Instructions to use IFM/K2-Horizon-0.9B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use IFM/K2-Horizon-0.9B with Transformers:

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

pipe = pipeline("text-generation", model="IFM/K2-Horizon-0.9B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("IFM/K2-Horizon-0.9B", trust_remote_code=True, device_map="auto")

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use IFM/K2-Horizon-0.9B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "IFM/K2-Horizon-0.9B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "IFM/K2-Horizon-0.9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/IFM/K2-Horizon-0.9B

SGLang

How to use IFM/K2-Horizon-0.9B 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 "IFM/K2-Horizon-0.9B" \
    --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": "IFM/K2-Horizon-0.9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "IFM/K2-Horizon-0.9B" \
        --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": "IFM/K2-Horizon-0.9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use IFM/K2-Horizon-0.9B with Docker Model Runner:
```
docker model run hf.co/IFM/K2-Horizon-0.9B
```

[superseded by #8 - opened as a discussion by mistake] Fix output_hidden_states / attentions returning None

by Avicennasis - opened Sep 30

Discussion

Avicennasis

Sep 30

Follow-up to #6 (full write-up, evidence and repro there).

Commit 1: K2HorizonModel.forward declares the hidden_states / attentions entries in _can_record_outputs, but never had the @capture_outputs decorator that reads the output_* flags and installs the capture hooks — so output_hidden_states=True returned None. Adds the import + decorator, mirroring stock Qwen3MoeModel.forward.
Commit 2: the attentions entry names K2HorizonAttention, but these checkpoints instantiate K2HorizonMoVAAttention, so output_attentions=True returned an empty tuple. Lists both classes (weights are produced under eager attention, as with stock models).

Verified on K2-Horizon-0.9B, transformers 5.17.0: as shipped → hidden_states=None; with this branch → 29 states for 28 layers ([0] == embeddings, [-1] == last_hidden_state), flag-off → None, greedy generation unchanged; attentions → 28 tensors with eager attention. The modeling file is byte-identical across all 12 full-model K2-Horizon repos, so the fix applies family-wide (tested here on the 0.9B).

Avicennasis changed discussion title from Fix `output_hidden_states` / `attentions` returning None (missing `@capture_outputs`) to [superseded by #8 - opened as a discussion by mistake] Fix output_hidden_states / attentions returning None Sep 30

Avicennasis

Sep 30

Opened this by mistake - the API flag I used created a plain discussion instead of a PR, and it cannot be converted after the fact. The real pull request with both commits is #8; background and evidence in #6.

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