home-assistant / addons

:heavy_plus_sign: Docker add-ons for Home Assistant
https://home-assistant.io/hassio/
Apache License 2.0
1.55k stars 1.5k forks source link

Whisper crash after version 1.0.1. #3403

Closed andregoncalvespires closed 9 months ago

andregoncalvespires commented 10 months ago

Describe the issue you are experiencing

Use HASSOS on a proxmox virtual machine. After upgrading Whisper to version 1.0.1, the error "FATAL: Your CPU does not support the Advanced Vector Extensions required by Whisper." started to be displayed in the addon log.

What type of installation are you running?

Home Assistant OS

Which operating system are you running on?

Home Assistant Operating System

Which add-on are you reporting an issue with?

Whisper

What is the version of the add-on?

1.0.1

Steps to reproduce the issue

  1. Use HASSOS in VM proxmox
  2. Instal or Update 1.0.1 whisper addon version
  3. Start addon

System Health information

System Information

version core-2024.1.2
installation_type Home Assistant OS
dev false
hassio true
docker true
user root
virtualenv false
python_version 3.11.6
os_name Linux
os_version 6.1.70-haos
arch x86_64
timezone America/Sao_Paulo
config_dir /config
Home Assistant Community Store GitHub API | ok -- | -- GitHub Content | ok GitHub Web | ok GitHub API Calls Remaining | 4481 Installed Version | 1.33.0 Stage | running Available Repositories | 1381 Downloaded Repositories | 38
Home Assistant Cloud logged_in | false -- | -- can_reach_cert_server | ok can_reach_cloud_auth | ok can_reach_cloud | ok
Home Assistant Supervisor host_os | Home Assistant OS 11.3 -- | -- update_channel | stable supervisor_version | supervisor-2023.12.0 agent_version | 1.6.0 docker_version | 24.0.7 disk_total | 30.8 GB disk_used | 17.6 GB healthy | true supported | true board | ova supervisor_api | ok version_api | ok installed_addons | Advanced SSH & Web Terminal (17.0.2), ESPHome (2023.12.5), Grafana (9.1.2), Home Assistant Google Drive Backup (0.105.2), InfluxDB (5.0.0), MQTT Explorer (browser-1.0.1), Piper (1.4.0), Samba share (12.2.0), Studio Code Server (5.15.0), Zigbee2MQTT (1.35.1-1), openWakeWord (1.8.2), Matter Server (5.0.2), Mosquitto broker (6.4.0), Whisper (1.0.1)
Dashboards dashboards | 2 -- | -- resources | 23 views | 38 mode | storage
Recorder oldest_recorder_run | 30 de dezembro de 2023 às 03:46 -- | -- current_recorder_run | 8 de janeiro de 2024 às 20:44 estimated_db_size | 1404.08 MiB database_engine | sqlite database_version | 3.41.2
Sonoff version | 3.5.4 (a4a8c5f) -- | -- cloud_online | 2 / 2 local_online | 2 / 2
Spotify api_endpoint_reachable | ok -- | --

Anything in the Supervisor logs that might be useful for us?

No response

Anything in the add-on logs that might be useful for us?

No response

Additional information

No response

### Tasks
emwjacobson commented 9 months ago

+1 for this. Whisper was running fine for me pre-update. Post update it won't start with the same "FATAL" error.

UPDATE: My CPUs in Proxmox were set to x86_64-v2-AES. I updated this to "host", though I think the v3 or v4 would also work, and Whisper was able to start again.

Bredick commented 9 months ago

I'm having the same problem, running haos vm on esxi on an atom c3758 CPU.

Disappointed that "Handle unsupported CPU configurations" means "arbitrarily make them not work anymore"

cwquek commented 9 months ago

+1 I am running on Proxmox, Whisper was running pre-update. After update to 1.0.1, it won't start with the same "FATAL" error.

Goopdegooo commented 9 months ago

Running fine with me on version 1.0.0, updated and [14:12:37] FATAL: Your CPU does not support the Advanced Vector Extensions required by Whisper. [14:12:37] INFO: Service exited with code 1 (by signal 0)

Now will not start, CPU N5105. No AVX support, but again was working fine on previous version 1.0.0.

g470258 commented 9 months ago

CPU Intel(R) Celeron(R) CPU 3965U. No AVX support, but was working fine on previous version 1.0.0. Now there is a cyclical error in the log, and the button to stop the addon is missing.

