Falcon40b prefill uses 8x8 core grid size, so the following environment variable needs to be set on a T3000 setup:
export WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml- To run the model for a single prompt, you can use the command line input:
pytest --disable-warnings -q -s --input-method=cli --cli-input="YOUR PROMPT GOES HERE!" models/demos/t3000/falcon40b/demo/demo.py`- A sample of input prompts for 32 users is provided in
models/demos/t3000/falcon40b/demo/input_data.json. - If you wish to run the model using a different set of input prompts you can provide a different path
--input-path.
pytest --disable-warnings -q -s --input-method=json --input-path='models/demos/t3000/falcon40b/demo/input_data.json' models/demos/t3000/falcon40b/demo/demo.py`- Weight caching: This model picks up certain configs and weights from huggingface pretrained model. The default model weights are the
tiiuae/falcon-40b-instructversion from huggingface. The first time you run the model, weights are downloaded, pre-processed, and stored on your machine. This might take a few hours. The second time you run the model on your machine, the weights are being read from cached files on your machine and it will be faster. - Max Context Length: The maximum context/sequence length for the demo is currently limited to 128 tokens. Support for context length 2048 is in testing.
- Batch Size: Currently only a batch size of 32 is supported.
- Token Generation Scheme: The model will first run in prefill mode on the input sequences to fill the KV cache and then in decode mode to generate the output tokens.