Nunca se me ha bloqueado el ordenador arrancando con los .run de Nvidia (alguna vez he tenido que esperar a un driver compatible con el nuevo kernel pero nada más).
En esta ocasión, el driver 595.71.05 (y otros que he probado) tienen una curiosidad. Me dejan recompilarlos para 7.0.7 (y anteriores) y hacen su rebuild initramfs, todo bien.
Pero cuando reinicio el ordenador con el kernel 7.0.7 el proceso de arranque se detiene. No sale kernel panic pero se detiene de golpe.
La info que he podido conseguir: sudo journalctl -k -b -2(-1 y 0 son con 6.19.12 a donde he tenido que volver)
En tu captura aparece nvidia-modeset: ERROR: GPU:0: Failed to initialize DMA, que apunta a que el módulo de NVIDIA sí carga, pero falla al inicializar la GPU durante el arranque. Ese tipo de fallo suele venir de una combinación de kernel nuevo + módulo NVIDIA no preparado para ese kernel.
Mira en el forum inglés si hay algún hilo reciente de ese problema, que aporta una solución.
O provisionalmente, desinstala el .run , añade el repo de Nvidia de oo e instala el driver. Hasta que salga un driver que solucione el problema.
Todo el mundo tiene problemas (o mucha gente) aunque parecen usar todos los paquetes .rpm de los repos Nvidia. A mi me sucede lo mismo con el .run (pantalla negra, magenta o azul incluidas).
Acabo de actualizar con RPM, con el kernel 7.07 aparece a cada rato el cartelito:
|Se han reiniciado los efectos de escritorio debido a un reinicio gráfico|
No parece grave, pero el botón de inicio del Gecko esta congelado y no funciona.
Y los CVE son bastante abundantes:
Para colmo ahora, con los buscadores web infectados con la IA hace imposible encontrar algo. Parece implementada para espiarnos y a alimentar a Palantir.
Logré mejorar el comportamiento, desactivando los efectos de escritorio, en la segunda pantalla. Había probado a desactivar todos los efectos, pero después no aparecía la pantalla de login.
Este es el dmesg:
[ 5.660305] [ T972] nvidia: loading out-of-tree module taints kernel.
[ 5.660311] [ T972] nvidia: module license 'NVIDIA' taints kernel.
[ 5.660313] [ T972] Disabling lock debugging due to kernel taint
[ 5.660315] [ T972] nvidia: module verification failed: signature and/or required key missing - tainting kernel
[ 5.660317] [ T972] nvidia: module license taints kernel.
[ 5.683937] [ T132] kauditd_printk_skb: 4 callbacks suppressed
[ 5.683940] [ T132] audit: type=1400 audit(1779382307.752:15): avc: denied { search } for pid=1047 comm="plymouthd" name="var" dev="nvme1n1p1" ino=262 scontext=system_u:system_r:plymouthd_t:s0 tcontext=system_u:object_r:unlabeled_t:s0 tclass=dir permissive=0
[ 5.683944] [ T132] audit: type=1400 audit(1779382307.752:16): avc: denied { search } for pid=1047 comm="plymouthd" name="var" dev="nvme1n1p1" ino=262 scontext=system_u:system_r:plymouthd_t:s0 tcontext=system_u:object_r:unlabeled_t:s0 tclass=dir permissive=0
[ 5.683946] [ T132] audit: type=1400 audit(1779382307.752:17): avc: denied { search } for pid=1047 comm="plymouthd" name="var" dev="nvme1n1p1" ino=262 scontext=system_u:system_r:plymouthd_t:s0 tcontext=system_u:object_r:unlabeled_t:s0 tclass=dir permissive=0
[ 6.068512] [ T972] nvidia-nvlink: Nvlink Core is being initialized, major device number 239
[ 6.072171] [ T972] nvidia 0000:0f:00.0: vgaarb: VGA decodes changed: olddecodes=io+mem,decodes=none:owns=io+mem
[ 6.117425] [ T972] NVRM: loading NVIDIA UNIX x86_64 Kernel Module 580.159.03 Fri Apr 24 06:16:47 UTC 2026
[ 6.196481] [ T972] nvidia_uvm: module uses symbols nvUvmInterfaceDisableAccessCntr from proprietary module nvidia, inheriting taint.
