branch: master
benchmark.yml
31278 bytesRaw
name: Benchmarks
env:
# TODO: this rescheduling makes gpt2, mixtral and llama unjitted slower
# TODO: very slow for llama 70B and resnet training 6 GPU
CAPTURE_PROCESS_REPLAY: "1"
ASSERT_PROCESS_REPLAY: "0"
PYTHONPATH: .
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
on:
push:
branches:
- master
- update_benchmark
- update_benchmark_staging
workflow_dispatch:
inputs:
run_process_replay:
description: "Run process replay tests"
required: false
default: false
type: boolean
jobs:
testmacbenchmark:
name: Mac Benchmark
runs-on: [self-hosted, macOS]
timeout-minutes: 20
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Symlink models and datasets
run: |
mkdir -p weights
ln -s ~/tinygrad/extra/disassemblers/applegpu extra/disassemblers/applegpu
ln -s ~/tinygrad/weights/sd-v1-4.ckpt weights/sd-v1-4.ckpt
ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: python3.11 test/external/process_replay/reset.py
- name: Run Stable Diffusion
run: JIT=1 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
- name: Run Stable Diffusion without fp16
run: JIT=1 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
- name: Run Stable Diffusion v2
run: JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
- name: Run SDXL
run: JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
- name: Run model inference benchmark
run: METAL=1 python3.11 test/external/external_model_benchmark.py
- name: Test speed vs torch
run: BIG=2 MPS=1 python3.11 test/test_speed_v_torch.py | tee torch_speed.txt
- name: Test tensor cores
run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
- name: Test AMX tensor cores
run: |
DEBUG=2 CPU=1 AMX=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
DEBUG=2 LLVM=1 AMX=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
- name: Run Tensor Core GEMM (float)
run: DEBUG=2 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
- name: Run Tensor Core GEMM (half)
run: DEBUG=2 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
- name: Run Tensor Core GEMM (bfloat16)
run: DEBUG=2 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
- name: Fuzz Padded Tensor Core GEMM
run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
- name: Run LLaMA
run: |
JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
- name: Run LLaMA with BEAM
run: JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
- name: Run quantized LLaMA
run: |
python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
#- name: Run LLaMA 7B on 4 (virtual) GPUs
# run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_four_gpu.txt
- name: Run GPT2
run: |
JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
JIT=1 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
- name: Run GPT2 w HALF
run: HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
- name: Run GPT2 w HALF/BEAM
run: HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
- name: Train MNIST
run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
- name: Run 10 CIFAR training steps
run: JIT=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
- name: Run 10 CIFAR training steps w HALF
run: JIT=2 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
#- name: Run 10 CIFAR training steps w BF16
# run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
- name: Run 10 CIFAR training steps w winograd
run: JIT=1 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
- uses: actions/upload-artifact@v4
with:
name: Speed (Mac)
path: |
onnx_inference_speed.csv
torch_speed.txt
llama_unjitted.txt
llama_jitted.txt
llama_beam.txt
llama_int8.txt
llama_nf4.txt
llama_four_gpu.txt
gpt2_unjitted.txt
gpt2_jitted.txt
gpt2_half.txt
gpt2_half_beam.txt
matmul.txt
matmul_half.txt
matmul_bfloat16.txt
sd.txt
sd_no_fp16.txt
sdv2.txt
sdxl.txt
beautiful_mnist.txt
train_cifar.txt
train_cifar_half.txt
train_cifar_bf16.txt
train_cifar_wino.txt
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py
testnvidiabenchmark:
name: tinybox green Benchmark
runs-on: [self-hosted, Linux, tinyboxgreen]
timeout-minutes: 30
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Print nvidia-smi
run: nvidia-smi
- name: Symlink models and datasets
run: |
mkdir -p weights
ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
mkdir -p extra/datasets
ln -s /raid/datasets/imagenet extra/datasets/imagenet
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: test/external/process_replay/reset.py
- name: Run model inference benchmark
run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
- name: Test speed vs torch
run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
- name: Test speed vs theoretical
run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
- name: Test benchmark allreduce
run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
- name: Test tensor cores
run: |
NV=1 ALLOW_TF32=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
PTX=1 ALLOW_TF32=1 NV=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
- name: Run Tensor Core GEMM (CUDA)
run: |
CUDA=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
CUDA=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
CUDA=1 ALLOW_TF32=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
- name: Run Tensor Core GEMM (PTX)
run: NV=1 PTX=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
- name: Run Tensor Core GEMM (NV)
run: NV=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
- name: Test NV=1
run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
- name: Test CUDA=1
run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
