 entries (1 minibatch(es)) have been sent to the engine
INFO PythonReader-1: 0 entries (1 minibatch(es)) have been sent to the engine
[08/07/23 15:49:45] INFO FilesystemReader-0: 6 entries (102 minibatch(es)) have been sent to the engine
[08/07/23 15:49:46] INFO PythonReader-1: 1 entries (126 minibatch(es)) have been sent to the engine
[08/07/23 15:49:50] INFO FilesystemReader-0: 0 entries (100 minibatch(es)) have been sent to the engine
[08/07/23 15:49:51] INFO PythonReader-1: 0 entries (101 minibatch(es)) have been sent to the engine
At first I suspected the delay is huggingface encoders being downloaded in the background, but its been way too long for that to bee the issue. Would be nice to see the logs for any such artefacts being downloaded for visibility
I ran the above on the suggested branch. Same result on the face of it.
I see 6 entries read by the engine which is what's expected (pathway-docs-small data). I couldn't reproduce this non response issue on WSL or Linux. I suspect though that the problem is not enough memory to run the embedding or inference models. Download bars are shown only on the first run before pathway is started as they are cached:
I reproduced the Seg fault error on colab. It happens when running one model on gpu (embedder by default runs on gpu if available) and the other on cpu (HFTextGenerationTask has device='cpu' by default). I'll add the device argument for the SentenceTransformerTask right away. Remember to clear the cache directory specified on .env Could you checkout PR50 @Pipboyguy
@mdmalhou Removes issue on my side. Thanks @mdmalhou
Logs don't show any activity. Contexfull works as expected