Clip fp8 to +/-240 on all targets. (#1172)

* clip fp8 to +/-240 on all targets

* if inputs to fp8 conversion are +/-inf, they remain unaltered

* increase tolerance for test_elementwise_layernorm to prevent false errors

* change the input values for gemm examples to floats

* reduce gemm example float input values to prevent errors

* increase the tolerance for gemm examples

[ROCm/composable_kernel commit: d0c7b45150]
This commit is contained in:
Illia Silin
2024-02-27 12:31:05 -08:00
committed by GitHub
parent d2750747d4
commit f0f3b65e2a
4 changed files with 16 additions and 13 deletions

View File

@@ -233,7 +233,7 @@ bool profile_elementwise_layernorm_impl(int do_verification,
y_dev.FromDevice(y.mData.data());
bool pass =
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 5e-3, 5e-3);
if(do_log)
{