pl752

Results 4 comments of pl752

Flash attention implementations do not handle case where value vectors have different dimension from query vectors

!unstale What about introducing configurations specific for Deepseek llms and adding cmake flags for enabling building these configurations, so default configuration build list isn't getting bloated, or adding flags for...

Eval: HIP: Llama-server multi-instance lockup

Okay, thank you for the reply. I will try to investigate further and give updates on this weird behaviour.

Eval: HIP: Llama-server multi-instance lockup

Here are some backtraces with debug symbols for amdhsa and hip runtime and with RelWithDbg build for some of the interesting threads Rocgdb bt output ``` slot launch_slot_: id 0...

AMDGPU is broken on 1.8

Also libomp.so is missing