Instructions to use mshapiro123/recurrent-qwen2.5-0.5b-full-block with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mshapiro123/recurrent-qwen2.5-0.5b-full-block with Transformers:

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

pipe = pipeline("text-generation", model="mshapiro123/recurrent-qwen2.5-0.5b-full-block", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("mshapiro123/recurrent-qwen2.5-0.5b-full-block", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mshapiro123/recurrent-qwen2.5-0.5b-full-block with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mshapiro123/recurrent-qwen2.5-0.5b-full-block"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mshapiro123/recurrent-qwen2.5-0.5b-full-block",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/mshapiro123/recurrent-qwen2.5-0.5b-full-block

SGLang

How to use mshapiro123/recurrent-qwen2.5-0.5b-full-block 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 "mshapiro123/recurrent-qwen2.5-0.5b-full-block" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mshapiro123/recurrent-qwen2.5-0.5b-full-block",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "mshapiro123/recurrent-qwen2.5-0.5b-full-block" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mshapiro123/recurrent-qwen2.5-0.5b-full-block",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use mshapiro123/recurrent-qwen2.5-0.5b-full-block with Docker Model Runner:
```
docker model run hf.co/mshapiro123/recurrent-qwen2.5-0.5b-full-block
```

mshapiro123 commited on Aug 16

Commit

fb37f02

verified ·

1 Parent(s): 5795c01

Render architecture figure in model card

Browse files

Files changed (3) hide show

.gitattributes +1 -0
README.md +1 -1
figure1_architecture_comparison.png +3 -0

.gitattributes CHANGED Viewed

@@ -3,3 +3,4 @@
 *.md text eol=lf
 *.py text eol=lf
 LICENSE text eol=lf

 *.md text eol=lf
 *.py text eol=lf
 LICENSE text eol=lf
+figure1_architecture_comparison.png filter=lfs diff=lfs merge=lfs -text

README.md CHANGED Viewed

@@ -15,7 +15,7 @@ The full-block recurrent-depth retrofit of Qwen2.5-0.5B-Instruct: the deep-chara
 ## Architecture
-![Architecture comparison from the paper](./figure1_architecture_comparison.svg)
 The 24-layer base model is split into a Prelude (layers 0-5), a weight-tied Recurrent Block (layers 6-17), and a Coda (layers 18-23). The block executes T times per pass, and a trained split re-entry bridge combines the carried state with the re-injected Prelude output under an identity-biased gate on loops 2 through T. At T = 1 the recurrent additions are bypassed and the model reproduces the base computation exactly. Trained parameters: the 12-layer block plus the bridge, 180,556,929 forward-active. The recurrent wrapper is custom architecture code: load with `trust_remote_code=True`, using the modeling code shipped in this repository.

 ## Architecture
+![Architecture comparison from the paper](./figure1_architecture_comparison.png)
 The 24-layer base model is split into a Prelude (layers 0-5), a weight-tied Recurrent Block (layers 6-17), and a Coda (layers 18-23). The block executes T times per pass, and a trained split re-entry bridge combines the carried state with the re-injected Prelude output under an identity-biased gate on loops 2 through T. At T = 1 the recurrent additions are bypassed and the model reproduces the base computation exactly. Trained parameters: the 12-layer block plus the bridge, 180,556,929 forward-active. The recurrent wrapper is custom architecture code: load with `trust_remote_code=True`, using the modeling code shipped in this repository.

figure1_architecture_comparison.png ADDED Viewed

Git LFS Details

SHA256: 5a17f61de76ea49395c5d5077dd5723675a79862752dc31cd0c6f5785b84ab27
Pointer size: 131 Bytes
Size of remote file: 559 kB