Scarriffle commented 9 months ago

Same Problem since updating to version 1.0.1, CPU: Intel Pentium J5005

8OND007 commented 9 months ago

same problem here running a ESXi v.7 u3 with ATOM CPU E3845 processor (4 CPU's). Did not had this problem with previous Whisper version. (small int8, 2 beams) Voice to Text sucks pretty bad locally. Even on the above ESXi VM running on Intel CPU's recognition time is up to 75 seconds.

I'm running a DuckDNS SSL HA. I have a 20/2 mbit internet connection. Maybe it takes long to "upload" the voice file to my SSL HA (even locally), because of max. 2 mbit upload. I only see 2 possible solution for local voice to text :

  1. use an external GPU/TPU device like google Coral to speed up conversion.
  2. use an external website service (API) for voice to text conversion.

Don't understand why it takes up to 75 seconds or more. Completely unusable so I'm switching back to my Google Nest device which is linked through HA's Google Assistant Integration. Year of the voice failed for me. It only "works" if you used the paid subscription Nabu Casa account which runs these voice to text services in the Cloud. (which is complete opposite to HA's existence sence "local only").

opensource4life commented 9 months ago

Same error here, CPU: AMD Ryzen 5 PRO 3400G. Worked fine till update 1.0.1

zombie241088 commented 9 months ago

Same here CPU: Intel® Celeron® Processor N5095 on mini pc with Proxmox

zeekens commented 9 months ago

Same here, but @emwjacobson trick of setting the cpu type to "host" on Proxmox seem to have fixed it

andregoncalvespires commented 9 months ago

I was already using it as a host, and the error for N5105 still occurs

HAinCA commented 9 months ago

It is extremely frustrating when you're a good techie and test something in your test environment, then go to deploy it in live only to have it fail because of an overnight update. An overnight update which is mandatory on new installs with no way to select an older version.

I understand there are issues with downgrades causing problems with database schema incompatibilities, but blocking old versions for fresh installs strikes me as almost malicious. I could have Whisper working on my live HA instance right now if I was permitted to install 1.0.0. Users should be able to remove an add-on and install the version of their choice - throw in a checkbox under a disclaimer or something, but don't block it entirely.

ninoger commented 9 months ago

I confirm. Same "Your CPU does not support the Advanced Vector Extensions required by Whisper" error after 1.0.1 update. Up to version 1.0.0 everything is fine. Home assistant is on mini pc N5095, version 2024.1.2, supervisor 2023.12.0, OS 11.3. No Proxmox or VM

dmspils commented 9 months ago

+1

frodo777 commented 9 months ago

Same Problem since updating to version 1.0.1. Proxmox intel i5 VM

emwjacobson commented 9 months ago

This change comes from commit 314ec9f54966ae14099bb874f1e5547b8cdd46fa introducing a new check for AVX instructions on the machine.

The program can obviously run without AVX, seeing as it has worked for users in the past without AVX-enabled CPUs.

Would it be possible to have it issue a Warning into the console that the CPU does not support AVX instructions, rather than outright exiting?

krzysiek2788 commented 9 months ago

Same issue truenas scale VM with Xeon e5-2640 v2

GrahamHollis commented 9 months ago

FATAL: Your CPU does not support the Advanced Vector Extensions required by Whisper. Since version 1.0.1. Proxmox VM running HAOS. Host CPU AMD Ryzen 7 5800U

MatteoD77 commented 9 months ago

Now will not start, CPU N5095 (Jusper Lake) that support AVX

GrahamHollis commented 9 months ago

I fixed this by changing the Proxmox VM CPU type from default 'kvm64' to 'host'. That allows AVX cpu extensions to be used by the VM.

agners commented 9 months ago

Hm, I actually don't have an affected machine, this got implemented based on user feedback, specifically https://github.com/home-assistant/addons/issues/3035#issuecomment-1862765838 and #3355.

Obviously, if it used to work on other machines without AVX, we should not exit then!

I wonder if those user were missing another instruction set as well, which lead to actual crashes (SIGILL). :thinking:

GrahamHollis commented 9 months ago

Hm, I actually don't have an affected machine, this got implemented based on user feedback, specifically #3035 (comment) and #3355.

Obviously, if it used to work on other machines without AVX, we should not exit then!

I wonder if those user were missing another instruction set as well, which lead to actual crashes (SIGILL). 🤔

It works without AVX but it is very slow. The change made was exclude CPU's that don't support AVX extensions because of this slow response. There is a hotfix in the works to change the exclusion to a waring. However it will be much better to to fix the underlying problem and change the VM cpu to support AVX instructions as discussed in this post https://forum.proxmox.com/threads/cpu-type-host-vs-kvm64.111165/

andregoncalvespires commented 9 months ago

Version 1.0.2 is working for me. Tks