When I open the profile on the Web UI and try looking into the HLO graph of a particular kernel that I see on a stream in the Trace view, whose respective XLA Ops is part of a command buffer, it will only show the command buffer name (something like command_buffer_x with x being a number) on the XLA Ops row. Therefore there is no individual XLA Ops name to put into the XLA Op Name field on the Graph type: HLO Graph view. Putting command_buffer_x into this field will just result in the UI saying Could not find HloInstruction or HloComputation names command_buffer_x. (I also tried changing XLA Modules field to different things already, according to this issue: #2458)
Is this a feature or a bug?
If it is a feature, is it recommended to set the --xla_gpu_enable_command_buffer="" for producing a profile with HLO graphs for all kernels?
When I open the profile on the Web UI and try looking into the HLO graph of a particular kernel that I see on a stream in the Trace view, whose respective XLA Ops is part of a command buffer, it will only show the command buffer name (something like
command_buffer_xwithxbeing a number) on the XLA Ops row. Therefore there is no individual XLA Ops name to put into theXLA Op Namefield on the Graph type: HLO Graph view. Puttingcommand_buffer_xinto this field will just result in the UI sayingCould not find HloInstruction or HloComputation names command_buffer_x.(I also tried changing XLA Modules field to different things already, according to this issue: #2458)Is this a feature or a bug?
If it is a feature, is it recommended to set the --xla_gpu_enable_command_buffer="" for producing a profile with HLO graphs for all kernels?