remove the different rtol / atol for openpilot CUDA in benchmark (#2907)

not sure what the issue was but seems to be fixed on master
2023-12-21 22:23:39 -05:00
parent a543d8bea8
commit 1bbeb3fe2f
1 changed files with 0 additions and 1 deletions
@@ -99,7 +99,6 @@ def benchmark_model(m, devices, validate_outs=False):

  if validate_outs:
    rtol, atol = 2e-3, 2e-3  # tolerance for fp16 models
-    if m == "openpilot" and 'CUDA' in devices: rtol, atol = 0.1, 0.1  # TODO: why is this broken?
    inputs = {k:Tensor(inp) for k,inp in np_inputs.items()}
    tinygrad_model = get_run_onnx(onnx_model)
    tinygrad_out = tinygrad_model(inputs)