Can you reproduce the configuration?
Keep a record of the model and code revision, framework and library versions, hardware, batch size, sequence or input sizes, precision, and parallelism settings. A comparison is easier to interpret when these are explicit.
Include the expected range of inputs in the evaluation. A check that exercises only the easiest input can leave the more demanding parts of a workload unexamined.
What does memory do during the check?
- Observe both CPU and GPU memory across representative steps.
- Record peak usage and whether usage continues growing across repeated steps.
- Investigate the allocations and call stacks associated with unexpected behavior.
- Record the effect of any proposed configuration change on both memory and throughput.
MemSage preflight checks can help identify configurations that risk memory failures or inefficient resource use. A check reduces uncertainty about the workload it exercised; it does not guarantee that every future run will succeed.
What should you agree on before a MemSage evaluation?
- Compatibility: confirm the exact framework, versions, and hardware with the team.
- Installation: obtain setup instructions for your environment and start on one workload.
- Overhead: compare the same workload with and without the diagnostic tooling.
- Data handling: confirm what is collected, where it is processed and stored, and the access and retention arrangements.
- Success criteria: define the memory question you need answered and the evidence that would make the evaluation useful.
What should you bring to the evaluation?
Bring your workload configuration, a representative failure, and the question you want to answer. See the MemSage evaluation overview for the next steps.
This is an evaluation checklist, not an installation manual or a list of guaranteed supported configurations.