- name: Run Stable Diffusion
run: NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
- name: Run SDXL
run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
- name: Run LLaMA
run: |
NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
- name: Run LLaMA with BEAM
run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
# - name: Run LLaMA 7B on 4 GPUs
# run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_four_gpu.txt
# - name: Run LLaMA 7B on 6 GPUs
# run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_six_gpu.txt
- name: Run LLaMA-3 8B BEAM
run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
- name: Run LLaMA-3 8B on 4 GPUs with BEAM
run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
# - name: Run LLaMA-3 8B on 6 GPUs
# run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
# - name: Run LLaMA-2 70B
# run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_2_70B.txt
- name: Run Mixtral 8x7B
run: time NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
- name: Run GPT2
run: |
NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
NV=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
- name: Run GPT2 w HALF
run: NV=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
- name: Run GPT2 w HALF/BEAM
run: NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
- uses: actions/upload-artifact@v4
with:
name: Speed (NVIDIA)
path: |
onnx_inference_speed.csv
torch_speed.txt
matmul.txt
matmul_bfloat16.txt
matmul_tf32.txt
matmul_ptx.txt
matmul_nv.txt
sd.txt
sdxl.txt
llama_unjitted.txt
llama_jitted.txt
llama_beam.txt
llama3_beam.txt
llama3_four_gpu.txt
llama3_six_gpu.txt
llama_2_70B.txt
mixtral.txt
gpt2_unjitted.txt
gpt2_jitted.txt
gpt2_half.txt
gpt2_half_beam.txt
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
testmorenvidiabenchmark:
name: tinybox green Training Benchmark
runs-on: [self-hosted, Linux, tinyboxgreen]
timeout-minutes: 20
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Symlink models and datasets
run: |
mkdir -p weights
ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
mkdir -p extra/datasets
ln -s /raid/datasets/imagenet extra/datasets/imagenet
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: test/external/process_replay/reset.py
- name: Fuzz Padded Tensor Core GEMM (NV)
run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
- name: Fuzz Padded Tensor Core GEMM (PTX)
run: NV=1 PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
- name: Train MNIST
run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
- name: Run 10 CIFAR training steps
run: NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
- name: Run 10 CIFAR training steps w HALF
run: NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
- name: Run 10 CIFAR training steps w BF16
run: NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
- name: Run 10 CIFAR training steps w winograd
run: NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
- name: Run full CIFAR training w 1 GPU
run: time NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
- name: Run full CIFAR training steps w 6 GPUS
run: time CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
- name: Run MLPerf resnet eval on training data
run: time NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
run: NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
run: NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
- name: Run 10 MLPerf Bert training steps (6 gpu)
# TODO: remove BERT_LAYERS once scheduler is fast
run: NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
- uses: actions/upload-artifact@v4
with:
name: Speed (NVIDIA Training)
path: |
beautiful_mnist.txt
train_cifar.txt
train_cifar_half.txt
train_cifar_bf16.txt
train_cifar_wino.txt
train_cifar_one_gpu.txt
train_cifar_six_gpu.txt
train_resnet.txt
train_resnet_one_gpu.txt
train_bert.txt
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
testamdbenchmark:
name: tinybox red Benchmark
runs-on: [self-hosted, Linux, tinybox]
timeout-minutes: 20
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Insert amdgpu
run: sudo modprobe amdgpu
- name: Symlink models and datasets
run: |
mkdir -p weights
ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
mkdir -p extra/datasets
ln -s /raid/datasets/imagenet extra/datasets/imagenet
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: test/external/process_replay/reset.py
- name: setup perflevel
run: |
examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
rocm-smi
- name: Show off tinybox
run: /opt/rocm/bin/rocm-bandwidth-test
# TODO: unstable on AMD
#- name: Run model inference benchmark
# run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
# TODO: unstable on AMD
#- name: Test speed vs torch
# run: |
# python3 -c "import torch; print(torch.__version__)"
# LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
- name: Test speed vs theoretical
run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
- name: Test tensor cores
run: AMD=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
- name: Run Tensor Core GEMM (AMD)
run: AMD=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
- name: Test AMD=1
run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
- name: Test HIP=1
run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
# TODO: AMD compiler bug causes this to fail
#- name: Fuzz Padded Tensor Core GEMM
# run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
- name: Remove amdgpu
run: sleep 5 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
- name: Test AM cold start time
