Triton self learning

environment

triton 3.4.0
torch 2.8.0+cu128
cuda 12.8
RTX5090 (capability 12.0)

Note

• There are some accuracy problems with matmul and linear.
• RTX5090 doesn't support certain features like blockwise scaled matmul.
• RTX5090 will encounter some OOM error. Try to reduce the problem size.
• Please check the corresponding version tag of triton tutorial, don't use main branch.
• Some perfermance reports are weird. Please check the correctness.
• The first param torch.testing.assert_close() is true value.
• Be careful of the dtype used for input, computation, accumulation and the output.
• Be careful of the alignment and edge case.
• Leave the ref link, it will help you in the future.
• Use ls ~/.triton/cache to check the kernel compilation cache.
• Triton kernel is useful for highly fused and customized operation.

set up

• linux bash shell settings

export PS1="\u@\h:\W> " (\W only display current directory)
export PS1="\u:\W> "

• query the device capability

import torch
torch.cuda.get_device_capability()

• check autodl github for github access source /etc/network_turbo

• debug

export TRITON_INTERPRET=1
unset TRITON_INTERPRET

link

triton_tutorial