Dharma-AI published a research preview on November 18 about predicting fine-tuning duration from 272 real SFT runs using H200 accelerators. The authors study the period from GPU allocation to completion, excluding queue time. The full paper is still described as forthcoming. Results do not establish universal accuracy for other hardware or organizations.
Compute duration versus job duration
The engineering question is what an operator is actually reserving. Theoretical operation counts describe only part of occupied time. A practical reservation must reach the artifact needed by the next task. If that endpoint is defined differently, two valid estimates can appear contradictory.
Our editorial example would record separate timestamps for resource allocation, the start of main work and confirmed final output. That makes variable overhead visible instead of assigning every delay to the model. It is an observability proposal, not a repetition of the published experiment.
A prediction must be useful before execution
The study uses features available before a run begins. That matters for scheduling: an estimate secretly depending on future progress cannot guide initial resource allocation. The authors identify limited data for very long runs and the need for recalibration on different infrastructure.
For a local evaluation, we would hold out later real jobs from those used for tuning. It would be useful to report not only average error but also how often duration is underestimated. In a dependent workflow, occasional large overruns may matter more than a small average difference.
Do not optimize time at the expense of output
Our interpretation is that shorter training is insufficient as a success criterion. Removing necessary evaluation can make a process finish earlier while leaving quality less well established. Duration should therefore be compared alongside predefined requirements for the resulting model.
A useful scheduler might provide a completion range and clearly flag cases outside familiar conditions. An unusually large job or changed software stack should be treated as a shift in circumstances, rather than receiving the same confidence by default.
The verified news is a published preview based on production observations. SIExpo has not trained the predictor or confirmed savings from deployment. The proposition worth evaluating locally is to measure the complete useful process rather than its nominal compute component alone.



