Visualization graph shows gray layers

[htwang14]

Hi, when I check the visualization of inference jobs (e.g., job jgzj9ejkp), I find many quantized layers to be gray and without a "compute unit" or "delegate" in the overview, as show in this screenshot:

But the "Run time layer analysis" do show all layers to be on NPU:

Is this simply a display issue on the visualization graph, or the compute units are actually unknown for those gray layers?

Thank you in advance!

[gustavla]

@htwang14 Thanks for filing this! This is indeed a bug.

Gray ops can legitimately happen when the ops are fused away and never end up running individually at the lowest level. Obviously, it can't skip entire Conv2D ops, so that they are missing is a name mismatch bug between the asset's op names and the op names reported by the runtime.

I have filed an internal issue for us to take a look. Thanks for reporting it!

[htwang14]

I see the Convolutions have 0 compute cycles while the ReLUs have much more than they should. I suppose this indicates the computes of convolutions are incorrectly fused into ReLUs when doing the statistics as you indicated? Are the total compute cycles and latency still accurate, or they skipped the cost of all the convolutions too?

[qaihm-bot]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[gustavla]

We are still tracking this internally, so even if this gets marked as stale and closes, we are still working on this.

[qaihm-bot]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[qaihm-bot]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.