Kaixi Chen

Deep Learning Compilers | Inference Perf @NVIDIA

[Kaixi Chen @matteo_chen]

@difficultyang Find the symm mem rank being stuck

[Kaixi Chen @matteo_chen]

@maharshii I didn’t know to code till 20

[Kaixi Chen @matteo_chen]

@ezyang This was a great analysis!

[Kaixi Chen @matteo_chen]

@drisspg Is this for CG?

[Kaixi Chen @matteo_chen]

@tenderizzation I might recognize that request

[Kaixi Chen @matteo_chen]

@tenderizzation Lmaooo

[Kaixi Chen @matteo_chen]

@encapsulated007 @henrylhtsang ~yeah

[Kaixi Chen @matteo_chen]

@auto_grad_ @pz_ai1 TORCH_LOGS=“recompiles” can dump all the issues. Then you should fix your code to avoid those triggers…or manually trigger them before the workload of interest if they’re contained/can’t refactor the code.

[Kaixi Chen @matteo_chen]

@pz_ai1 @auto_grad_ I had a similar issue, I ended up warming up all the possible shapes before the actual run 🤭

[Kaixi Chen @matteo_chen]

@tenderizzation 😂

[Kaixi Chen @matteo_chen]

Good boy

[Kaixi Chen @matteo_chen]

I really thought the Hadamard transform would stay buried in my quantum physics class forever…

…but no, it came back from the dead just to personally make my prefill stage dog slow.

[Kaixi Chen @matteo_chen]

@tenderizzation Really

[Kaixi Chen @matteo_chen]

@jcz42 Could you share which issues you’ve been finding with torch compile on blackwell?

[Kaixi Chen @matteo_chen]

@jsuarez @mohitwt_ Is there a repro open source?

[Kaixi Chen @matteo_chen]

@tenderizzation Hope at least it was the old gr86

[Kaixi Chen @matteo_chen]

@jackminong Yeah, seen it. Thanks!

[Kaixi Chen @matteo_chen]

@tenderizzation Looking fwd to see this in DTensor

[Kaixi Chen @matteo_chen]

Brothers it’s time to master profiling not coding