| 25 Jan 2025 |
 laurynasp | * based on journalctl -u wyoming-faster-whisper.., to process simple queries like "what time is it?" it takes
~30s on pi 4B
~4min on my desktop Ryzen 7 2700 (ryzen runs old nixos version, but I would be surprisied if older software would have such a huge impact) | 07:49:43 |
| laurynasp | * based on journalctl -u wyoming-faster-whisper.., to process simple queries like "what time is it?" with "medium-int8" model it takes
~30s on pi 4B
~4min on my desktop Ryzen 7 2700 (ryzen runs old nixos version, but I would be surprisied if older software would have such a huge impact) | 07:51:38 |
| laurynasp | Similarly, which models do you guys use for faster-whisper? | 07:52:21 |
 uep | I'm running on a Lenovo m920q tiny.. usff PC | 08:02:16 |
| uep | It's.. rather slow and I haven't looked into it yet, really only had time for some initial random play | 08:03:13 |
| uep | I went off to unexpose several hundred entities so it wouldn't spend forever trying to match them all | 08:04:42 |
| uep | Haven't really done much more testing since doing that | 08:05:03 |
| laurynasp | yeah, I haven't done much testing either, I just thought I could reduce 30s by using beefier cpu, and was suprised that it takes almost 10 times longer outside raspberry 4.
I tested against framework's Ryzen 7 7840U, and it is also 3-4min...
Upon closer investigations, faster-whisper on both Ryzen machine printed error wyoming-faster-whisper[2771726]: Error in cpuinfo: failed to parse processor information from /proc/cpuinfo, so perhaps to downgrades to single thread or smth
| 08:12:21 |
| laurynasp | * yeah, I haven't done much testing either, I just thought I could reduce 30s by using beefier cpu, and was suprised that it takes almost 10 times longer outside raspberry 4.
I have just tested against framework's Ryzen 7 7840U, and it is also 3-4min...
Upon closer investigations, faster-whisper on both Ryzen machine printed error wyoming-faster-whisper[2771726]: Error in cpuinfo: failed to parse processor information from /proc/cpuinfo, so perhaps to downgrades to single thread or smth
| 08:12:58 |
| laurynasp | * yeah, I haven't done much testing either, I just thought I could reduce 30s by using beefier cpu, and was suprised that it takes almost 10 times longer outside raspberry 4.
I have just tested against framework's Ryzen 7 7840U, and it is also 3-4min...
Upon closer investigations, faster-whisper on both Ryzen machines printed error wyoming-faster-whisper[2771726]: Error in cpuinfo: failed to parse processor information from /proc/cpuinfo, so perhaps it downgrades to single thread or smth
| 08:24:37 |
| uep | 
Download image.png | 08:28:49 |
| uep | it has these things in it, and it's running in a nixos-container so no access to whatever (measly integrated) gpu either | 08:29:35 |
 @hexa:lossy.network | laurynasp: it should be able to access /proc/cpuinfo since https://github.com/NixOS/nixpkgs/pull/372898 | 09:28:54 |
| @hexa:lossy.network | can you check systemctl show -P ProcSubset on your wyoming-faster-whisper systemd unit and report the value back? | 09:29:37 |
| @hexa:lossy.network | also check the new option description for model https://search.nixos.org/options?channel=unstable&show=services.wyoming.faster-whisper.servers.%3Cname%3E.model&from=0&size=50&sort=relevance&type=packages&query=faster-whisper | 09:38:26 |
| @hexa:lossy.network | * also check the new option description for model | 09:40:08 |
| laurynasp | value for systemctl show -P ProcSubset wyoming-faster-whisper-whisper.service:
all on pi 4 that works well (i.e. no cpu error, and proceeses in 30s), it is running nixos-unstable from yesterday, so it has the the PR in it
pid on both ryzen systems, both of which report cpu error, and both systems are older than the PR | 09:57:22 |
| laurynasp | Is there a lifehack to check if a PR has reached nixos-24.11? or is it just grepping in git branch/commits? | 09:58:25 |
| laurynasp | I see the PR has backport to release-24.11, so I guess it is on the stable branch now too | 09:59:58 |
| laurynasp | I will update 24.11 and report back | 10:00:14 |
| @hexa:lossy.network | nixpk.ga | 10:04:22 |
| @hexa:lossy.network | * | 10:04:26 |
|  @number42:matrix.rezel.net left the room. | 10:07:21 |
| laurynasp | hexa: Just retested with most recent 24.11 - that PR did wonders! Reduced from ~4min to 8 seconds, around a factor of 40x! | 10:56:48 |
| @hexa:lossy.network | cool | 10:57:04 |
| @hexa:lossy.network | might want to try a smaller model to make the pipeline more responsive | 10:57:31 |
| laurynasp | will start experimenting later today - thanks for the hint to look at the updated description for model option | 10:58:50 |
| laurynasp | * will start experimenting later today - thanks for the hint to look at the updated description for model option - its very useful | 10:59:13 |
 dotlambda | How do I search HACS without having it installed? | 21:56:17 |
| @hexa:lossy.network | https://github.com/hacs/default/blob/master/integration | 21:57:36 |
