Out of VRAM with SD 3.5 Large FP8 scaled, but no issue in ComfyUI

[p-try]

Hello erverybody,

I wonder if somebody here may have an idea about a little problem.

I have an AMD Radeon RX 6800 XT with 16 GB of VRAM and I'm trying to set up image generation using SD 3.5. I am using this model:
https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/sd3.5_large_fp8_scaled.safetensors

This is the model configuration:

name: sd3.5-large-fp8-scaled-safetensors
backend: stablediffusion-ggml
parameters:
  model: sd3.5_large_fp8_scaled.safetensors
step: 20
cfg_scale: 4.5
low_vram: true
options:
- "clip_l_path:clip_l.safetensors"
- "clip_g_path:clip_g.safetensors"
- "t5xxl_path:t5xxl_fp8_e4m3fn_scaled.safetensors"
- "sampler:euler"

LocalAI reports that it is out of VRAM:

localai-1  | 12:26PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr ggml_cuda_init: found 1 ROCm devices:
localai-1  | 12:26PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr   Device 0: AMD Radeon RX 6800 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32
localai-1  | 12:26PM INF Success ip=127.0.0.1 latency="33.8µs" method=GET status=200 url=/readyz
localai-1  | 12:27PM INF Success ip=127.0.0.1 latency="86.485µs" method=GET status=200 url=/readyz
localai-1  | 12:27PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr Created context: OK
localai-1  | 12:27PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr Generating image
localai-1  | 12:27PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr ggml_backend_cuda_buffer_type_alloc_buffer: allocating 1664.00 MiB on device 0: cudaMalloc failed: out of memory
localai-1  | 12:27PM DBG GRPC(sd3.5-large-fp8-scaled-safetensors-127.0.0.1:33845): stderr ggml_gallocr_reserve_n: failed to allocate ROCm0 buffer of size 1744830464

However, when I generate an image using the same model files in ComfyUI, everything works just fine (and quite fast). It DOES complain about VRAM but only during VAE decoding:

got prompt
Using scaled fp8: fp8 matrix mult: False, scale input: True
model weight dtype torch.float16, manual cast: None
model_type FLOW
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
Using scaled fp8: fp8 matrix mult: False, scale input: False
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load SD3ClipModel_
loaded completely 15122.8 6228.190093994141 True
/root/ComfyUI/comfy/ops.py:80: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:331.)
  return torch.nn.functional.linear(input, weight, bias)
Requested to load SD3
loaded completely 11615.620722656251 7683.561706542969 True
100%|██████████| 20/20 [01:30<00:00,  4.51s/it]
Requested to load AutoencodingEngine
loaded completely 3787.8148437500004 319.7467155456543 True
Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.
Prompt executed in 166.86 seconds

Now I'm wondering:

• Maybe the model only narrowly fits into VRAM and the LocalAI stack has a little more overhead?
• Maybe the model does not fit into VRAM but ComfyUI manages to offload some to the regular RAM which LocalAI doesn't?
• Something else?

Can somebody point me in the right direction?

Thank you very much in advance!