run on update_benchmark too (#1460)

* run on update_benchmark too

* amd inference test

* name it better

* add 10 CIFAR training steps

.github/workflows/benchmark.yml

@@ -4,18 +4,36 @@ on:
   push:
     branches:
       - master
+      - update_benchmark
 
 jobs:
   testmacbenchmark:
     name: Mac Benchmark
-    runs-on: self-hosted
+    runs-on: [self-hosted, macOS]
 
     steps:
     - name: Checkout Code
       uses: actions/checkout@v3
-    - name: Run metal ops test
+    - name: Run model inference benchmark
       run: python3 test/external/external_model_benchmark.py
+    - name: Run 10 CIFAR training steps
+      run: STEPS=10 python3 examples/hlb_cifar10.py
     - uses: actions/upload-artifact@v3
       with:
-        name: ONNX Inference Speed
+        name: ONNX Inference Speed (Mac)
+        path: onnx_inference_speed.csv
+  testamdbenchmark:
+    name: AMD Benchmark
+    runs-on: [self-hosted, Linux]
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v3
+    - name: Run model inference benchmark
+      run: python3 test/external/external_model_benchmark.py
+    - name: Run 10 CIFAR training steps
+      run: STEPS=10 python3 examples/hlb_cifar10.py
+    - uses: actions/upload-artifact@v3
+      with:
+        name: ONNX Inference Speed (AMD)
         path: onnx_inference_speed.csv

test/external/external_model_benchmark.py

@@ -3,9 +3,11 @@ import pathlib
 import time
 import onnx
 import torch
+torch.set_num_threads(1)
 from onnx2torch import convert
 from extra.utils import download_file
 from extra.onnx import get_run_onnx
+from tinygrad.helpers import OSX
 from tinygrad.tensor import Tensor
 from tinygrad.lazy import Device
 
@@ -53,7 +55,7 @@ def benchmark_model(m):
   np_inputs = {k:torch.randn(shp).numpy() for k,shp in input_shapes.items()}
   assert len(input_shapes) < 20
 
-  for device in ["METAL", "CLANG"]:
+  for device in ["METAL" if OSX else "GPU", "CLANG"]:
     Device.DEFAULT = device
     inputs = {k:Tensor(inp) for k,inp in np_inputs.items()}
     tinygrad_model = get_run_onnx(onnx_model)
@@ -69,9 +71,10 @@ def benchmark_model(m):
   torch_inputs = [torch.tensor(x) for x in np_inputs.values()]
   benchmark(m, "torch_cpu", lambda: torch_model(*torch_inputs))
 
-  torch_mps_model = torch_model.to('mps')
-  torch_mps_inputs = [x.to('mps') for x in torch_inputs]
-  benchmark(m, "torch_mps", lambda: torch_mps_model(*torch_mps_inputs))
+  torch_device = "mps" if OSX else "cuda"
+  torch_mps_model = torch_model.to(torch_device)
+  torch_mps_inputs = [x.to(torch_device) for x in torch_inputs]
+  benchmark(m, f"torch_{torch_device}", lambda: torch_mps_model(*torch_mps_inputs))
 
   if open_csv is None:
     open_csv = csv.DictWriter(open('onnx_inference_speed.csv', 'w', newline=''), fieldnames=list(CSV.keys()))