CGO-26-AE

Install Environment

Init Git Submodules

git submodule init
git submodule update

Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

Install NV Env if you are on GH200

UV_SKIP_WHEEL_FILENAME_CHECK=1
uv sync --extra cuda
. .venv/bin/activate

Install AMD Env if you are on MI300X

UV_SKIP_WHEEL_FILENAME_CHECK=1
uv sync --extra rocm
. .venv/bin/activate

Instructions to Reproduce Experiments

Before proceeding, make sure you have activated python env by . .venv/bin/activate.

Major Results

Figure 13 and Figure 14

1. cd fig13_14
2. Start experiments by python run_experiment_nv.py if on NVIDIA or python run_experiment_amd.py if on AMD. The results will be both displayed in terminal and stored in a json file.
3. Plot Figure 13 by python fig13.py. The script will automatically read json file and plot AMD part of the figure.
4. Plot Figure 14 by python fig14.py.
5. Check generated fig13.pdf and fig14.pdf

Figure 15

1. cd fig15
2. Start experiments by bash run_exp_and_visualize.sh
3. Plot Figure 15 by python fig15.py. The generated figure is fig15.pdf

Optional Results

Figure 8

cd fig8
bash run.sh

The proton profile will be visualized and printed in the terminal outputs.

Figure 9

cd fig9
bash run.sh

Data and figures would be generated under fig9/triton_kernels/bench/logs/gpt-oss-x2

Figure 10

cd fig10
bash run.sh

Note

This experiment does not work with ROCM 7.0 due to ROCM bugs, and it requires the user to install ROCM 6.4. Please try uv pip install https://download.pytorch.org/whl/rocm6.4/torch-2.9.0%2Brocm6.4-cp313-cp313-manylinux_2_28_x86_64.whl

Similar to figure 9, data and figures would be generated under fig10/triton_kernels/bench/logs/gpt-oss-x2

Figure 12

cd fig12
bash run.sh

Similar to figure 8, the proton profile will be visualized and printed in the terminal.