#Unable to use Invoke offline with Anima

1 messages · Page 1 of 1 (latest)

pastel lily
#

TL;DR I am unable to use Invoke offline with Anima.

I am using the latest release cut 2 of Invokeai. I have all the relevant models downloaded for Anima (Anima model, qwen_image_vae, t5_base_encoder, and qwen_3_06b_base). The generation works perfectly fine with it when I am connected to the internet. However, if I disconnect from the internet and start a new Invoke session completely offline, I get an error for the qwen_3_06b_base: "Warning --> Unknown Qwen3 variant: embed_hidden_size=1024, layers=28. Attempting to detect configuration from weights...

(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen3-4B/resolve/main/tokenizer_config.json"...

Note that if I am connected to the internet for the initial image generation and then disconnect there is no issue. The issue only happens when I close Invoke and open a new session. It seems to try to locate the tokenizer_config.json which is not included when the model is downloaded by itself and trying to contact Hugging face for it. I was trying to download the tokenizer_config.json to try and get it to look at it locally but unsure where to put it/how to have it look at it locally. Any way for me to get around this issue?

finite bobcat
#

You can download that file with hf to have it in Cache to be used by invoke ich (i think this can work, not tested)

pastel lily
#

I tried downloading it via huggingface in the model manager by pasting Qwen/Qwen3-4B as per (https://huggingface.co/Qwen/Qwen3-4B/tree/main), but because the name of the safetensors is not Qwen3-0.6B-Base it doesn't recognize it. I tried renaming it and using the model manager to explicitly point to that specific file and it recognizes it then but I think it doesn't see the rest of the downloaded files in the folder with it so it again tries to reach back to huggingface.

finite bobcat
#

I mean you need to download the tokenizer for qwen3-4b with hf

#

Try to download Qwen/Qwen3.5-4B::tokenizer_config.json

pastel lily
#

Sorry, I am a bit new to this but did you mean manually or through the model manager by pasting "Qwen/Qwen3.5-4B::tokenizer_config.json"?

finite bobcat
#

that should not work with the UI. give me a sec

pastel lily
#

Currently I have this in my models folder after downloading via model manager

image.png
finite bobcat
#

ok you need to open the dev console in the launcher and paste this: hf download Qwen/Qwen3-4B tokenizer_config.json

#

that should download that file into the hf cache

pastel lily
#

Sweet, let me try that now

pastel lily
#

I ran the code, it added this to the following path: .cache/huggingface/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c/tokenizer_config.json. Then I downloaded the qwen_3_06b_base.safetensors separately manually from circlestone-labs/Anima (as the invoke model manager doesn't let me redownload it from the starter model because it assumes I already have it even though I had already deleted it before). I then tried it offline and get the same error. Do I need to download the safetensors directly from Qwen/Qwen3-4B for it to work? If it does, then Qwen/Qwen3-4B has 3 safetensor files and I don't know which one to download

finite bobcat
pastel lily
#

In that case that is indeed the same model that I downloaded separately after running the command in the dev console. After downloading that file I used URL/Local path in the model manager to point to that file and it recongized it. I then tried doing the generation offline and get the same trying to reach huggingface error

finite bobcat
#

i may take a look at that or maybe @zinc ruin can take a look at that.

pastel lily
#

thanks for your help!

zinc ruin