llama3_70b

Llama3-70B Demo

How to Run

For Users on TT-VPN

If you have access to TT-VPN, you can copy the weights directly to your local machine using the following SCP commands:

1. Copying repacked Llama3-70B weights:

   scp -r 10.230.36.208:/home/llama3-data-repacked/llama-3-70b/ <repacked_output_dir>

2. Copying Llama3-70B tokenizer:

   scp -r 10.230.36.208:/home/llama3-data-repacked/tokenizer.model <path_to_checkpoint_dir>

For Users without TT-VPN Access

If you do not have access to TT-VPN, follow these steps to download the weights directly from Meta and use the repacking script:

1. Download the Llama3-70B weights from Meta (https://llama.meta.com/):

2. Repack the weights:

   # This concatenates the sharded checkpoints and makes it easier for us to load.
   python models/demos/t3000/llama2_70b/scripts/repack_weights.py <path_to_checkpoint_dir> <repacked_output_dir>

Running the Demo

After setting up the repacked weights and tokenizer, you can run the demo using the commands below:

1. Install requirements for reference/llama folder:

   pip install blobfile

2. Prepare the weight cache directory:

   # Make a directory for us to cache weights into. This speeds up subsequent runs.
   mkdir <weight_cache_dir>

3. Set up environment variables:

   export LLAMA_CKPT_DIR=<repacked_output_dir>
   export LLAMA_TOKENIZER_PATH=<path_to_checkpoint_dir> # Path needs to include the tokenizer.model file
   export LLAMA_CACHE_PATH=<weight_cache_dir>
   
   # Example:
   # export LLAMA_CKPT_DIR="/home/llama3-data-repacked/llama-3-70b/"
   # export LLAMA_TOKENIZER_PATH="/home/llama3-data-repacked/tokenizer.model"
   # export LLAMA_CACHE_PATH="/home/llama3-data-cache/weights-cache"

4. Cache the weights (first-time setup only):

   # Build a full 80 layer model to cache the weights. This will take some time.
   pytest -svv models/demos/t3000/llama2_70b/tests/test_llama_model.py::test_LlamaModel_inference[decode-8chip-T3000-80L]

5. Run the demo:

   # Run the demo using sampling decode
   pytest -svv models/demos/t3000/llama3_70b/demo/demo.py::test_LlamaModel_demo[sampling-tt-70b-80L]

Details

• Batch Size: Supports batch size 32.
• Input File: Uses ./demo/data/multi_prompt.json.
• Model Configuration: Utilizes a pretrained model.
• Hardware Requirements: Runs on an 8-chip T3000 machine using tensor parallelism. The host machine must have at least 512 GB of memory.
• Model Functionality: Implements decode-to-prefill strategy, where prompts are processed token-by-token to produce KV caches, followed by token generation in decode mode.

Ensure you follow these guidelines to successfully run the Llama3-70B demo.

Benchmarking

For best performance results, please do the following:

• Set environment variable TT_METAL_ASYNC_DEVICE_QUEUE=1.
• Ensure the cpu frequency governor is set to performance.