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

Libraries

How to use mshapiro123/recurrent-qwen2.5-0.5b-natural-keeper 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-natural-keeper", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("mshapiro123/recurrent-qwen2.5-0.5b-natural-keeper", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mshapiro123/recurrent-qwen2.5-0.5b-natural-keeper 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-natural-keeper"
# 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-natural-keeper",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/mshapiro123/recurrent-qwen2.5-0.5b-natural-keeper

SGLang

How to use mshapiro123/recurrent-qwen2.5-0.5b-natural-keeper 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-natural-keeper" \
    --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-natural-keeper",
		"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-natural-keeper" \
        --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-natural-keeper",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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

recurrent-qwen2.5-0.5b-natural-keeper / figure1_architecture_comparison.svg

mshapiro123

Add paper links and architecture figure

7cc6097 verified about 2 months ago

preview code

Raw History Blame Contribute Delete

17.2 kB

	<?xml version="1.0" encoding="UTF-8"?>
	<svg viewBox="0 0 1280 900" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Paper One architecture comparison diagram">
	<defs>
	<marker id="arr" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
	<path d="M0,0 L10,5 L0,10 z" fill="#4B5563"/>
	</marker>
	<marker id="arrB" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
	<path d="M0,0 L10,5 L0,10 z" fill="#2563EB"/>
	</marker>
	<marker id="arrA" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
	<path d="M0,0 L10,5 L0,10 z" fill="#B45309"/>
	</marker>
	</defs>

	<!-- ===================== HEADER ===================== -->
	<text x="30" y="34" font-size="19" font-weight="700" fill="#111827">The base model and its recurrent retrofits</text>
	<text x="30" y="56" font-size="12" fill="#6B7280">Qwen2.5-0.5B-Instruct · split at layers 6 and 18 · forced depth: loops = task depth (no learned halting)</text>

	<!-- legend -->
	<rect x="880" y="24" width="14" height="10" fill="#E8E8E6" stroke="#7A7A75"/>
	<text x="900" y="33" font-size="9.5" fill="#4B5563">pretrained, unchanged</text>
	<rect x="880" y="40" width="14" height="10" fill="#EFF4FF" stroke="#2563EB"/>
	<text x="900" y="49" font-size="9.5" fill="#4B5563">trained — full-block budget (180.6M)</text>
	<rect x="880" y="56" width="14" height="10" fill="#FEF3E2" stroke="#B45309"/>
	<text x="900" y="65" font-size="9.5" fill="#4B5563">trained — adapter budget (6.01M, base frozen)</text>

	<!-- ===================== PANEL 1: BASE MODEL ===================== -->
	<rect x="20" y="80" width="400" height="540" rx="10" fill="#F6F6F3" stroke="#E2E2DE"/>
	<text x="36" y="108" font-size="13.5" font-weight="700" fill="#111827">Base model — dense</text>

	<rect x="70" y="130" width="300" height="36" rx="6" fill="#FFFFFF" stroke="#9CA3AF"/>
	<text x="220" y="152" font-size="10.5" fill="#1F2937" text-anchor="middle">tokens</text>
	<line x1="220" y1="166" x2="220" y2="188" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="70" y="190" width="300" height="248" rx="6" fill="#E8E8E6" stroke="#7A7A75" stroke-width="1.5"/>
	<text x="220" y="288" font-size="11" fill="#333" text-anchor="middle" font-weight="600">24 decoder layers — 0 to 23</text>
	<text x="220" y="306" font-size="9.5" fill="#6B7280" text-anchor="middle">executed once, in order</text>
	<text x="220" y="322" font-size="9.5" fill="#6B7280" text-anchor="middle">depth fixed by the architecture</text>
	<text x="220" y="352" font-size="8.5" fill="#6B7280" text-anchor="middle">24 × 14.91M = 357.9M layer parameters</text>
	<text x="220" y="366" font-size="8.5" fill="#6B7280" text-anchor="middle">tied embedding, LM head, and norms: 136.1M</text>
	<text x="220" y="380" font-size="8.5" fill="#6B7280" text-anchor="middle">494.0M unique parameters</text>
	<line x1="220" y1="438" x2="220" y2="452" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="70" y="454" width="300" height="40" rx="6" fill="#FFFFFF" stroke="#9CA3AF"/>
	<text x="220" y="478" font-size="10.5" fill="#1F2937" text-anchor="middle">LM head — token logits</text>

