problem loading koboldcpp-oldpc pony xl and ernie

[chk777]

Hi! My kobold version 1.114.1 - pony_diffusion_v6_x model won't load. When loading, I specify the L and G encoders, but the program resets them. I also can't load the model.
ernie-image-turbo model It asks me to specify a clip for it, although according to the architecture, it should already be in the model. I tried adding a lightweight pony_clip_l.safetensors, but it didn't work. Please help.

[wbruna]

ernie-image-turbo model It asks me to specify a clip for it, although according to the architecture, it should already be in the model. I tried adding a lightweight pony_clip_l.safetensors, but it didn't work. Please help.

ERNIE does need a text encoder: Ministral-3-3B-Instruct-2512. It also needs the Flux 2 VAE.

Could you tell us from where you got the pony_diffusion_v6_x, clip_l and clip_g files?

[chk777]

Не могли бы вы сказать, откуда вы взяли файлы pony_diffusion_v6_x, clip_l и clip_g?
from here https://huggingface.co/morikomorizz/Pony-Diffusion-V6-XL-GGUF/
Thanks for the tip about ERNIE

[chk777]

ERNIE does need a text encoder: Ministral-3-3B-Instruct-2512. It also needs the Flux 2 VAE.

but this not work on my comp. that's what writes:

Chat completion heuristic: Mistral Tekken

ImageGen Init - Load Model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf
With Custom VAE: /home/models/ernie/diffusion_pytorch_model.safetensors
Selecting CPU as CLIP image generation device

Error: Image generation failed to setup!
Make sure you have ALL files required (e.g. VAE, T5, Clip...) or baked into the model!
Load Image Model OK: False

Could not load image model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf"

[wbruna]

Chat completion heuristic: Mistral Tekken

Image generation is completely separate from LLM generation: you need to specify the text model with --sdclip1 (or the Image screen in the launcher).

[chk777]

--sdclip1 - this command is ignored, that write in the log "Selecting CPU as CLIP image generation device", on Pony xl, on Erny, on Z-image...

[LostRuins]

Set --sdclipdevice

[chk777]

Set --sdclipdevice

not working, this one write

ImageGen Init - Load Model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf
With Custom VAE: /home/models/ernie/diffusion_pytorch_model.safetensors
Offloading weights to system RAM
Selecting CPU as CLIP image generation device

[LostRuins]

Please share your launch commands and full log

[chk777]

CUDA_VISIBLE_DEVICES=6,7 ~/koboldcpp/koboldcpp-linux-x64-oldpc --skiplauncher --port 5002 --sdoffloadcpu --model /home/models/ernie/Ministral-3-3B-Instruct-2512-Q4_K_M.gguf --sdmodel /home/models/ernie/ernie-image-turbo-Q4_0.gguf --sdvae /home/models/ernie/diffusion_pytorch_model.safetensors --sdthreads 4 --usecuda all --lowvram

This model does not use SWA
Automatic RoPE Scaling: Using model internal value.
llama_context: constructing llama_context
llama_context: setting new yarn_attn_factor = 1.0000 (mscale == 1.0, mscale_all_dim = 1.0)
llama_context: n_seq_max = 1
llama_context: n_ctx = 8448
llama_context: n_ctx_seq = 8448
llama_context: n_batch = 1024
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = enabled
llama_context: kv_unified = true
llama_context: freq_base = 1000000.0
llama_context: freq_scale = 0.0625
llama_context: n_rs_seq = 0
llama_context: n_ctx_seq (8448) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
set_abort_callback: call
llama_context: CUDA_Host output buffer size = 0.50 MiB
llama_kv_cache: CPU KV buffer size = 858.00 MiB
llama_kv_cache: size = 858.00 MiB ( 8448 cells, 26 layers, 1/1 seqs), K (f16): 429.00 MiB, V (f16): 429.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 128
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 128
llama_context: enumerating backends
llama_context: backend_ptrs.size() = 3
sched_reserve: reserving ...
sched_reserve: max_nodes = 1896
sched_reserve: reserving full memory module
sched_reserve: worst-case: n_tokens = 512, n_seqs = 1, n_outputs = 1
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve: CUDA0 compute buffer size = 82.25 MiB
sched_reserve: CUDA1 compute buffer size = 262.00 MiB
sched_reserve: CUDA_Host compute buffer size = 20.27 MiB
sched_reserve: graph nodes = 838
sched_reserve: graph splits = 55
sched_reserve: reserve took 15.11 ms, sched copies = 1
Threadpool set to 3 threads and 3 blasthreads...
attach_threadpool: call
Starting model warm up, please wait a moment...
Load Text Model OK: True
Chat completion heuristic: Mistral Tekken

ImageGen Init - Load Model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf
With Custom VAE: /home/models/ernie/diffusion_pytorch_model.safetensors
Offloading weights to system RAM
Selecting CPU as CLIP image generation device

Error: Image generation failed to setup!
Make sure you have ALL files required (e.g. VAE, T5, Clip...) or baked into the model!
Load Image Model OK: False

Could not load image model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf

[chk777]

I've already tried many different methods and models. I need to run at least Pony XL. But of course I want Ernie with its new architecture.

[wbruna]

@chk777 , as I've commented before, --model is for LLM usage; you need to pass Ministral in --sdclip1 instead. Try:

CUDA_VISIBLE_DEVICES=6,7 ~/koboldcpp/koboldcpp-linux-x64-oldpc --skiplauncher --port 5002 --sdoffloadcpu --sdclip1 /home/ppk2/models/ernie/Ministral-3-3B-Instruct-2512-Q4_K_M.gguf --sdmodel /home/models/ernie/ernie-image-turbo-Q4_0.gguf --sdvae /home/models/ernie/diffusion_pytorch_model.safetensors --sdthreads 4 --usecuda all

About Pony XL: I suggest you avoid that quant at first, it may have been generated in a way that's not compatible with our backend. How much memory do you have in your cards? Not in total, each one of them?

[chk777]

8 gb old vram pascal

[chk777]

is for LLM usage; you need to pass Ministral in --sdclip1 instead.

I also tried this parameter. It's about the same.

[wbruna]

is for LLM usage; you need to pass Ministral in --sdclip1 instead.

I also tried this parameter. It's about the same.

The log should've been completely different, unless you've kept --model too (you shouldn't, at least not on the same GPU). Is it now printing the model on loading, like this?

ImageGen Init - Load Model: /home/models/ernie/ernie-image-turbo-Q4_0.gguf
With Custom VAE: /home/models/ernie/diffusion_pytorch_model.safetensors
With Custom Clip-1 Model: /home/models/ernie/Ministral-3-3B-Instruct-2512-Q4_K_M.gguf <== this line

(that last path is also slightly different than before: I've copied your original path with the ppk2 directory on my suggestion, so double check if it's the correct one)

Could you post the command-line args + launching log with --debug?

[chk777]

Thank you very much! Im happy! All worked!