[ 6.197164] [ T909] systemd-journald[909]: Received client request to flush runtime journal.
[ 6.243170] [ T909] systemd-journald[909]: /var/log/journal/9c8378856a69425e88eb29f0faef0ec1/system.journal: Realtime clock jumped backwards relative to last journal entry, rotating.
[ 6.243175] [ T909] systemd-journald[909]: Rotating system journal.
[ 6.429864] [ T972] nvidia-modeset: Loading NVIDIA Kernel Mode Setting Driver for UNIX platforms 580.159.03 Fri Apr 24 05:49:32 UTC 2026
[ 6.434996] [ T972] [drm] [nvidia-drm] [GPU ID 0x00000f00] Loading driver
[ 6.439213] [ T1069] SGI XFS with ACLs, security attributes, realtime, quota, no debug enabled
[ 6.441005] [ T1062] XFS (nvme0n1p2): Mounting V5 Filesystem b6aa5297-de48-4aac-a789-4b99ddb19e6e
[ 6.450255] [ T1062] XFS (nvme0n1p2): Ending clean mount
[ 6.904735] [ T983] snd_hda_codec_ca0132 hdaudioC0D1: ca0132 DSP downloaded and running
[ 7.333830] [ T1230] device-mapper: uevent: version 1.0.3
[ 7.334089] [ T1230] device-mapper: ioctl: 4.50.0-ioctl (2025-04-28) initialised: dm-devel@lists.linux.dev
[ 7.346453] [ T1182] nct6775: Found NCT6798D or compatible chip at 0x2e:0x290
[ 7.359514] [ T1194] nvme nvme0: using unchecked data buffer
[ 7.477660] [ T1346] NET: Registered PF_QIPCRTR protocol family
[ 7.862650] [ T972] [drm] Initialized nvidia-drm 0.0.0 for 0000:0f:00.0 on minor 1
[ 7.880652] [ T972] nvidia 0000:0f:00.0: vgaarb: deactivate vga console
[ 7.961054] [ T972] fbcon: nvidia-drmdrmfb (fb0) is primary device
[ 7.961056] [ T972] fbcon: Deferring console take-over
[ 7.961060] [ T972] nvidia 0000:0f:00.0: [drm] fb0: nvidia-drmdrmfb frame buffer device
[ 8.239324] [ T250] input: HDA Creative Mic as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input11
[ 8.239408] [ T250] input: HDA Creative Line Out Front as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input12
[ 8.239464] [ T250] input: HDA Creative Line Out CLFE as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input13
[ 8.239518] [ T250] input: HDA Creative Headphone as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input14
[ 8.405179] [ C7] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 9.502171] [ C7] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 10.000163] [ C7] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 10.707004] [ T182] igb 0000:08:00.0 enp8s0: igb: enp8s0 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: RX/TX
[ 11.063647] [ T1744] NET: Registered PF_PACKET protocol family
[ 15.946407] [ C3] NVRM: GPU at PCI:0000:0f:00: GPU-2d9a31aa-5972-4726-5a61-6092856686b1
[ 15.946410] [ C3] NVRM: Xid (PCI:0000:0f:00): 8, pid=1713, name=sddm-greeter-qt, channel 0x0000000a
[ 16.919099] [ C7] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 20.126065] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 21.423041] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 21.739959] [ T2018] block nvme0n1: No UUID available providing old NGUID
[ 24.138432] [ C11] NVRM: Xid (PCI:0000:0f:00): 8, pid=2012, name=ksplashqml, channel 0x0000000d
[ 25.270001] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 27.838974] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 28.299603] [ T116] BTRFS info (device nvme1n1p1): qgroup scan completed (inconsistency flag cleared)
[ 29.774953] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 31.045940] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 32.330413] [ C14] NVRM: Xid (PCI:0000:0f:00): 8, pid=2012, name=ksplashqml, channel 0x0000000d
[ 34.261907] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 34.895901] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 35.549894] [ C8] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 39.399855] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 40.693838] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 41.989824] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 42.629818] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 47.125772] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 48.422866] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 48.714244] [ C11] NVRM: Xid (PCI:0000:0f:00): 8, pid=2009, name=kwin_wayland, channel 0x0000000a
[ 53.565705] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 54.205706] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 55.493791] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 56.133687] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 57.414663] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 63.189603] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 70.265246] [ T3125] NOTICE: Automounting of tracing to debugfs is deprecated and will be removed in 2030
[ 70.887522] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 72.165511] [ C11] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 73.290700] [ C14] NVRM: Xid (PCI:0000:0f:00): 8, pid=2009, name=kwin_wayland, channel 0x0000000a
Al final he bloqueado los paquetes de kernel (kernel-default, kernel-default-devel, kernel-devel, kernel-macros y kernel-syms) y el sudo zypper dup de hoy lo he hecho sin actualizar el kernel (manteniendo el 6.19.12).
