Keren Zhou

Triton int64 overflow bug

2

To compute `(a * b) % c`, if `a * b` is greater than `max(int32_t)`, then the final result is incorrect. Reproducer: https://github.com/Jokeren/triton-samples/blob/main/int64-mod.py

Potential problems in testing.py

6

Hi @ptillet , I started playing triton a few days ago. I have read through your tutorials on vec-add and matrix multiplications. I should say triton is very convenient for...

- int8 matmul is skipped because of precision problems - If the kernel is in the warmup mode, we skip interpreter execution - All interpreter-related tests have been moved under...

Should be 2 instead of 1, right?

Perf numbers comparing against torch inductor and blade disc?

2

Just curious about the perf comparison between this work and others that also adopt the `compile` interface to optimize the computation graph. See https://github.com/alibaba/BladeDISC/releases/tag/v0.3.0

Add missing cython dependency

8

Otherwise pip install will ignore cython