	<text x="220" y="530" font-size="10" fill="#374151" text-anchor="middle" font-weight="600">more computation requires more tokens</text>
	<text x="220" y="548" font-size="9.5" fill="#6B7280" text-anchor="middle">(the scratchpad recipe serializes its reasoning here)</text>

	<!-- ===================== PANEL 2: FULL-BLOCK RETROFIT ===================== -->
	<rect x="440" y="80" width="400" height="540" rx="10" fill="#F6F6F3" stroke="#E2E2DE"/>
	<text x="456" y="108" font-size="13.5" font-weight="700" fill="#111827">Full-block retrofit</text>

	<rect x="470" y="130" width="230" height="36" rx="6" fill="#FFFFFF" stroke="#9CA3AF"/>
	<text x="585" y="152" font-size="10.5" fill="#1F2937" text-anchor="middle">tokens</text>
	<line x1="585" y1="166" x2="585" y2="188" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="470" y="190" width="230" height="52" rx="6" fill="#E8E8E6" stroke="#7A7A75"/>
	<text x="585" y="207" font-size="10.5" fill="#333" text-anchor="middle" font-weight="600">Prelude — layers 0–5</text>
	<text x="585" y="221" font-size="8.5" fill="#6B7280" text-anchor="middle">pretrained, unchanged → output p</text>
	<text x="585" y="235" font-size="8.5" fill="#6B7280" text-anchor="middle">6 layers · 89.5M</text>
	<line x1="585" y1="242" x2="585" y2="250" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<!-- p re-injection path into bridge -->
	<path d="M700,212 L770,212 L770,254" fill="none" stroke="#2563EB" stroke-width="1.3" marker-end="url(#arrB)"/>
	<text x="736" y="205" font-size="8.5" fill="#1E40AF" text-anchor="middle">p, re-injected</text>

	<rect x="470" y="252" width="230" height="116" rx="6" fill="#EFF4FF" stroke="#2563EB" stroke-width="1.5"/>
	<text x="585" y="274" font-size="10.5" fill="#1E40AF" text-anchor="middle" font-weight="600">Recurrent Block — layers 6–17</text>
	<text x="585" y="293" font-size="9" fill="#1E40AF" text-anchor="middle">12 layers, weight-tied · 12 × 14.91M = 178.9M</text>
	<text x="585" y="310" font-size="9" fill="#1E40AF" text-anchor="middle">every block weight trained</text>
	<text x="585" y="340" font-size="8.5" fill="#2563EB" text-anchor="middle">loop 1 input is p · later loops input u from the bridge</text>

	<!-- feedback loop through the bridge -->
	<line x1="700" y1="330" x2="714" y2="330" stroke="#2563EB" stroke-width="1.4" marker-end="url(#arrB)"/>
	<text x="707" y="324" font-size="8" fill="#1E40AF" text-anchor="middle">h</text>
	<rect x="716" y="256" width="108" height="100" rx="6" fill="#EFF4FF" stroke="#2563EB" stroke-width="1.5"/>
	<text x="770" y="274" font-size="8.5" fill="#1E40AF" text-anchor="middle" font-weight="700">split re-entry bridge</text>
	<text x="770" y="292" font-size="8.5" fill="#1E40AF" text-anchor="middle">W<tspan dy="2" font-size="6.5">p</tspan><tspan dy="-2">·p + W</tspan><tspan dy="2" font-size="6.5">s</tspan><tspan dy="-2">·h</tspan></text>
	<text x="770" y="308" font-size="8.5" fill="#1E40AF" text-anchor="middle">identity-biased gate</text>
	<text x="770" y="324" font-size="8" fill="#2563EB" text-anchor="middle">→ u, the next loop's input</text>
	<text x="770" y="344" font-size="8.5" fill="#1E40AF" text-anchor="middle" font-weight="600">1.61M trained</text>
	<line x1="716" y1="272" x2="702" y2="272" stroke="#2563EB" stroke-width="1.4" marker-end="url(#arrB)"/>
	<text x="709" y="264" font-size="8" fill="#1E40AF" text-anchor="middle">u</text>
	<text x="770" y="370" font-size="8.5" fill="#6B7280" text-anchor="middle">loops 2…T</text>