Ya no actualizo el kernel hasta que lea en algún sitio que se ha arreglado sin añadir cosas raras a los parámetros del kernel (cosas que no me han funcionado como los diferentes initcall_blacklist=loquesea.
Hoy he vuelto a actualizar, pero entes no cargaba el entorno gráfico con el kernel 6.19. Tal vez sean incompatibles los kernel-sysms y firmware del más nuevo. Tal vez tendría que reinstalar los paquetes complementarios del kernel 6.19.
Con el kernel 7.0.9-1 parece un poco más estable:
[ 5.096929] [ T953] nvidia: loading out-of-tree module taints kernel.
[ 5.096935] [ T953] nvidia: module license ‘NVIDIA’ taints kernel.
[ 5.096936] [ T953] Disabling lock debugging due to kernel taint
[ 5.096939] [ T953] nvidia: module verification failed: signature and/or required key missing - tainting kernel
[ 5.096940] [ T953] nvidia: module license taints kernel.
[ 5.100946] [ T167] input: HDA NVidia HDMI/DP,pcm=3 as /devices/pci0000:00/0000:00:03.1/0000:0f:00.1/sound/card1/input7
[ 5.101037] [ T167] input: HDA NVidia HDMI/DP,pcm=7 as /devices/pci0000:00/0000:00:03.1/0000:0f:00.1/sound/card1/input8
[ 5.101097] [ T167] input: HDA NVidia HDMI/DP,pcm=8 as /devices/pci0000:00/0000:00:03.1/0000:0f:00.1/sound/card1/input9
[ 5.101171] [ T167] input: HDA NVidia HDMI/DP,pcm=9 as /devices/pci0000:00/0000:00:03.1/0000:0f:00.1/sound/card1/input10
[ 5.571027] [ T953] nvidia-nvlink: Nvlink Core is being initialized, major device number 239
[ 5.574891] [ T953] nvidia 0000:0f:00.0: vgaarb: VGA decodes changed: olddecodes=io+mem,decodes=none:owns=io+mem
[ 5.610877] [ T908] systemd-journald[908]: Received client request to flush runtime journal.
[ 5.625285] [ T953] NVRM: loading NVIDIA UNIX x86_64 Kernel Module 580.159.03 Fri Apr 24 06:16:47 UTC 2026
[ 5.662500] [ T953] nvidia_uvm: module uses symbols nvUvmInterfaceDisableAccessCntr from proprietary module nvidia, inheriting taint.
