Closed Ldpe2G closed 4 months ago
@Ldpe2G Thanks for reporting the issue. reduceOp is a known issue on Navi GPUs. We are working on it. Will keep you posted.
@joviliast What is the current status of reduction op on Navi3?
@zhanglx13 It should be supported for now.
Tried to reproduce this case: On https://github.com/ROCm/triton/tree/triton-mlir and https://github.com/openai/triton/tree/main results are exactly the same on the different sizes:
Thank you Illia. @Ldpe2G I'll close this one. We are merging in the upstream triton. For further issues, feel free to open tickets at https://github.com/openai/triton/issues
Environment
1e2fd0dd1a986a79fc3d4b3d4cfc3396098eb418
2.3.0.dev20240107+rocm5.7
Issue Description
I have encountered an issue where the sum computed using Triton's JIT does not match the sum computed by PyTorch's
torch.sum
. This discrepancy occurs when running the provided code snippet, which is intended to calculate the sum of a random tensor.To Reproduce
Here is a minimal code snippet to reproduce the issue:
Expected Behavior
I would expect the
y_triton
tensor to be very close toy_torch
tensor, with differences within the acceptable tolerance set bytorch.allclose
(atol=1e-2, rtol=0).Actual Behavior
However, the assertion fails, which indicates that the sums do not match within the specified tolerances. The actual values returned are:
y_triton
: 33.0685y_torch
: 32.5651Could you please look into this discrepancy?
Thank you for your assistance.