run: time AMD=1 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
- name: Test AM warm start time
run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
- name: Run Stable Diffusion
run: AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
- name: Run SDXL
run: AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
- name: Run LLaMA 7B
run: |
AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
- name: Run LLaMA 7B with BEAM
run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
# - name: Run LLaMA 7B on 4 GPUs
# run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_four_gpu.txt
# - name: Run LLaMA 7B on 6 GPUs
# run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_six_gpu.txt
- name: Run LLaMA-3 8B BEAM
run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
- name: Run LLaMA-3 8B on 4 GPUs with BEAM
run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
# - name: Run LLaMA-3 8B on 6 GPUs
# run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
- name: Restore amdgpu
run: sudo modprobe amdgpu
# - name: Run LLaMA-2 70B
# run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_2_70B.txt
- name: Run Mixtral 8x7B
run: time AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
- name: Run GPT2
run: |
AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
AMD=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
- name: Run GPT2 w HALF
run: AMD=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
- name: Run GPT2 w HALF/BEAM
run: AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
- uses: actions/upload-artifact@v4
with:
name: Speed (AMD)
path: |
onnx_inference_speed.csv
torch_speed.txt
llama_unjitted.txt
llama_jitted.txt
llama_beam.txt
llama3_beam.txt
llama3_four_gpu.txt
llama3_six_gpu.txt
llama_2_70B.txt
gpt2_unjitted.txt
gpt2_jitted.txt
gpt2_half.txt
gpt2_half_beam.txt
matmul.txt
matmul_amd.txt
sd.txt
sdxl.txt
mixtral.txt
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
testmoreamdbenchmark:
name: tinybox red Training Benchmark
runs-on: [self-hosted, Linux, tinybox]
timeout-minutes: 30
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: Remove amdgpu
run: sudo rmmod amdgpu || true
- name: Symlink models and datasets
run: |
mkdir -p weights
ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
mkdir -p extra/datasets
ln -s /raid/datasets/imagenet extra/datasets/imagenet
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: test/external/process_replay/reset.py
- name: Train MNIST
run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
- name: Run 10 CIFAR training steps
run: AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
- name: Run 10 CIFAR training steps w HALF
run: AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
- name: Run 10 CIFAR training steps w BF16
run: AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
- name: Run 10 CIFAR training steps w winograd
run: AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
- name: Run full CIFAR training w 1 GPU
run: time AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
- name: Run full CIFAR training steps w 6 GPUS
run: time AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
- name: Run MLPerf resnet eval
run: time AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
run: AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
run: AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
- name: Run 10 MLPerf Bert training steps (6 gpu)
# TODO: remove BERT_LAYERS once scheduler is fast
run: AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
- uses: actions/upload-artifact@v4
with:
name: Speed (AMD Training)
path: |
beautiful_mnist.txt
train_cifar.txt
train_cifar_half.txt
train_cifar_bf16.txt
train_cifar_wino.txt
train_cifar_one_gpu.txt
train_cifar_six_gpu.txt
train_resnet.txt
train_resnet_one_gpu.txt
train_bert.txt
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
testqualcommbenchmark:
name: comma Benchmark
runs-on: [self-hosted, Linux, comma]
timeout-minutes: 20
defaults:
run:
shell: bash -o pipefail {0}
if: github.repository_owner == 'tinygrad'
steps:
- name: Checkout Code
uses: actions/checkout@v4
- name: setup staging db
if: github.ref == 'refs/heads/update_benchmark_staging'
run: |
echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
- name: reset process replay
run: test/external/process_replay/reset.py
- name: validate openpilot 0.9.7
run: PYTHONPATH=. FLOAT16=0 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
- name: benchmark openpilot 0.9.4
run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_4.txt
- name: benchmark openpilot 0.9.7
run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
- name: benchmark openpilot w IMAGE=2 0.9.4
run: PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_4.txt
- name: benchmark openpilot w IMAGE=2 0.9.7
run: PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
- name: openpilot compile3 0.9.7
run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx
- name: openpilot compile3 0.9.7+ tomb raider
run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/e8bea2c78ffa92685ece511e9b554122aaf1a79d/selfdrive/modeld/models/supercombo.onnx
- name: openpilot dmonitoring compile3 0.9.7
run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/dmonitoring_model.onnx
- name: Run process replay tests
run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
- uses: actions/upload-artifact@v4
with:
name: Speed (comma)
path: |
openpilot_compile_0_9_4.txt
openpilot_compile_0_9_7.txt
openpilot_0_9_4.txt
openpilot_0_9_7.txt
openpilot_image_0_9_4.txt
openpilot_image_0_9_7.txt