Instructions to use AlonMal/granite-4.1-3b-sr with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AlonMal/granite-4.1-3b-sr with Transformers:

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

pipe = pipeline("text-generation", model="AlonMal/granite-4.1-3b-sr")
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("AlonMal/granite-4.1-3b-sr", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AlonMal/granite-4.1-3b-sr with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AlonMal/granite-4.1-3b-sr"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AlonMal/granite-4.1-3b-sr",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AlonMal/granite-4.1-3b-sr

SGLang

How to use AlonMal/granite-4.1-3b-sr 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 "AlonMal/granite-4.1-3b-sr" \
    --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": "AlonMal/granite-4.1-3b-sr",
		"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 "AlonMal/granite-4.1-3b-sr" \
        --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": "AlonMal/granite-4.1-3b-sr",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AlonMal/granite-4.1-3b-sr with Docker Model Runner:
```
docker model run hf.co/AlonMal/granite-4.1-3b-sr
```

granite-4.1-3b-sr

ibm-granite/granite-4.1-3b converted to the Shadow Residual (SR) architecture — the base model for training SR adapters with stock PEFT LoRA.

SR runs two streams through every decoder layer: a frozen base stream (the unadapted base model) and a trainable adapter stream (LoRA deltas plus a per-layer low-rank cross-stream injection from the base). The weights here are the base model's own weights, re-laid-out for SR (fused projections split into separate q/k/v and gate/up projections). With no adapter attached the model produces the base model's outputs.

Usage

Loading needs the shadow-residual package installed (it registers the shadow_residual model type with transformers; no trust_remote_code).

import shadow_residual.shadow_residual  # registers the SR model type
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AlonMal/granite-4.1-3b-sr")
tokenizer = AutoTokenizer.from_pretrained("AlonMal/granite-4.1-3b-sr")

Attach a trained SR adapter with stock PEFT:

from peft import PeftModel

model = PeftModel.from_pretrained(model, "<adapter>")

The frozen cross-stream sites are all-zero by construction and are not stored; they are re-created on load.

Downloads last month: 343

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for AlonMal/granite-4.1-3b-sr

Base model

ibm-granite/granite-4.1-3b

Finetuned

(23)

this model