Instructions to use kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK")
model = AutoModelForMultimodalLM.from_pretrained("kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK

SGLang

How to use kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK with Docker Model Runner:
```
docker model run hf.co/kaivoss/diffusiongemma-26B-A4B-it-ALWAYS-THINK
```

Correct the README: the finetune works, and the <|think|> caveat was wrong

by Akicou - opened Sep 20

base: refs/heads/main

←

from: refs/pr/4

Discussion Files changed

+44

-32

Akicou

Sep 20

Supersedes PR #3, which had two errors.

It said the <|think|> marker is 'necessary but not sufficient' because the base model still produced an empty thought block every time. That measurement came from a pod running the STOCK template, with no marker present - a no-marker result attributed to the marker. Measured properly, 14 held-out prompts each: stock template 0/14 non-empty thought blocks, this template 14/14.
It said the sriq finetune failed and blamed the architecture. The real cause was a missing loss term. NeMo Automodel's recipe trains total = diffusion_CE + encoder_loss_weight * encoder_AR_CE; we had only the diffusion half, plus a bug that applied the logit softcap twice (forward already softcaps). With both fixed, 800 steps at r=16/alpha=32/lr=1.5e-4 took mean CJK of generation from 0.0% to 26.6%, with all 14 of 14 rows producing Chinese where base produced none.

Also added the gotchas that cost the most time: do not merge the adapter (Gemma4ClippableLinear), keep the decoder block-causal via decoder_attention_mask, and transformers 5.11 does not expose encoder_last_hidden_state from the top-level forward.

Correct the README: the finetune works, and the <|think|> caveat was wrong819678a5

kaivoss changed pull request status to merged Sep 20

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment