Issue with `runtime_observation_count`

[lqwk]

❓ Questions and Help

As mentioned in our previous communications, I tried adding debugging logs into the code to print out the runtime_observation_count and corresponding callstacks to debug why we are seeing only 1 observation for starting runtimes.

See the debug logs in this branch.

Repro

The command I tested this with is:

python -m llvm_autotuning.tune -m \
    experiment=my-exp \
    outputs=/tmp/logs \
    executor.cpus=32 \
    num_replicas=1 \
    autotuner=nevergrad \
    autotuner.optimization_target=runtimeseries \
    autotuner.search_time_seconds=120 \
    benchmarks=single_benchmark_for_testing

and I got the following logs

Details

• I verified that the runtime_observation_count is setup properly when we create the optimization target: debug logs

• I saw that the previous_runtimes had a correct number of observations of 30 but during reset, the observation count was 1: debug logs and I couldn't figure out why.

• I also saw that there was a time when the runtime_observation_count was set to 1: debug logs. The logs of the corresponding callstack is also printed here.

[ChrisCummins]

Hey, I'm going to need your help debugging this, as logs and stacktraceds aren't useful to me without knowing the details of your code, and the repro command requires changes that aren't in upstream CompilerGym.

Could you please reduce your repro example to a standalone snippet of code that runs against an unmodified CompilerGym package and demonstrates the issue? Then I can run it on my end and see what the underlying problem is.

Cheers,
Chris