[ 5.719593] [ T969] snd_hda_codec_ca0132 hdaudioC0D1: ca0132 DSP downloaded and running
[ 5.817611] [ T1065] SGI XFS with ACLs, security attributes, realtime, quota, no debug enabled
[ 5.873622] [ T1056] XFS (nvme1n1p2): Mounting V5 Filesystem b6aa5297-de48-4aac-a789-4b99ddb19e6e
[ 5.882070] [ T1056] XFS (nvme1n1p2): Ending clean mount
[ 5.893081] [ T953] nvidia-modeset: Loading NVIDIA Kernel Mode Setting Driver for UNIX platforms 580.159.03 Fri Apr 24 05:49:32 UTC 2026
[ 5.897816] [ T953] [drm] [nvidia-drm] [GPU ID 0x00000f00] Loading driver
[ 6.658886] [ T145] input: HDA Creative Mic as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input11
[ 6.658969] [ T145] input: HDA Creative Line Out Front as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input12
[ 6.659043] [ T145] input: HDA Creative Line Out CLFE as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input13
[ 6.659099] [ T145] input: HDA Creative Headphone as /devices/pci0000:00/0000:00:01.2/0000:01:00.0/0000:02:04.0/0000:0a:00.0/sound/card0/input14
[ 6.784661] [ T1183] nct6775: Found NCT6798D or compatible chip at 0x2e:0x290
[ 6.804937] [ T1241] device-mapper: uevent: version 1.0.3
[ 6.805061] [ T1241] device-mapper: ioctl: 4.50.0-ioctl (2025-04-28) initialised: dm-devel@lists.linux.dev
[ 6.819756] [ T1197] nvme nvme0: using unchecked data buffer
[ 6.983220] [ T1339] NET: Registered PF_QIPCRTR protocol family
[ 7.390742] [ T953] [drm] Initialized nvidia-drm 0.0.0 for 0000:0f:00.0 on minor 1
[ 7.409511] [ T953] nvidia 0000:0f:00.0: vgaarb: deactivate vga console
[ 7.488467] [ T953] fbcon: nvidia-drmdrmfb (fb0) is primary device
[ 7.488470] [ T953] fbcon: Deferring console take-over
[ 7.488474] [ T953] nvidia 0000:0f:00.0: [drm] fb0: nvidia-drmdrmfb frame buffer device
[ 7.890013] [ C5] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 8.989974] [ C5] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 10.234849] [ T254] igb 0000:08:00.0 enp8s0: igb: enp8s0 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: RX/TX
[ 10.588572] [ T1760] NET: Registered PF_PACKET protocol family
[ 14.163912] [ C5] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 15.498229] [ C0] NVRM: GPU at PCI:0000:0f:00: GPU-2d9a31aa-5972-4726-5a61-6092856686b1
[ 15.498235] [ C0] NVRM: Xid (PCI:0000:0f:00): 8, pid=1728, name=sddm-greeter-qt, channel 0x0000000a
[ 19.447857] [ C0] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 22.545845] [ T2081] block nvme0n1: No UUID available providing old NGUID
[ 23.690256] [ C14] NVRM: Xid (PCI:0000:0f:00): 8, pid=2058, name=ksplashqml, channel 0x0000000d
[ 24.395792] [ C0] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 29.006741] [ C0] hid-generic 0003:051D:0002.0003: Event data for report 32 was too short (4 vs 2)
[ 31.883853] [ C14] NVRM: Xid (PCI:0000:0f:00): 8, pid=2055, name=kwin_wayland, channel 0x0000000a
±----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.03 Driver Version: 580.159.03 CUDA Version: 13.0 |
±----------------------------------------±-----------------------±---------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 4060 Ti Off | 00000000:0F:00.0 On | N/A |
| 0% 29C P8 6W / 165W | 785MiB / 8188MiB | 1% Default |
| | | N/A |
±----------------------------------------±-----------------------±---------------------+
±----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 1602 G /usr/bin/Xorg.bin 15MiB |
| 0 N/A N/A 1958 G /usr/bin/ksecretd 2MiB |
| 0 N/A N/A 2055 G /usr/bin/kwin_wayland 25MiB |
| 0 N/A N/A 2158 G /usr/bin/Xwayland 7MiB |
| 0 N/A N/A 2237 G /usr/bin/ksmserver 2MiB |
| 0 N/A N/A 2239 G /usr/bin/kded6 2MiB |
| 0 N/A N/A 2344 G /usr/bin/plasmashell 116MiB |
