Checked 24 September 2026 · vLLM-Omni H3 capacity check
Two RTX 5090s can run full H3. The documented job still takes 8:38 for five seconds of video.
The update: vLLM-Omni says one 50-step H3 text-to-video-and-audio request completed on two RTX 5090 GPUs at 1344×768: 124 frames at 24 fps, about 5.17 seconds of output, returned in 8 minutes 38 seconds of client end-to-end time. This is a full-H3 capacity result, not a live-video route.
What the project actually validated
Project says: at vLLM-Omni commit ae6577ea, one 50-step T2VA request completed without an out-of-memory failure on two RTX 5090 GPUs. The source reports a valid H.264 video with 32 kHz stereo AAC, a full ffmpeg decode check, and an approximate sampled peak of 22.6 GiB per GPU.
The source calls this a single end-to-end validation, not a warmed multi-run latency benchmark. Its GPU reading is a sampled nvidia-smi peak, not a CUDA allocator maximum. Treat the result as a documented setup lead, not a general 5090 benchmark.
The clock makes this an offline route
124 frames at 24 fps equal about 5.17 seconds of media. Eight minutes 38 seconds equal 518 seconds of client end-to-end time. That is about 100 seconds of wait for each output second.
This can be useful for a controlled full-H3 render when the GPU and host are already available. It does not establish a warm response, first playable media, first visible frame, queue behavior, concurrent capacity, retries, stream delivery, interactive response, or a continuous channel.
Two GPUs do not remove the host requirement
Project says: its two-GPU distributed-layerwise-offload path needs at least 200 GiB of available system RAM; it recommends 384 GiB. The recipe also says each FL2VA or Ref2VA checkpoint partition is about 134 GiB in BF16 safetensors, roughly 135 GiB on disk, and warns against running both partitions together on a host only sized to the minimum.
The documented RTX 5090 profile starts at 1344×768 with 20 resident DiT blocks. The listed two-RTX-4090 profile is a capacity-proxy starting point at 1024×576, not a validated 4090 timing result. GPU VRAM is only one part of this setup.
Do not transfer this row to a transition or a FastH3 headline
The measured target workload is T2VA. It does not prove a two-5090 FL2VA or Ref2VA run, including a first/last-frame transition. It also cannot be compared directly with vLLM-Omni's separate warmed eight-B300 FastH3 result: that is a four-forward dedicated FastH3 T2VA profile with a different timing boundary and hardware class.
Smallest honest next test
If you are authorised to try this route, reproduce the documented T2VA control before changing the task. Record model and software revisions, GPU topology, host RAM, free storage, DLO setting, cold ready time, request acceptance, complete MP4, decoded media, and first visible frame. Then test first/last-frame conditioning separately. A valid T2VA MP4 does not prove a frame-anchored workflow.