Summary
Local video generation on Linux exhausts system RAM and swap before normal GPU inference begins.
The problem is reproducible in both LTX Desktop 1.2.5 and 1.2.6.
It occurs with both LTX 2.3 Fast and LTX 2.5 Fast.
The GPU still has almost all VRAM available when the system runs out of host memory.
Environment
- LTX Desktop: reproduced on 1.2.5 and 1.2.6
- OS: Debian 13 x86_64
- GPU: NVIDIA RTX PRO 2000 Blackwell
- VRAM: 16 GB
- System RAM: 32 GB
- Swap: 11 GB
- Generation mode: local
Reproduction
- Start LTX Desktop.
- Select either LTX 2.3 Fast or LTX 2.5 Fast.
- Start a local video generation.
- Monitor system RAM, swap, GPU VRAM and GPU utilization.
Observed behavior
During model loading:
- System RAM rises to almost the full available memory (~32.7 / 33.5 GB).
- Swap reaches 100%.
- GPU VRAM remains almost unused, around 1.0-1.1 GB used.
- Roughly 14.9 GB of VRAM is still free.
- GPU utilization remains very low.
- Normal GPU inference never starts.
- The Linux kernel triggers a global OOM and kills the LTX Desktop python3 backend.
Kernel log:
chrome invoked oom-killer: gfp_mask=0x140cca(GFP_HIGHUSER_MOVABLE|__GFP_COMP), order=0, oom_score_adj=300
oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=user.slice,mems_allowed=0,global_oom,task_memcg=/user.slice/user-1000.slice/user@1000.service/app.slice/app-ltx-desktop-45487.scope,task=python3,pid=45565,uid=1000
Out of memory: Killed process 45565 (python3) total-vm:137650648kB, anon-rss:29779888kB, file-rss:59608kB, shmem-rss:384kB, UID:1000 pgtables:147036kB oom_score_adj:100
The LTX python3 backend was using approximately 28.4 GiB of anonymous resident RAM when it was killed.
Version difference
LTX Desktop 1.2.5
The same host-memory exhaustion occurred.
The system could become effectively unresponsive and could require a reboot.
LTX Desktop 1.2.6
The generation still fails because host RAM and swap are exhausted.
However, the system now remains responsive after the LTX backend is killed, so a reboot is no longer required.
Models tested
The problem was reproduced with:
- LTX 2.5 Fast
- LTX 2.3 Fast
Comparison with older version
LTX Desktop 1.0.5 with LTX 2.3 rev1.1 worked on the same machine.
Expected behavior
The machine meets the documented Linux local-generation requirements:
- NVIDIA CUDA GPU
- 16 GB VRAM
- 32 GB RAM
Local generation should either complete successfully or fail gracefully without exhausting all host RAM and swap.
This does not appear to be a VRAM-capacity problem because most of the 16 GB VRAM remains free when the global OOM occurs.
Questions
Summary
Local video generation on Linux exhausts system RAM and swap before normal GPU inference begins.
The problem is reproducible in both LTX Desktop 1.2.5 and 1.2.6.
It occurs with both LTX 2.3 Fast and LTX 2.5 Fast.
The GPU still has almost all VRAM available when the system runs out of host memory.
Environment
Reproduction
Observed behavior
During model loading:
Kernel log:
The LTX python3 backend was using approximately 28.4 GiB of anonymous resident RAM when it was killed.
Version difference
LTX Desktop 1.2.5
The same host-memory exhaustion occurred.
The system could become effectively unresponsive and could require a reboot.
LTX Desktop 1.2.6
The generation still fails because host RAM and swap are exhausted.
However, the system now remains responsive after the LTX backend is killed, so a reboot is no longer required.
Models tested
The problem was reproduced with:
Comparison with older version
LTX Desktop 1.0.5 with LTX 2.3 rev1.1 worked on the same machine.
Expected behavior
The machine meets the documented Linux local-generation requirements:
Local generation should either complete successfully or fail gracefully without exhausting all host RAM and swap.
This does not appear to be a VRAM-capacity problem because most of the 16 GB VRAM remains free when the global OOM occurs.
Questions