github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-10 07:28:15 -05:00
Code Issues Packages Projects Releases Wiki Activity

Onnx Layer/Group/RMS/Batch-Norm ReduceL2 fp32 intermediates for fp16 (#12109)

Browse Source 
* match onnx spec

* use least_upper_dtype

* promote the square

* just cast before the square
This commit is contained in:
Sieds Lykles
2025-10-24 12:26:11 +02:00
committed by GitHub
parent 0bde87d8d7
commit e1f8c82938
2 changed files with 20 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
test/external/external_test_onnx_ops.py vendored
View File

@@ -272,6 +272,10 @@ class TestMainOnnxOps(TestOnnxOps):
def test_qlinearmatmul_2D_int8_float32(self): self._run_qlinearmatmul_test(np.int8, np.float32, 2)
def test_qlinearmatmul_3D_int8_float32(self): self._run_qlinearmatmul_test(np.int8, np.float32, 3)
def test_reduce_l2_half(self):
inputs = {"data": np.random.randn(1, 1, 32, 32, 32).astype(np.half)*100}
self.helper_test_single_op("ReduceL2", inputs, {}, ["reduced"])
class TestTrainingOnnxOps(TestOnnxOps):
# NOTE: ORT doesn't actually support training ops on cpu so we test using functions provided by onnx
DOMAIN = AI_ONNX_PREVIEW_TRAINING_DOMAIN
@@ -487,4 +491,4 @@ class TestContribOnnxOps(TestOnnxOps):
self.helper_test_single_op("QLinearGlobalAveragePool", inputs, attributes, outputs)
if __name__ == "__main__":
unittest.main()
unittest.main()