mamba

Mamba

Inference Demo

This demo is designed to run Mamba-2.8b on a wormhole_b0 card and generate outputs for a set of prompts. We used finetuned version state-spaces/mamba-2.8b-slimpj for quality outputs. Follow the instructions below to run the demo successfully.

How to Run

To get the best performance during decode we can use the 8x8 core grid. To enable it run the following command:

export WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml

To run the demo using pre-written prompts for a batch of 32 users run:

pytest --disable-warnings -q -s --input-method=json --input-path='models/demos/wormhole/mamba/demo/prompts.json' models/demos/wormhole/mamba/demo/demo.py

To run the demo using custom input prompts, you can provide a different path to the input prompts file for e.g.:

pytest --disable-warnings -q -s --input-method=json --input-path='path_to_input_prompts.json' models/demos/wormhole/mamba/demo/demo.py

Any sequence length is supported. We currently only support JSON file with strictly 1 user prompt or 32 user prompts with same token length. Check the models/demos/wormhole/mamba/demo/prompts.json file for reference.

The prefill graph is not currently integrated into the demo. Therefore we currently process the prompt a single token at a time using the decode graph.

---

Unit Tests

These unit tests are designed to test the Mamba model and its components. The tests are written using the pytest framework.

Navigate to the tt-metal directory

cd tt-metal

SSM Block

pytest -svv models/demos/wormhole/mamba/tests/test_mamba_ssm.py

Mamba Block

pytest -svv models/demos/wormhole/mamba/tests/test_mamba_block.py

Residual Block

pytest -svv models/demos/wormhole/mamba/tests/test_residual_block.py

Full Model

Note : input embedding layer and TopK are on CPU

pytest -svv models/demos/wormhole/mamba/tests/test_mamba_model.py::test_inference

Performance Tests

These tests are designed to evaluate device-side and host performance of Mamba model. The tests are written using the pytest framework. Navigate to the tt-metal directory

End-to-End Model Performance

pytest -svv models/demos/wormhole/mamba/tests/test_mamba_perf.py -m models_performance_bare_metal

Device-Side Performance

Build with profiler support enabled (use the build script ./build_metal.sh --enable-profiler) and run the following command to test device-side performance:

pytest -svv models/demos/wormhole/mamba/tests/test_mamba_perf.py -m models_device_performance_bare_metal

This will also generate device and host profiling logs in directory generated/profiler/reports/ttnn_mamba