	<line x1="585" y1="368" x2="585" y2="382" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="470" y="384" width="230" height="52" rx="6" fill="#E8E8E6" stroke="#7A7A75"/>
	<text x="585" y="401" font-size="10.5" fill="#333" text-anchor="middle" font-weight="600">Coda — layers 18–23</text>
	<text x="585" y="415" font-size="8.5" fill="#6B7280" text-anchor="middle">pretrained, unchanged</text>
	<text x="585" y="429" font-size="8.5" fill="#6B7280" text-anchor="middle">6 layers · 89.5M</text>
	<line x1="585" y1="436" x2="585" y2="450" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="470" y="452" width="230" height="40" rx="6" fill="#EFF4FF" stroke="#2563EB" stroke-width="1.5"/>
	<text x="585" y="476" font-size="10" fill="#1E40AF" text-anchor="middle" font-weight="600">decoded state at loop T — the answer</text>

	<text x="640" y="530" font-size="10" fill="#374151" text-anchor="middle" font-weight="600">trained: block 178.9M + bridge 1.61M = 180.6M forward-active</text>
	<text x="640" y="548" font-size="9.5" fill="#6B7280" text-anchor="middle">T = 1 bypasses the bridge and recurrent additions entirely,</text>
	<text x="640" y="562" font-size="9.5" fill="#6B7280" text-anchor="middle">reproducing the base computation exactly</text>

	<!-- ===================== PANEL 3: ADAPTER RETROFIT ===================== -->
	<rect x="860" y="80" width="400" height="540" rx="10" fill="#F6F6F3" stroke="#E2E2DE"/>
	<text x="876" y="108" font-size="13.5" font-weight="700" fill="#111827">Adapter retrofit — same surgery</text>

	<rect x="890" y="130" width="230" height="36" rx="6" fill="#FFFFFF" stroke="#9CA3AF"/>
	<text x="1005" y="152" font-size="10.5" fill="#1F2937" text-anchor="middle">tokens</text>
	<line x1="1005" y1="166" x2="1005" y2="188" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="890" y="190" width="230" height="52" rx="6" fill="#E8E8E6" stroke="#7A7A75"/>
	<text x="1005" y="207" font-size="10.5" fill="#333" text-anchor="middle" font-weight="600">Prelude — layers 0–5</text>
	<text x="1005" y="221" font-size="8.5" fill="#6B7280" text-anchor="middle">frozen → output p</text>
	<text x="1005" y="235" font-size="8.5" fill="#6B7280" text-anchor="middle">6 layers · 89.5M</text>
	<line x1="1005" y1="242" x2="1005" y2="250" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<!-- p re-injection path into bridge -->
	<path d="M1120,212 L1190,212 L1190,254" fill="none" stroke="#B45309" stroke-width="1.3" marker-end="url(#arrA)"/>
	<text x="1156" y="205" font-size="8.5" fill="#7C3E06" text-anchor="middle">p, re-injected</text>

