🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4117

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure

As of commit a1b96e4 with merge base eb64bfb ():

This comment was automatically generated by Dr. CI and updates every 15 minutes.

can we have a test?

can we have a test?

@vkuzo @slayton58 correct me if i'm wrong but i don't think this bug is reachable from our pytorch code? to reach a biased scale of 254 we need a very large value with exponent bits of 254, and we only can pass bf16 or fp32 (same dynamic range, so let's use fp32).

the max fp32 value is (2 - (2**-23)) * (2**127) which has exponent 254. However, when using RCEIL scaling (only path that uses exp2f_rcp) this is divided by 448 and converted with the cvt.sat.finite inline ptx and this results in scale.view(uint8) of 247, which is below the target value of 254 to hit this code path.

tl;dr We can't create large enough test inputs for the kernel to hit this path since they aren't representable in fp32 or bf16.

@vkuzo @slayton58 ok i determined we can pass "inf" which will result in biased exponent of 254 and returns scales.view(uint8) == 254. So i guess this case isn't reachable via finite values but is special case for infinity?

@vkuzo @slayton58 ok i determined we can pass "inf" which will result in biased exponent of 254 and returns scales.view(uint8) == 254. So i guess this case isn't reachable via finite values but is special case for infinity?

Sounds about right - even looking at MXFP4 we'd still only hit 252, which is still < 254. I think we'd need a format where max(fmt) == 1 to hit this, which we don't have.