| 0 N/A N/A 2363 G /usr/bin/kaccess 2MiB |
| 0 N/A N/A 2364 G …it-kde-authentication-agent-1 2MiB |
| 0 N/A N/A 2366 G …ibexec/xdg-desktop-portal-kde 2MiB |
| 0 N/A N/A 2502 G /usr/bin/python3 2MiB |
| 0 N/A N/A 2508 G /usr/bin/kdeconnectd 2MiB |
| 0 N/A N/A 2521 G /usr/libexec/DiscoverNotifier 2MiB |
| 0 N/A N/A 2523 G /usr/bin/kalendarac 2MiB |
| 0 N/A N/A 2720 G /usr/bin/akonadi_control 2MiB |
| 0 N/A N/A 2790 G …bin/akonadi_archivemail_agent 2MiB |
| 0 N/A N/A 2793 G …konadi_followupreminder_agent 2MiB |
| 0 N/A N/A 2797 G …/akonadi_maildispatcher_agent 2MiB |
| 0 N/A N/A 2798 G …/bin/akonadi_mailfilter_agent 2MiB |
| 0 N/A N/A 2799 G /usr/bin/akonadi_mailmerge_agent 2MiB |
| 0 N/A N/A 2802 G /usr/bin/akonadi_migration_agent 2MiB |
| 0 N/A N/A 2805 G …akonadi_newmailnotifier_agent 2MiB |
| 0 N/A N/A 2806 G /usr/bin/akonadi_sendlater_agent 2MiB |
| 0 N/A N/A 2809 G …/akonadi_unifiedmailbox_agent 2MiB |
| 0 N/A N/A 3411 G /usr/lib64/firefox/firefox 228MiB |
| 0 N/A N/A 5604 G /usr/bin/konsole 2MiB |
±----------------------------------------------------------------------------------------+
Ahora tira el error de kwin: “Se han reiniciado los efectos del escritorio debido a un reinicio gráfico” entre los 2 y 5 minutos.
Así que desbloqueé el kernel, actualicé a 7.0.9, reinicié, instalé el driver 610.43.02 y reinicié de nuevo.
SDDM aparece (¡bien!) pero increíblemente lento. Entonces, si SDDM va lento, KDE Plasma irá peor.
En lugar de entrar en KDE Plasma, reinicié y añadí en los parámetros del kernel en Grub initcall_blacklist=sysfb_init. SDDM aparece rápido (y tengo TTY y los mensajes de inicio disponibles) pero cuando entro en KDE Plasma, Kwin falla.
Inicié un juego en Steam y Kwin volvió a fallar.
Nota: Parece que el driver 610.43.02 tiene problemas con algunos compositores de Wayland y parece que KDE Plasma 6.7 solucionará este problema. Los desarrolladores están trabajando para arreglar 6.6:
Entonces… “Probaré con el 595.71.05 nuevamente pero con el parámetro del kernel”, me dije.
Hice el proceso y todo va bien (gracias a initcall_blacklist=sysfb_init). Pero antes de 7.0.9 este parámetro no me funcionaba.
Yo intenté migrar ayer del driver 580.159.03-54.1 a 595.71.05-2.2 con kernel 7.0.9-1, pero no hay un sustituto directo a nvidia-driver-G06-kmp-default-580.159.03_k7.0.7_1-54.1, lo más parecido es el nvidia-video-G07-595.71.05-13.1.
Terminaba con la pantalla negra cuando debía iniciar Plasma-Wayland y lo único que aparecía, era el puntero del mouse.
Opté por probar con el .run y me descargó la versión 595.71.05, con idéntico resultado.
Opté por hacer un rollback a la última actualización que todavía funcionaba.
¿Será que la Arquitectura NVIDIA Ada Lovelace no soporta los drivers más recientes?
A mi me ha ido con el kernel 7.0.9 (no con los anteriores).
Y sí, hay algún problema con el driver Nvidia con openSUSE desde que salió el kernel 7.0. Lee el enlace que menciono aquí (en inglés y ya ha superado las 600 respuestas de largo)
Yo era fan de AMD, y me burlaba cada vez que los Nvidi-osos tenían problemas…
Hasta que intenté implementar ROCM, y me encontré que para poder implementarlo, debía comprar las tarjetas de video más caras de AMD, llegando a la conclusión de que era más práctico y barato migrar a Nvidia y ejecutar CUDA nativo. Me compré una RTX 4060 Ti. Primero había comprado la 4060 a secas, pero no tiene CUDA habilitado, para Laptops si, está habilitado. Me quedé corto con los 8 GB de memoria para cargar los modelos Large para transcribir audio:
Recientemente encontré un modelo large-v3-turbo que carga bien, porque está en fp16, en vez de fp32, como el large-v3 a secas.