	<rect x="890" y="252" width="230" height="116" rx="6" fill="#E8E8E6" stroke="#7A7A75" stroke-width="1.5"/>
	<text x="1005" y="270" font-size="10.5" fill="#333" text-anchor="middle" font-weight="600">Recurrent Block — layers 6–17</text>
	<text x="1005" y="285" font-size="8.5" fill="#6B7280" text-anchor="middle">12 layers, weight-tied, frozen — 178.9M</text>
	<!-- LoRA composition mini-diagram -->
	<rect x="906" y="294" width="88" height="18" rx="3" fill="#FFFFFF" stroke="#7A7A75"/>
	<text x="950" y="307" font-size="8" fill="#4B5563" text-anchor="middle">W — frozen</text>
	<rect x="906" y="318" width="112" height="18" rx="3" fill="#FEF3E2" stroke="#B45309"/>
	<text x="962" y="331" font-size="8" fill="#7C3E06" text-anchor="middle">ΔW = A·B — rank 16, trained</text>
	<line x1="994" y1="303" x2="1043" y2="312" stroke="#4B5563" stroke-width="1.1" marker-end="url(#arr)"/>
	<line x1="1018" y1="327" x2="1043" y2="320" stroke="#B45309" stroke-width="1.1" marker-end="url(#arrA)"/>
	<circle cx="1052" cy="316" r="9" fill="#FFFFFF" stroke="#374151"/>
	<text x="1052" y="320" font-size="10" fill="#111827" text-anchor="middle" font-weight="700">+</text>
	<line x1="1061" y1="316" x2="1078" y2="316" stroke="#4B5563" stroke-width="1.1" marker-end="url(#arr)"/>
	<text x="1096" y="313" font-size="8" fill="#374151" text-anchor="middle">every block</text>
	<text x="1096" y="323" font-size="8" fill="#374151" text-anchor="middle">projection</text>
	<text x="1005" y="354" font-size="8.5" fill="#7C3E06" text-anchor="middle">ΔW ≈ 4.40M over 84 projections — applied at every loop</text>

	<!-- feedback loop through the bridge -->
	<line x1="1120" y1="330" x2="1134" y2="330" stroke="#B45309" stroke-width="1.4" marker-end="url(#arrA)"/>
	<text x="1127" y="324" font-size="8" fill="#7C3E06" text-anchor="middle">h</text>
	<rect x="1136" y="256" width="108" height="100" rx="6" fill="#FEF3E2" stroke="#B45309" stroke-width="1.5"/>
	<text x="1190" y="274" font-size="8.5" fill="#7C3E06" text-anchor="middle" font-weight="700">split re-entry bridge</text>
	<text x="1190" y="292" font-size="8.5" fill="#7C3E06" text-anchor="middle">W<tspan dy="2" font-size="6.5">p</tspan><tspan dy="-2">·p + W</tspan><tspan dy="2" font-size="6.5">s</tspan><tspan dy="-2">·h</tspan></text>
	<text x="1190" y="308" font-size="8.5" fill="#7C3E06" text-anchor="middle">identity-biased gate</text>
	<text x="1190" y="324" font-size="8" fill="#B45309" text-anchor="middle">→ u, the next loop's input</text>
	<text x="1190" y="344" font-size="8.5" fill="#7C3E06" text-anchor="middle" font-weight="600">1.61M trained</text>
	<line x1="1136" y1="272" x2="1122" y2="272" stroke="#B45309" stroke-width="1.4" marker-end="url(#arrA)"/>
	<text x="1129" y="264" font-size="8" fill="#7C3E06" text-anchor="middle">u</text>
	<text x="1190" y="370" font-size="8.5" fill="#6B7280" text-anchor="middle">loops 2…T</text>

	<line x1="1005" y1="368" x2="1005" y2="382" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="890" y="384" width="230" height="52" rx="6" fill="#E8E8E6" stroke="#7A7A75"/>
	<text x="1005" y="401" font-size="10.5" fill="#333" text-anchor="middle" font-weight="600">Coda — layers 18–23</text>
	<text x="1005" y="415" font-size="8.5" fill="#6B7280" text-anchor="middle">frozen</text>
	<text x="1005" y="429" font-size="8.5" fill="#6B7280" text-anchor="middle">6 layers · 89.5M</text>
	<line x1="1005" y1="436" x2="1005" y2="450" stroke="#4B5563" stroke-width="1.4" marker-end="url(#arr)"/>

	<rect x="890" y="452" width="230" height="40" rx="6" fill="#FEF3E2" stroke="#B45309" stroke-width="1.5"/>
	<text x="1005" y="476" font-size="10" fill="#7C3E06" text-anchor="middle" font-weight="600">decoded state at loop T — the answer</text>

	<text x="1060" y="530" font-size="10" fill="#374151" text-anchor="middle" font-weight="600">trained: LoRA ≈ 4.40M + bridge 1.61M = 6.01M forward-active (3.3%)</text>
	<text x="1060" y="548" font-size="9.5" fill="#6B7280" text-anchor="middle">base weights untouched — the adapter is detachable</text>
	<text x="1060" y="562" font-size="9.5" fill="#6B7280" text-anchor="middle">and recovery of the base model is guaranteed</text>

