Instructions to use datasysdev/ann-sparseattention with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use datasysdev/ann-sparseattention with Transformers:

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("datasysdev/ann-sparseattention", dtype="auto")

llama-cpp-python

How to use datasysdev/ann-sparseattention with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="datasysdev/ann-sparseattention",
	filename="gguf/Qwen3-4B-Instruct-2507-F16-ann-6layer-k128-v2.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use datasysdev/ann-sparseattention with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf datasysdev/ann-sparseattention:F16
# Run inference directly in the terminal:
llama cli -hf datasysdev/ann-sparseattention:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf datasysdev/ann-sparseattention:F16
# Run inference directly in the terminal:
llama cli -hf datasysdev/ann-sparseattention:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf datasysdev/ann-sparseattention:F16
# Run inference directly in the terminal:
./llama-cli -hf datasysdev/ann-sparseattention:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf datasysdev/ann-sparseattention:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf datasysdev/ann-sparseattention:F16

Use Docker

docker model run hf.co/datasysdev/ann-sparseattention:F16

LM Studio
Jan
Ollama
How to use datasysdev/ann-sparseattention with Ollama:
```
ollama run hf.co/datasysdev/ann-sparseattention:F16
```

Unsloth Studio

How to use datasysdev/ann-sparseattention with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for datasysdev/ann-sparseattention to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for datasysdev/ann-sparseattention to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for datasysdev/ann-sparseattention to start chatting

How to use datasysdev/ann-sparseattention with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf datasysdev/ann-sparseattention:F16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "datasysdev/ann-sparseattention:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use datasysdev/ann-sparseattention with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf datasysdev/ann-sparseattention:F16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default datasysdev/ann-sparseattention:F16

Run Hermes

hermes

Atomic Chat new
Docker Model Runner
How to use datasysdev/ann-sparseattention with Docker Model Runner:
```
docker model run hf.co/datasysdev/ann-sparseattention:F16
```

Lemonade

How to use datasysdev/ann-sparseattention with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull datasysdev/ann-sparseattention:F16

Run and chat with the model

lemonade run user.ann-sparseattention-F16

List all available models

lemonade list

datasysdev commited on May 8

Commit

dd33ae2

verified ·

1 Parent(s): 0989d52

Document dynamic index proxy

Browse files

Files changed (1) hide show

README.md +28 -0

README.md CHANGED Viewed

@@ -268,6 +268,34 @@ about 3x in favor of learned projections. This supports the theoretical
 scaling claim only; production speed claims still require GPU-resident
 retrieval and KV-cache/decode integration.
 ### Compute / quality knobs (FLOP-counted)
 `L = 4096`. Compute reduction is the attention scoring step, `≈ L / K`.

 scaling claim only; production speed claims still require GPU-resident
 retrieval and KV-cache/decode integration.
+### Dynamic-index proxy
+The current ANN wrapper is prefill-only (`use_cache=False`), so a true
+generation-time dynamic-index benchmark still requires cache integration. As a
+first capability proxy, `dynamic_index_proxy.py` splits clean block-causal eval
+sequences into a prefill prefix and decode-like suffix, then compares learned
+retrieval mass under two masks:
+- dynamic index: suffix queries can retrieve from all same-segment prior keys;
+- static index: suffix queries can retrieve from prefill keys plus a 256-token
+  recent local suffix window, but not older suffix keys.
+On the clean d128 block-causal checkpoint, using K=128, prefill length 1024,
+local window 256, and 8 eval batches:
+| Setting | Teacher mass captured |
+|---|---:|
+| Dynamic proxy | 0.972 |
+| Static proxy | 0.928 |
+| Static teacher mass available | 0.954 |
+| Dynamic - static | +0.044 |
+Per-layer dynamic-minus-static mass ranges from +0.022 (L04) to +0.058 (L08).
+This does not establish task accuracy or decode latency, but it shows that a
+frozen prefill-plus-local index loses measurable teacher-attention mass on
+decode-like suffix queries. The raw result is in
+`artifacts/dynamic_proxy_8b.json`.
 ### Compute / quality knobs (FLOP-counted)
 `L = 4096`. Compute reduction is the attention scoring step, `≈ L / K`.