	<!-- ===================== BOTTOM STRIP: THE FIVE ARMS ===================== -->
	<text x="30" y="666" font-size="13.5" font-weight="700" fill="#111827">The five registered comparison arms (Section 9)</text>

	<rect x="20" y="682" width="232" height="140" rx="10" fill="#EFF4FF" stroke="#2563EB" stroke-width="1.3"/>
	<text x="36" y="706" font-size="11" font-weight="700" fill="#1E40AF">Arm A — recurrent 0.5B</text>
	<text x="36" y="726" font-size="9.5" fill="#1F2937">Full-block budget</text>
	<text x="36" y="742" font-size="9.5" fill="#1F2937">Latent loops at forced depth</text>
	<text x="36" y="758" font-size="9.5" fill="#1F2937">One decoded state, no generation</text>
	<text x="36" y="806" font-size="9" fill="#6B7280" font-style="italic">primary registered system</text>

	<rect x="268" y="682" width="232" height="140" rx="10" fill="#FEF3E2" stroke="#B45309" stroke-width="1.3"/>
	<text x="284" y="706" font-size="11" font-weight="700" fill="#7C3E06">Arm E — recurrent 0.5B</text>
	<text x="284" y="726" font-size="9.5" fill="#1F2937">Adapter budget, base frozen</text>
	<text x="284" y="742" font-size="9.5" fill="#1F2937">Identical surgery and protocol</text>
	<text x="284" y="758" font-size="9.5" fill="#1F2937">Tests whether 180M is required</text>
	<text x="284" y="806" font-size="9" fill="#6B7280" font-style="italic">parameter-efficiency control</text>

	<rect x="516" y="682" width="232" height="140" rx="10" fill="#FFFFFF" stroke="#9CA3AF" stroke-width="1.3"/>
	<text x="532" y="706" font-size="11" font-weight="700" fill="#374151">Arm B — dense 0.5B</text>
	<text x="532" y="726" font-size="9.5" fill="#1F2937">Direct-answer SFT</text>
	<text x="532" y="742" font-size="9.5" fill="#1F2937">No sequential computation</text>
	<text x="532" y="806" font-size="9" fill="#6B7280" font-style="italic">primary preregistered control</text>

	<rect x="764" y="682" width="232" height="140" rx="10" fill="#FFFFFF" stroke="#9CA3AF" stroke-width="1.3"/>
	<text x="780" y="706" font-size="11" font-weight="700" fill="#374151">Arm C — dense 0.5B</text>
	<text x="780" y="726" font-size="9.5" fill="#1F2937">Serialized-scratchpad SFT</text>
	<text x="780" y="742" font-size="9.5" fill="#1F2937">Writes its reasoning in tokens</text>
	<text x="780" y="758" font-size="9.5" fill="#1F2937">Token axis instead of depth axis</text>
	<text x="780" y="806" font-size="9" fill="#6B7280" font-style="italic">strongest dense control</text>

	<rect x="1012" y="682" width="232" height="140" rx="10" fill="#FFFFFF" stroke="#9CA3AF" stroke-width="1.3"/>
	<text x="1028" y="706" font-size="11" font-weight="700" fill="#374151">Arm D — dense 1.5B</text>
	<text x="1028" y="726" font-size="9.5" fill="#1F2937">Direct-answer SFT</text>
	<text x="1028" y="742" font-size="9.5" fill="#1F2937">3× the parameters, one pass</text>
	<text x="1028" y="758" font-size="9.5" fill="#1F2937">Tests scale as a substitute for depth</text>
	<text x="1028" y="806" font-size="9" fill="#6B7280" font-style="italic">scale control</text>

	<text x="640" y="850" font-size="9.5" fill="#6B7280" text-anchor="middle">All arms evaluated on identical frozen rows under the same reader (Section 4.2).</text>
	<text x="640" y="878" font-size="9.5" fill="#9CA3AF" text-anchor="middle">Latent Space Reasoning program · Paper One · architecture comparison · 2026-07-27</text>
	</svg>